A complex problem knowledge externalization modeling system and method based on human-computer dialogue

Abstract

The application discloses a complex problem knowledge externalization modeling system and method based on human-computer conversation, and belongs to the technical field of computers. Through compliance preliminary screening and structured coding of multi-source cross-modal knowledge, a dynamic conflict graph with double time markers is constructed, and the conflict detection sensitivity is dynamically adjusted based on the user cognitive state, solving the problems of traditional human-computer conversation system knowledge fusion black box and cross-modal conflict detection deficiency. The decision logic is encapsulated as a visual decision path graph through a dynamic arbitration engine, and the presentation complexity is adjusted according to the user cognitive state, realizing the transparency and interpretability of the arbitration process. At the same time, the user intervention result is obtained in real time and strategy recalculation is triggered, combining with the ethical pre-check and example library mechanism, a closed-loop system of deep coupling of conflict detection, dynamic arbitration, user intervention and ethical compliance is constructed, promoting the co-evolution of each module and improving the credibility and intervenability of the system.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a system and method for externalizing knowledge of complex problems based on human-computer dialogue. Background Technology

[0002] Currently, human-computer dialogue systems typically employ knowledge fusion technology to integrate multi-source knowledge and generate answers when dealing with complex problems. However, the fusion process is often encapsulated within the model's internal computation, forming a difficult-to-interpret "knowledge fusion black box." Users can only passively receive the final result without understanding the decision-making basis, leading to insufficient system credibility and operability.

[0003] While related technologies involve knowledge base construction and feedback iteration, conflict detection largely relies on single-modal semantic matching, making it difficult to discover implicit conflicts between cross-modal knowledge. The detection results are disconnected from the subsequent arbitration process, lacking a two-way interaction mechanism, and the detection module cannot dynamically adjust parameters based on arbitration feedback. Furthermore, user feedback is typically a post-hoc, batch-based indirect optimization, isolated from the real-time decision-making process; each user intervention fails to be systematically transformed into reusable knowledge.

[0004] On the other hand, with increasingly stringent requirements for ethical review of artificial intelligence, existing solutions generally treat ethical compliance as an "additional module" added after the fact, rather than embedding it into the core algorithm design, leading to conflicts between compliance assurance and system performance. How to construct a deeply coupled closed-loop system that integrates conflict detection, dynamic arbitration, user intervention, and ethical compliance, enabling each module to evolve collaboratively, has become an urgent technical challenge to be solved. Summary of the Invention

[0005] This application provides a system and method for externalizing knowledge in complex problems based on human-computer dialogue, and the technical solution is as follows: On the one hand, a method for externalizing knowledge of complex problems based on human-computer dialogue is provided, the method comprising: The system acquires complex target questions, multi-source cross-modal knowledge data, dialogue context features, user cognitive state features, and data traceability authorization information in human-computer dialogue scenarios. After compliance screening and structured coding, it obtains a standardized knowledge set, a user cognitive state vector, and a compliance tag set. Based on the standardized knowledge set and the compliance tag set, a cross-modal semantic unified mapping and conflict identification are performed through a cross-modal conflict graph builder to construct a dynamic conflict graph with dual temporal labels. The knowledge node weights of the dynamic conflict graph are constrained by the compliance tag set, and the conflict detection sensitivity of the dynamic conflict graph is dynamically adjusted based on the user cognitive state vector. Based on the dynamic conflict graph, the dialogue context features, the user cognitive state vector, and the compliance tag set, the decision logic, knowledge source weights, and conflict resolution paths corresponding to the arbitration strategy are encapsulated into a visual decision path graph through the dynamic arbitration engine. The presentation complexity of the visual decision path graph is adjusted based on the user cognitive state vector, and the arbitration strategy needs to undergo ethical pre-checking. Based on the visualized decision path graph, user intervention results are obtained, and the intervention results are parsed into standardized intervention paradigms and transmitted back in real time to trigger recalculation of arbitration strategies. The intervention paradigms are stored in the example library after ethical pre-checking. The example library is used to optimize the arbitration strategy generation logic for similar conflict scenarios in the future, and the dynamic conflict graph and the user cognitive state vector are updated simultaneously based on the intervention results. Attached Figure Description

[0006] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0007] Figure 1 This is a flowchart of a method for externalizing knowledge of complex problems based on human-computer dialogue, provided in an embodiment of this application. Figure 2 This is a flowchart of another method for externalizing knowledge of complex problems based on human-computer dialogue, provided in an embodiment of this application; Figure 3 This is a flowchart of another method for externalizing knowledge of complex problems based on human-computer dialogue, provided in an embodiment of this application. Figure 4 This is a flowchart of another method for externalizing knowledge of complex problems based on human-computer dialogue, provided in an embodiment of this application. Figure 5 This is a flowchart of another method for externalizing knowledge of complex problems based on human-computer dialogue, provided in an embodiment of this application. Figure 6 This is a schematic diagram of the structure of a complex problem knowledge externalization modeling system based on human-computer dialogue provided in an embodiment of this application. Detailed Implementation

[0008] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0009] In this application, the terms "first," "second," etc., are used to distinguish identical or similar items with essentially the same function. It should be understood that there is no logical or temporal dependency between "first," "second," and "nth," nor are there any restrictions on quantity or execution order.

[0010] It should be noted that the information (including but not limited to user device information, user personal information, etc.), data (including but not limited to data used for analysis, data stored, data displayed, etc.) and signals involved in this application are all authorized by the user or fully authorized by all parties, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0011] Traditional human-computer dialogue systems often create an inexplicable "black box" during the knowledge fusion process when handling complex problems, leaving users unable to understand the basis for their decisions and resulting in insufficient system credibility and operability. Related technologies rely heavily on single-modal semantic matching for conflict detection, making it difficult to discover implicit cross-modal conflicts. Furthermore, the detection and arbitration processes are disconnected, lacking bidirectional interaction. User feedback is typically provided as post-hoc optimization, isolated from the real-time decision-making process. In addition, ethical compliance is often added as an add-on module rather than embedded in the core algorithm design, leading to conflicts between compliance assurance and system performance. How to construct a deeply coupled closed-loop system that enables the co-evolution of its modules has become a pressing technical challenge.

[0012] To address this, this application proposes a method for externalizing knowledge in complex problems based on human-computer dialogue, see [link to relevant documentation]. Figure 1 ,include: 101. Obtain the target complex problem, multi-source cross-modal knowledge data, dialogue context features, user cognitive state features and data traceability authorization information in the human-computer dialogue scenario. After compliance screening and structured coding, obtain a standardized knowledge set, user cognitive state vector and compliance label set.

[0013] 102. Based on the standardized knowledge set and the compliance tag set, cross-modal semantic unified mapping and conflict identification are performed through the cross-modal conflict graph builder to construct a dynamic conflict graph with dual temporal labels. The knowledge node weights of the dynamic conflict graph are constrained by the compliance tag set, and the conflict detection sensitivity of the dynamic conflict graph is dynamically adjusted based on the user's cognitive state vector.

[0014] 103. Based on the dynamic conflict graph, the dialogue context features, the user cognitive state vector, and the compliance tag set, the decision logic, knowledge source weights, and conflict resolution paths corresponding to the arbitration strategy are encapsulated into a visual decision path graph through the dynamic arbitration engine. The presentation complexity of the visual decision path graph is adjusted based on the user cognitive state vector. The arbitration strategy needs to undergo ethical pre-checking.

[0015] 104. Based on the visualized decision path map, obtain the user intervention results, parse the intervention results into a standardized intervention paradigm, and transmit it back in real time to trigger the recalculation of the arbitration strategy. After ethical pre-checking, store the intervention paradigm in the example library. This example library is used to optimize the arbitration strategy generation logic for similar conflict scenarios in the future, and simultaneously update the dynamic conflict map and the user cognitive state vector based on the intervention results.

[0016] For ease of understanding, the following explains some key terms in this embodiment: Complex problems in human-computer dialogue scenarios: These refer to questions posed by users during human-computer interaction that require the system to deeply understand, integrate multiple information sources, and reason in order to provide a satisfactory answer. These problems typically involve multiple domains and information types, and may contain potential conflicts or uncertainties. Multi-source cross-modal knowledge data: This refers to knowledge information originating from different channels (e.g., text, images, audio, video, structured databases, etc.) and presented in diverse formats (e.g., textual, visual, auditory). This data needs to be integrated and utilized by the system to solve complex problems. Dialogue context features: These refer to the characteristics reflected in all interactive information prior to the current dialogue turn during human-computer dialogue, including but not limited to historical dialogue content, the order of user questions, and the system's response history, used to help the system understand the context of the current question. User cognitive state features: These refer to the cognitive state exhibited by the user during interaction with the system, such as the user's level of professional knowledge, the urgency of the current task, emotional state, and level of attention. These features guide the system in providing personalized responses and information presentation. Data source traceability authorization information: This refers to metadata related to the source of knowledge data, including the data collection source, collection timestamp, authorization certificate of the data provider, and scope of use, used to ensure the legality, compliance, and timeliness of knowledge data. Standardized knowledge set: This refers to a collection of knowledge units that have undergone compliance screening and structured coding, possessing a unified format, unified semantics, and conforming to preset specifications. This knowledge set is the foundation for the system's knowledge reasoning and conflict detection. User cognitive state vector: This refers to the numerical representation obtained by fusing and encoding user cognitive state features. This vector can quantify dimensions such as the user's professional level and task urgency, used to dynamically adjust system behavior, such as conflict detection sensitivity or information presentation complexity. Compliance tag set: This refers to the set of compliance identifiers associated with knowledge units in the standardized knowledge set. These identifiers originate from the legality verification of the data source traceability authorization information, used to constrain the weight of knowledge nodes and ensure the compliance of knowledge use. Cross-modal conflict graph builder: This refers to the functional module responsible for semantically mapping and conflict identification of multi-source cross-modal knowledge, and constructing a dynamic conflict graph. It can handle potential conflicts between different modalities of data and assign temporal tags to knowledge units. Dynamic Conflict Graph: This refers to a knowledge network represented by a graph structure, where nodes represent knowledge units and edges represent conflict relationships between them. This graph has dual temporal markers; the weights of knowledge nodes are constrained by a set of compliant tags, and the sensitivity of conflict detection can be dynamically adjusted based on the user's cognitive state vector. Dual temporal markers refer to the two timestamps assigned to each knowledge unit in the dynamic conflict graph: an event validity timestamp (the valid time range of the knowledge content itself) and a system entry timestamp (the time the knowledge was included in the system), used to manage the timeliness and version of knowledge.Dynamic Arbitration Engine: This module is responsible for generating arbitration strategies and encapsulating them into a visual decision path diagram based on a dynamic conflict graph, dialogue context features, user cognitive state vectors, and compliance tag sets. It can adjust the arbitration logic according to real-time status. Visual Decision Path Diagram: This diagram presents the arbitration decision-making process graphically, including decision logic, knowledge source weight allocation, and conflict resolution paths. The complexity of this graph can be adjusted based on the user's cognitive state vector to improve interpretability and user intervention capability. Standardized Intervention Paradigm: This is a unified format representation obtained by structuring and encoding user intervention operations (such as node weight adjustment, path selection, knowledge injection, etc.) through the visual decision path diagram. This paradigm is used for real-time feedback to trigger strategy recalculation. Example Library: This knowledge base stores standardized intervention paradigms that have undergone ethical pre-checking. This library is used to accumulate user intervention experience, optimize the arbitration strategy generation logic for similar future conflict scenarios, and achieve continuous learning and evolution of the system.

[0017] This method addresses the shortcomings of traditional human-computer dialogue systems, such as the black box nature of knowledge fusion and insufficient cross-modal conflict detection, by performing compliance screening and structured encoding on multi-source cross-modal knowledge, constructing a dynamic conflict graph with dual temporal tags, and dynamically adjusting the conflict detection sensitivity based on the user's cognitive state. A dynamic arbitration engine encapsulates the decision logic into a visual decision path graph, adjusting the presentation complexity according to the user's cognitive state, thus achieving transparency and interpretability in the arbitration process. Simultaneously, real-time acquisition of user intervention results triggers policy recalculation. Combined with ethical pre-screening and an example library mechanism, a closed-loop system deeply coupled with conflict detection, dynamic arbitration, user intervention, and ethical compliance is constructed, promoting the co-evolution of each module and enhancing the system's credibility and operability.

[0018] However, in the implementation of the above-mentioned methods, the lack of an effective modal alignment and feature unification mechanism for multi-source cross-modal data, the logic of knowledge screening guided by the fusion encoding of user cognitive state, and the comprehensive legality verification and compliance verification process for data traceability authorization information leads to low quality of preprocessed data, insufficient identification of cross-modal semantic conflicts, and insufficient integration of user personalized needs, which in turn affects the accuracy of subsequent conflict detection and the credibility of arbitration decisions.

[0019] To address this, this application further proposes a method for acquiring complex target questions, multi-source cross-modal knowledge data, dialogue context features, user cognitive state features, and data source authorization information in human-computer dialogue scenarios. After compliance screening and structured coding, a standardized knowledge set, user cognitive state vector, and compliance label set are obtained. (See [link to relevant documentation]). Figure 2 The method includes: 201. Perform semantic analysis on the complex problem to extract the problem intent and key entities, and obtain the core elements of the problem.

[0020] 202. Based on the core elements of the problem, modal alignment and feature unification are performed on the multi-source cross-modal knowledge data to obtain aligned knowledge items that are semantically related to the core elements of the problem.

[0021] 203. The context features of the dialogue and the user's cognitive state features are fused and encoded to generate a user cognitive state vector to guide knowledge selection.

[0022] 204. Based on the core elements of the problem and the user's cognitive state vector, candidate knowledge items are selected from the aligned knowledge items, and the legality of the data traceability authorization information is verified to generate a data source compliance identifier for each candidate knowledge item. Then, the candidate knowledge item is verified for compliance based on the data source compliance identifier. The verified knowledge items are encoded into the standardized knowledge set, and the user's cognitive state vector is output. The user's cognitive state vector is used to dynamically adjust the conflict detection sensitivity in subsequent steps. At the same time, the data source compliance identifier is aggregated into a compliance tag set associated with the standardized knowledge set according to preset rules.

[0023] For example, semantic parsing of a complex problem to extract its intent and key entities, thus obtaining its core elements, refers to using natural language processing (NLP) technology to deeply understand the complex problem posed by the user and transform it into structured information that can be processed by a machine. The question intent refers to the user's fundamental purpose in asking the question or the core problem they hope to solve. Key entities refer to the specific objects, concepts, or attributes involved in the question. Through semantic parsing, the system can accurately grasp user needs, providing precise guidance for subsequent knowledge retrieval and processing.

[0024] Based on the core elements of the problem, modal alignment and feature unification are performed on the multi-source cross-modal knowledge data to obtain aligned knowledge items that are semantically associated with the core elements of the problem. This aims to address the differences in representation and semantic gaps between multi-source knowledge data (such as text, images, audio, and video). By mapping data from different modalities to a unified semantic space and extracting features related to the core elements of the problem, effective comparison and fusion of knowledge from different sources and in different forms can be ensured. Aligned knowledge items are knowledge units that, after modal alignment and feature unification, can be semantically associated with the core elements of the problem.

[0025] The dialogue context features and the user's cognitive state features are fused and encoded to generate a user cognitive state vector to guide knowledge selection. The dialogue context features reflect historical information of the current dialogue, such as previous question-and-answer rounds, topics mentioned by the user, and system suggestions. The user cognitive state features describe non-static attributes such as the user's current knowledge level, mood, task urgency, and preferences. Fusion encoding integrates this heterogeneous information into a unified vector representation that reflects the user's current cognitive and demand state. This user cognitive state vector plays a crucial guiding role in the subsequent knowledge selection process, ensuring that the selected knowledge better matches the user's personalized needs and the current dialogue context. Based on the core elements of the question and the user cognitive state vector, candidate knowledge items are selected from the aligned knowledge items. This step, based on the user's current question focus and cognitive state, initially filters out the most relevant and potentially useful knowledge fragments from a large amount of aligned knowledge. The core elements of the question ensure the direct relevance of knowledge to the user's question, while the user cognitive state vector incorporates personalized and contextual considerations, making the selection results more targeted. Candidate knowledge items are a subset of knowledge that, after initial screening, will proceed to the subsequent compliance verification and coding stages.

[0026] The data source authorization information is validated for legality to generate a data source compliance identifier for each candidate knowledge item. This data source authorization information includes key metadata such as the source of the knowledge data, collection time, and usage permissions. Legality validation verifies this information to ensure that the acquisition and use of knowledge comply with laws, regulations, ethical standards, and user authorization agreements. The data source compliance identifier is a structured representation of the validation results; it provides proof of the legality and compliance of each candidate knowledge item and serves as the basis for subsequent compliance checks.

[0027] Based on the compliance identifier of the data source, the candidate knowledge item undergoes compliance verification. The verified knowledge items are encoded into the standardized knowledge set. The compliance verification process involves judging the compliance of candidate knowledge items based on the data source compliance identifier and pre-defined compliance policy rules. Only knowledge items that pass the verification are considered trustworthy, usable, and ethically compliant. The standardized knowledge set is a high-quality collection of knowledge that has undergone rigorous screening and encoding. It not only contains the knowledge content itself but also includes its compliance attributes, providing reliable input for subsequent conflict detection and arbitration.

[0028] The system outputs a user cognitive state vector, which is used to dynamically adjust the conflict detection sensitivity in subsequent steps. This user cognitive state vector, after generation, is not only used for knowledge filtering but also serves as important contextual information passed to the subsequent conflict detection module. Its function is to dynamically adjust the stringency of conflict detection based on factors such as the user's expertise level and the urgency of the task. For example, for experienced users or urgent tasks, the system may increase the sensitivity of conflict detection to avoid potentially misleading information. Conversely, for novice users or non-urgent tasks, the sensitivity may be appropriately reduced to minimize unnecessary interference.

[0029] Simultaneously, the compliance identifiers from the data source are aggregated into a compliance tag set associated with the standardized knowledge set according to preset rules. This compliance tag set is a collection that summarizes and abstracts the data source compliance identifiers of all knowledge items in the standardized knowledge set. It is not simply a list of identifiers for each knowledge item, but rather, based on preset aggregation rules (e.g., statistically analyzing the distribution of different compliance dimensions, calculating the overall compliance score, and identifying major compliance risk types), it forms a tag set that macroscopically reflects the compliance status of the entire knowledge set. This tag set can serve as a constraint in the subsequent conflict graph construction, influencing the weight of knowledge nodes and the logic of conflict detection.

[0030] Through the aforementioned technical solutions, this application addresses the issues of modality alignment and feature unification in the preprocessing stage of multi-source cross-modal data, ensuring the comparability and consistency of knowledge across different modalities at the semantic level. This resolves the problem of insufficient cross-modal semantic conflict identification and provides unified, high-quality knowledge input for subsequent conflict detection. Simultaneously, by fusing and encoding dialogue context features and user cognitive state features, the knowledge selection process fully considers users' personalized needs and the current dialogue context, improving the accuracy of knowledge selection and user experience, and providing a crucial basis for dynamically adjusting conflict detection sensitivity. Furthermore, by verifying the legality and compliance of data source authorization information, the legality, credibility, and timeliness of knowledge data are guaranteed from the source, effectively mitigating ethical and compliance risks and improving the overall quality and reliability of the standardized knowledge set.

[0031] In some of the embodiments described above in this application, modal alignment and feature unification are proposed to obtain aligned knowledge items that are semantically related to the core elements of the problem. However, in the implementation process, the potential conflict tendency between different modal knowledge items is not effectively identified, which may lead to the omission of implicit conflicts in subsequent conflict detection, affecting the accurate construction of dynamic conflict graphs and arbitration efficiency.

[0032] To address this, this application further proposes a method for performing modal alignment and feature unification on multi-source cross-modal knowledge data based on core problem elements to obtain aligned knowledge items semantically associated with the core problem elements. This method includes: performing modality-specific encoding on each modality data in the multi-source cross-modal knowledge data to obtain an initial modal feature set; extracting key modal features semantically related to the current problem from the initial modal feature set based on the core problem elements and the user's cognitive state vector; projecting these key modal features into a shared semantic space, and constraining the relative positions of different modal features in the shared semantic space during the projection process using a conflict-oriented contrastive loss function, which is used to shorten the distance between semantically consistent features and widen the distance between potentially conflicting features; and generating the aligned knowledge item with a conflict-sensitive label based on the feature distribution in the shared semantic space, which is used to identify potential conflict tendencies between different modal knowledge items.

[0033] For example, modal-specific encoding is performed on each modality of the multi-source cross-modal knowledge data to obtain an initial modal feature set. This aims to transform different forms of raw data (such as text, images, and audio) into a unified numerical feature representation while preserving the unique information of each modality. For instance, for text data, word embedding models (such as Word2Vec and GloVe) or Transformer-based encoders (such as BERT and RoBERTa) can be used to convert it into a vector representation. For image data, convolutional neural networks (such as ResNet and VGG) or visual Transformers (such as ViT) can be used to extract its visual features. For audio data, Mel-frequency cepstral coefficients (MFCCs) or recurrent neural networks (RNNs) can be used for feature extraction. This approach lays the foundation for subsequent cross-modal fusion processing.

[0034] Based on the core elements of the problem and the user's cognitive state vector, key modal features relevant to the current problem's semantics are extracted from the initial modal feature set. This step aims to filter out the most relevant and valuable parts from a massive set of initial features based on the user's current focus and cognitive state, reducing interference from irrelevant information and improving processing efficiency and accuracy. For example, an attention mechanism can be used, where the core elements of the problem and the user's cognitive state vector serve as queries, and the initial modal feature set serves as keys and values. Attention weights are calculated to highlight features highly relevant to the problem's semantics. Alternatively, a gating mechanism can be used to dynamically adjust the activation level of features based on the core elements of the problem and the user's cognitive state vector, thereby achieving the extraction of key features.

[0035] The key modal features are projected into a shared semantic space, and during the projection process, a conflict-oriented contrastive loss function constrains the relative positions of different modal features within this shared semantic space. The shared semantic space is an abstract, modality-independent feature space that allows for direct comparison of semantically related knowledge from different modalities. The conflict-oriented contrastive loss function is the core of this step, actively learning and distinguishing between semantically consistent and potentially conflicting features. For example, a Siamese network or triplet network structure can be constructed. For semantically consistent cross-modal feature pairs, the loss function will bring them closer together in the shared space. Conversely, for cross-modal feature pairs identified as potentially conflicting, the loss function will increase the distance between them, thus achieving pre-identification and separation of conflicts at the feature level.

[0036] Based on the feature distribution in the shared semantic space, the aligned knowledge item is generated with a conflict-sensitive label. This label identifies potential conflict tendencies between knowledge items of different modalities. In the shared semantic space, after training with a conflict-oriented contrastive loss function, the feature distribution of the knowledge item reflects its semantic consistency or conflict tendency. For example, its conflict sensitivity can be determined based on its clustering in the shared space or its distance from other knowledge items, and a label can be attached, such as a Boolean value, a conflict probability score, or a conflict type label. These labels explicitly indicate the potential conflict risk of the knowledge item, providing important predictive information for subsequent dynamic conflict graph construction and arbitration.

[0037] Through the aforementioned technical solution, this application proactively identifies and labels potential conflict tendencies between different modal knowledge items in the early stages of modality alignment and feature unification by introducing a conflict-oriented contrastive loss function. This solves the problem of traditional methods failing to effectively identify implicit conflicts and avoids omissions in conflict detection. By bringing semantically consistent features closer together and pushing potentially conflicting features further apart in a shared semantic space, this application can more accurately capture subtle differences and contradictions between cross-modal knowledge, thereby generating aligned knowledge items with clear conflict-sensitive labels. These labels provide rich predictive information for the subsequent construction of a dynamic conflict graph, making the construction of the conflict graph more accurate and improving the efficiency and reliability of the subsequent arbitration process, providing a more robust and interpretable foundation for the externalization modeling of complex problem knowledge.

[0038] In some of the embodiments described above in this application, candidate knowledge items are selected based on the core elements of the problem and the user's cognitive state vector to improve the accuracy of knowledge selection. However, in the implementation process, the selection threshold is fixed or the specific dimensions of the user's cognitive state are not considered, and the data source credibility is not combined for correction, which may result in the selection results being inaccurate or unsuitable for different user scenarios.

[0039] To address this, this application further proposes a method for selecting candidate knowledge items from aligned knowledge items based on the core elements of the problem and the user's cognitive state vector. This method includes: determining the semantic matching degree between each aligned knowledge item and the core elements of the problem to obtain an initial relevance score; dynamically determining a screening threshold based on the user's cognitive state vector, which includes a user's professional level dimension and a task urgency dimension, wherein the screening threshold is negatively correlated with the user's professional level dimension and positively correlated with the task urgency dimension; identifying aligned knowledge items with an initial relevance score greater than the screening threshold as preliminary screening knowledge items; and converting the data source compliance identifier corresponding to each preliminary screening knowledge item into a credibility weight based on a preset credibility mapping rule, using this credibility weight to weight-correct the initial relevance score of the preliminary screening knowledge item, and selecting a preset number of preliminary screening knowledge items from high to low according to the weighted corrected scores as candidate knowledge items.

[0040] For example, when determining the semantic matching degree between each aligned knowledge item and the core elements of the question to obtain an initial relevance score, the aim is to evaluate the degree of association between each aligned knowledge item and the user's current complex question's core intent. One approach is to use deep learning-based semantic similarity models, such as BERT and RoBERTa pre-trained language models, to encode the core elements of the question and the aligned knowledge items as vectors, and then calculate the cosine similarity or Euclidean distance between them as the semantic matching degree. Another approach is to use a method based on keyword matching and ontology knowledge graphs, calculating the overlap between keywords in the core elements of the question and entities and concepts in the knowledge items, and combining this with the semantic distance between concepts in the knowledge graph to calculate the matching degree.

[0041] When dynamically determining the filtering threshold based on the user's cognitive state vector, this vector includes a user expertise level dimension and a task urgency dimension. The filtering threshold is negatively correlated with the user expertise level dimension and positively correlated with the task urgency dimension. This step aims to flexibly adjust the strictness of knowledge filtering according to the user's specific cognitive state and task requirements. One implementation method is to use a pre-trained regression model or lookup table. This model takes the user expertise level and task urgency dimensions as input and outputs a dynamic filtering threshold. For example, when the user's expertise level is high, the model outputs a lower threshold, allowing more potentially relevant knowledge to pass through. When the task urgency is high, the model outputs a higher threshold to ensure that only highly relevant, high-quality knowledge is filtered out. Another implementation method is to use a rule engine-based approach, defining a series of rules, such as "if the user's expertise level is 'expert' and the task urgency is 'low,' then the threshold is 0.6. If the user's expertise level is 'novice' and the task urgency is 'high,' then the threshold is 0.8," etc., dynamically matching the corresponding threshold according to the value of the user's cognitive state vector.

[0042] When an aligned knowledge item with an initial relevance score greater than the screening threshold is identified as a preliminary screening item, its function is to perform initial, relevance-based screening, eliminating knowledge items with low relevance to the question. One implementation is to directly perform numerical comparisons, marking all aligned knowledge items with initial relevance scores higher than the current dynamic screening threshold as preliminary screening items. Another implementation is to combine fuzzy logic, allowing a small tolerance range around the threshold, so that knowledge items with scores slightly below the threshold can also be included in the preliminary screening under specific conditions (e.g., when other dimensions perform well).

[0043] Based on a preset credibility mapping rule, the compliance identifier of each initially screened knowledge item is converted into a credibility weight. This credibility weight is then used to weight and correct the initial relevance score of the initially screened knowledge item. A preset number of initially screened knowledge items are selected as candidate knowledge items according to the weighted corrected scores, from high to low. The purpose is to introduce the credibility information of the data source to evaluate and rank the initially screened knowledge items, ensuring that the selected candidate knowledge items are not only relevant but also reliable. One implementation is that the preset credibility mapping rule can be a lookup table or a function that maps the data source compliance identifier (e.g., including authorization status, source credibility, timeliness, etc.) to a credibility weight between 0 and 1. Weighting correction can be achieved by multiplying the initial relevance score by the credibility weight or by performing a weighted average. For example, corrected score = initial relevance score × credibility weight. Another implementation is that the preset credibility mapping rule can be a machine learning-based model that is trained based on the correlation between historical data source compliance identifiers and knowledge item quality, outputting more refined credibility weights. Weighted correction can also employ more complex fusion algorithms, such as linear combinations or nonlinear functions, to fuse relevance scores and confidence weights to generate a corrected score. The corrected scores are then sorted in descending order, and a predetermined number (e.g., Top-K) of knowledge items are selected as candidate knowledge items.

[0044] Through the aforementioned technical solution, this application can dynamically adjust the rigor of knowledge screening based on the user's current cognitive state (such as professional level and task urgency), and combine the data source compliance of knowledge items with credibility weighting, thereby improving the accuracy and reliability of screening results while ensuring knowledge relevance. For example, the introduction of dynamic screening thresholds allows the system to adapt to the professional backgrounds and task needs of different users, avoiding knowledge omissions or redundancies that may be caused by fixed thresholds. For example, for professional users, the system can appropriately lower the threshold to provide more comprehensive information. For urgent tasks, the threshold is raised to prioritize providing highly relevant and reliable information. In addition, by converting data source compliance identifiers into credibility weights and weighting and correcting the relevance scores, it ensures that the selected candidate knowledge items are not only highly relevant to the problem, but also come from credible and compliant data sources, avoiding decision-making risks caused by unreliable information sources, and improving the credibility and decision-making quality of the human-computer dialogue system in handling complex problems.

[0045] In some of the embodiments described above in this application, a method is proposed to verify the legality of data traceability authorization information to generate a data source compliance identifier, in order to ensure the compliance of knowledge items and support subsequent conflict detection and arbitration. However, in its implementation, the verification process may only rely on a single dimension of information, which cannot fully cover multiple factors such as source reliability, authorization validity and timeliness, resulting in the generated compliance identifier being inaccurate and incomplete, affecting the overall system's data compliance assurance and conflict resolution efficiency.

[0046] To address this, this application further proposes a method for verifying the legality of the data tracing authorization information to generate a data source compliance identifier for each candidate knowledge item. Specifically, this includes: extracting the collection source, collection timestamp, and user authorization credential for each candidate knowledge item from the data tracing authorization information; verifying whether the candidate knowledge item conforms to preset authorization validity rules based on the user authorization credential to obtain an authorization status identifier; verifying whether the candidate knowledge item originates from a preset trusted source list based on the collection source to obtain a source trusted identifier; determining whether the candidate knowledge item exceeds a preset timeliness threshold based on the collection timestamp to obtain a timeliness status identifier; and combining and encoding the authorization status identifier, the source trusted identifier, and the timeliness status identifier to generate a data source compliance identifier containing multi-dimensional compliance information.

[0047] This process involves extracting the collection source, collection timestamp, and user authorization credential for each candidate knowledge item from the data traceability authorization information. The aim is to obtain the fundamental metadata required for compliance verification from the raw data. The collection source is used to determine the reliability of the data, the collection timestamp is used to assess the timeliness of the data, and the user authorization credential is used to confirm the data usage rights. By comprehensively extracting this key information, a foundation is laid for subsequent multi-dimensional compliance verification. For example, this information can be obtained by parsing the metadata fields of the data item or the associated log records. For example, for structured data, the predefined "source," "timestamp," and "authorization_token" fields can be read directly. For unstructured data, it may be necessary to identify and extract relevant entities from the text description using Natural Language Processing (NLP) technology. Furthermore, it is also possible to query the traceability chain information of a specific data item by making interface calls with the data management system or blockchain traceability platform, thereby obtaining its original collection source, precise collection timestamp, and the user authorization credential associated with that data item.

[0048] Based on the user's authorization credentials, the system verifies whether the candidate knowledge item conforms to the preset authorization validity rules, obtaining an authorization status identifier. This step verifies whether the use of the candidate knowledge item has obtained legal authorization, ensuring the compliance of data use. The authorization status identifier is a key indicator for measuring whether a knowledge item has usage rights. Preset authorization validity rules may include checking whether the signature of the authorization credentials is valid, whether the authorization credentials have expired, and whether the authorization scope covers the current use case of the knowledge item. For example, the system can call an authorization management module, submit the user's authorization credentials to the module for verification, and the module will return the verification result according to the preset encryption algorithm and authorization policy. Alternatively, the system can integrate with external identity authentication and authorization services (such as OAuth 2.0 or OpenID Connect), sending the user's authorization credentials to the service for verification. The service will return a boolean value or a status code containing authorization details, based on which an authorization status identifier is generated.

[0049] Verifying whether a candidate knowledge item originates from a pre-defined list of trusted sources based on the data source yields a source trust identifier. This step aims to assess the reliability of the data source for the candidate knowledge item, preventing the introduction of information from untrusted or malicious sources, thereby ensuring the quality and credibility of the knowledge. The source trust identifier reflects the reliability of the data source. The pre-defined list of trusted sources can be a whitelist, containing authoritative data sources (such as data released by official institutions, well-known academic databases, etc.) that have undergone manual review or system evaluation. After receiving a data source, the system compares it with this whitelist; if a match is found, it is marked as trusted. Alternatively, a dynamic list of trusted sources based on a reputation scoring mechanism can be used. Each data source has a reputation score, which is dynamically adjusted based on factors such as historical data quality, update frequency, and citation frequency. When the reputation score of a data source exceeds a certain threshold, it is considered trustworthy.

[0050] Based on the data collection timestamp, the system determines whether the candidate knowledge item exceeds a preset timeliness threshold, obtaining a timeliness status identifier. This step checks the timeliness of the candidate knowledge item to ensure that the knowledge used is up-to-date and not expired, avoiding decision-making errors due to the use of outdated information. The timeliness status identifier indicates whether the knowledge item is within a valid time frame. The preset timeliness threshold can be set according to the characteristics of different types of knowledge. For example, news knowledge may require a 24-hour validity period, while scientific principles have lower timeliness requirements. The system compares the data collection timestamp with the current time; if the time difference exceeds the threshold, it is marked as expired. Alternatively, an adaptive timeliness threshold can be adopted, considering the update frequency and domain characteristics of the knowledge item. For example, for financial data, the timeliness threshold might be measured in minutes, while for medical guidelines, it might be measured in years. The system can maintain a timeliness rule base, dynamically matching the corresponding timeliness threshold for judgment based on the type of knowledge item.

[0051] The authorization status identifier, the trusted source identifier, and the timeliness status identifier are combined and encoded to generate a data source compliance identifier containing multi-dimensional compliance information. This step integrates the various compliance indicators obtained from previous independent verifications into a unified and comprehensive data source compliance identifier, facilitating subsequent overall compliance assessment and processing of knowledge items by the system. The combined encoding can use a bitmask, mapping different identifiers to different positions in binary, using an integer value to represent multi-dimensional compliance information. For example, the first bit represents the authorization status, the second bit represents the trusted source, and the third bit represents the timeliness status. Alternatively, structured data formats (such as JSON objects or XML) can be used for encoding, storing the authorization status identifier, trusted source identifier, and timeliness status identifier as different fields in a single data structure. For example, {"authorized":true,"trusted_source":true,"timely":false}.

[0052] The above technical solution extracts the collection source, collection timestamp, and user authorization credential for each candidate knowledge item from the data traceability authorization information. This ensures that the verification basis covers three key dimensions: source reliability, time validity, and user authorization, avoiding information omissions caused by relying on a single attribute. Verifying whether candidate knowledge items conform to preset authorization validity rules based on user authorization credentials and obtaining an authorization status identifier directly verifies the legality of authorization, strengthening the compliance of knowledge item usage. Verifying whether candidate knowledge items originate from a preset trusted source list and obtaining a source trust identifier based on the collection source enhances the assessment of source trustworthiness and prevents the introduction of unreliable data sources. Determining whether candidate knowledge items exceed a preset timeliness threshold based on the collection timestamp and obtaining a timeliness status identifier addresses data freshness issues and ensures timeliness compliance. The authorization status identifier, source trust identifier, and timeliness status identifier are combined and encoded to generate a data source compliance identifier containing multi-dimensional compliance information. Integrating all dimensions of information into a unified identifier facilitates efficient subsequent use. Overall, the systematic processing of multi-dimensional compliance information is achieved through step-by-step, refined verification.

[0053] By generating data source compliance identifiers containing multi-dimensional compliance information, this application can provide more refined and comprehensive input for subsequent compliance verification. This means that when encoding verified knowledge items into standardized knowledge sets, the associated compliance tag set will possess richer dimensional information. This multi-dimensional compliance tag set can more accurately constrain the weights of knowledge nodes in the dynamic conflict graph, making the sensitivity adjustment of conflict detection more intelligent and precise, thereby improving the credibility, compliance, and decision-making efficiency of the entire knowledge externalization modeling method when dealing with complex problems.

[0054] In some of the embodiments described above in this application, a standard knowledge set is generated by performing compliance verification on candidate knowledge items based on data source compliance identifiers. However, in its implementation, the verification process lacks systematic and comprehensive decision-making rules, which may lead to inaccurate or inconsistent verification results, affecting the reliability and ethical compliance of the knowledge set.

[0055] To address this, this application further proposes a method for compliance verification of candidate knowledge items based on data source compliance identifiers, and encoding the verified knowledge items into a standardized knowledge set. This method includes: parsing the data source compliance identifier corresponding to each candidate knowledge item, extracting the authorization status identifier, source trust identifier, and validity status identifier; making a comprehensive judgment based on the authorization status identifier, source trust identifier, and validity status identifier based on compliance policy rules, and determining candidate knowledge items whose judgment results meet a preset compliance threshold as verified knowledge items; and structurally encoding the verified knowledge items to generate standardized knowledge units including the knowledge content ontology and associated compliance attributes. The associated compliance attributes are mapped from the data source compliance identifier, and all standardized knowledge units constitute a standardized knowledge set.

[0056] For example, in the steps of parsing the data source compliance identifier corresponding to each candidate knowledge item and extracting the authorization status identifier, source trust identifier, and validity status identifier, its role is to provide multi-dimensional and fine-grained input information for compliance verification, ensuring the comprehensiveness and accuracy of the verification. This step can be implemented in several ways. For example, a predefined parser or regular expression can be used to directly extract the authorization status identifier, source trust identifier, and validity status identifier from specific fields or encoding formats of the data source compliance identifier. If the data source compliance identifier is a structured data format (such as a JSON string), the corresponding information can be found through key-value pairs. Alternatively, a pre-defined API interface or parsing module can be called, taking the data source compliance identifier as input, which will automatically unpack and return the authorization status identifier, source trust identifier, and validity status identifier. For example, the data source compliance identifier might be an encrypted or encoded string, requiring a specific decoding function to obtain the internal compliance attributes.

[0057] The step of comprehensively judging the authorization status identifier, source trust identifier, and timeliness status identifier based on compliance policy rules, and determining candidate knowledge items whose judgment results meet the preset compliance threshold as verified knowledge items, plays a role in introducing a systematic decision-making mechanism to uniformly evaluate multi-dimensional compliance attributes, avoid subjective judgment, and improve the objectivity and consistency of verification. This step can be implemented by constructing a rule engine-based decision-making system, where compliance policy rules are stored in the form of "IF-THEN" statements.

[0058] In the step of structurally encoding validated knowledge items to generate standardized knowledge units including knowledge content ontology and associated compliance attributes, where associated compliance attributes are mapped from data source compliance identifiers, and all standardized knowledge units constitute a standardized knowledge set, the role is to transform validated knowledge items into a unified and standardized format, facilitating subsequent system processing, storage, and utilization, and ensuring the close association and traceability between knowledge content and compliance attributes. This step can be implemented as follows: Define a unified knowledge unit data structure (such as XML, JSON, or Protobuf format) containing a "knowledge content ontology" field and an "associated compliance attribute" field. During encoding, the original knowledge content is filled into the "knowledge content ontology" field, and information such as the authorization status identifier, source trust identifier, and timeliness status identifier extracted and mapped from the data source compliance identifier is filled into the "associated compliance attribute" field. Alternatively, a knowledge graph can be constructed, with validated knowledge items as nodes in the graph, their knowledge content ontology as node attributes, and associated compliance attributes as specific relationships or metadata attributes associated with that knowledge node. For example, attributes such as "hasAuthorizationStatus", "hasSourceCredibility", and "hasTimelinessStatus" can be added to each knowledge node, and their values ​​are directly derived from the mapping of the data source compliance identifier.

[0059] Through the aforementioned technical solution, this application extracts authorization status, source trustworthiness, and timeliness status identifiers from the data source compliance identifier corresponding to each candidate knowledge item, ensuring a comprehensive, multi-dimensional, and fine-grained assessment of knowledge item compliance and avoiding inaccuracies caused by missing information. Based on compliance policy rules, a comprehensive judgment is made using these extracted identifiers, and preset compliance thresholds are set, providing a systematic and objective decision-making mechanism for the verification process and resolving the problem of inaccurate or inconsistent verification results. This rule-driven verification method avoids the limitations of subjective judgment and improves the reliability and consistency of verification results. The verified knowledge items are structured and encoded to generate standardized knowledge units that include the knowledge content ontology and associated compliance attributes, ensuring that the associated compliance attributes are directly mapped from the data source compliance identifier, thereby tightly binding the knowledge content with its compliance metadata. This not only makes the source, authorization, and timeliness of each knowledge unit clearly traceable, greatly improving knowledge traceability and management efficiency, but also guarantees the quality and credibility of the standardized knowledge set from the source.

[0060] In some of the solutions mentioned above in this application, a comprehensive judgment decision is proposed to verify the compliance of candidate knowledge items. However, in its implementation, because the compliance strategy remains unchanged, it is impossible to dynamically adjust the weights and judgment logic according to changes in the user's cognitive state or the dialogue context, resulting in a lack of flexibility and scenario adaptability in compliance verification, which may lead to misjudgments or inefficiency. To address this, this application further proposes a method for comprehensive judgment decision based on compliance strategy rules for authorization status identifiers, source trust identifiers, and timeliness status identifiers. This method includes: obtaining a compliance strategy rule base, which includes multiple judgment strategy templates for different application scenarios. Each judgment strategy template defines a weight coefficient combination for the authorization status identifier, source trust identifier, and timeliness status identifier, as well as compliance judgment logic. Based on the user's cognitive state vector or the dialogue context features, a target judgment strategy template is dynamically selected from the compliance strategy rule base. Based on the weight coefficient combination defined in the target judgment strategy template, a weighted fusion calculation is performed on the authorization status identifier, the source trust identifier, and the timeliness status identifier to obtain a comprehensive compliance score. Based on the compliance judgment logic defined in the target judgment strategy template, the comprehensive compliance score is compared with the threshold range included in the compliance judgment logic to obtain the judgment decision result.

[0061] For example, the compliance policy rule base can be acquired. This rule base includes multiple judgment policy templates for different application scenarios. Each judgment policy template defines the weight coefficient combination of authorization status identifier, source trust identifier, and timeliness status identifier, as well as the compliance judgment logic. It aims to provide a predefined and configurable set of rules for compliance verification. This rule base can be stored in a database, organized in the form of structured data tables. Each judgment policy template corresponds to one record, containing a description of its applicable scenario, the weight coefficient of each identifier (e.g., the weight of the authorization status identifier is 0.5, the weight of the source trust identifier is 0.3, and the weight of the timeliness status identifier is 0.2), and the compliance judgment logic (e.g., a comprehensive compliance score greater than 0.7 is considered compliant). Alternatively, this rule base can also be managed through configuration files (such as XML or JSON format). Each file represents a judgment policy template, clearly defining the weight configuration and judgment thresholds for different compliance dimensions (authorization status, source trust, and timeliness status).

[0062] Based on the user's cognitive state vector or the features of the dialogue context, the system dynamically selects a target judgment strategy template from the compliance policy rule base. The aim is to intelligently match the most suitable compliance verification strategy according to the actual situation of the current human-computer dialogue. For example, the system can select a judgment strategy template with stricter requirements for the credibility of data sources when the user's professional level is high, based on the user's professional level reflected in the cognitive state vector. When the user's professional level is low, a template that focuses more on authorization status indicators is selected to ensure that the provided information is authorized without issue. Furthermore, based on dialogue context features, such as the urgency or sensitivity of the dialogue, a template with faster judgment speed but potentially slightly relaxed minor compliance conditions can be selected in urgent scenarios, while a more comprehensive and strict template can be selected in non-urgent scenarios. This dynamic selection mechanism can be implemented through a rule engine, matching the user's cognitive state vector or dialogue context features with the applicable scenarios of the judgment strategy template according to preset matching rules. Alternatively, it can be implemented through a machine learning model, such as a classifier, to predict and select the most suitable judgment strategy template based on the input user cognitive state vector and dialogue context features.

[0063] Based on the weighted coefficient combination defined in the target judgment strategy template, this method performs a weighted fusion calculation on the authorization status identifier, the source trust identifier, and the validity period identifier to obtain a comprehensive compliance score. The aim is to quantify compliance information from multiple dimensions into a unified indicator. In practice, a weighted summation method can be used. The authorization status identifier, source trust identifier, and validity period identifier are each converted into a numerical value (e.g., compliance = 1, non-compliance = 0, or a more refined score), then multiplied by the corresponding weighted coefficient defined in the target judgment strategy template, and the products are summed to obtain the comprehensive compliance score. For example, if the authorization status identifier is "authorized" (converted to 1), the source trust identifier is "trusted" (converted to 1), and the validity period identifier is "not expired" (converted to 1), and the template weights are 0.5, 0.3, and 0.2 respectively, then the comprehensive compliance score is 1×0.5 + 1×0.3 + 1×0.2 = 1.0. Another approach is to use fuzzy logic fusion, which represents the compliance level of each identifier as a fuzzy set, and performs fusion calculation through fuzzy operators (such as fuzzy AND and fuzzy OR) and weights to obtain a fuzzy compliance score between 0 and 1.

[0064] Based on the compliance judgment logic defined in the target judgment strategy template, this system compares the comprehensive compliance score with the threshold range included in the compliance judgment logic to obtain the judgment result. Its purpose is to make a compliance judgment based on the quantified comprehensive compliance score. The compliance judgment logic can be a simple threshold judgment; for example, if the comprehensive compliance score is greater than or equal to 0.8, it is judged as "compliant," otherwise as "non-compliant." It can also be a more complex range judgment; for example, 0.9-1.0 is "fully compliant," 0.7-0.9 is "basically compliant," 0.5-0.7 is "partially compliant," and below 0.5 is "non-compliant." Furthermore, the compliance judgment logic can include additional conditional judgments; for example, even if the comprehensive compliance score is high, if the authorization status is marked as "unauthorized," it is directly judged as "non-compliant." Through this comparison, the system can output a clear judgment result to guide subsequent knowledge processing procedures.

[0065] Through the aforementioned technical solution, a compliance policy rule base containing multiple judgment strategy templates is constructed. The system dynamically selects the most suitable template based on user cognitive state vectors or dialogue context features, achieving adaptive adjustment of compliance verification logic. For example, when dealing with complex issues in the healthcare field, the system can dynamically select a judgment strategy template with extremely high requirements for the authority and timeliness of data sources, based on the user's professional background (user cognitive state vector) or the urgency of the dialogue (dialogue context features), thus ensuring that the knowledge provided to the user is highly reliable and up-to-date. When handling general consultation questions, a more lenient template may be selected to improve response efficiency. This dynamic adjustment mechanism avoids misjudgments caused by policy mismatches and improves the accuracy of compliance verification.

[0066] In some of the solutions mentioned above in this application, a dynamic conflict graph with dual temporal labels is proposed to be constructed based on a standardized knowledge set and a compliance label set for cross-modal semantic unified mapping and conflict identification. However, in this process, due to the lack of fine integration of time factors and compliance attributes, the accuracy and dynamism of conflict detection are insufficient. Specifically, it is unable to effectively handle the timeliness differences of knowledge, the impact of compliance credibility on feature representation, and the real-time decay requirement of conflict edge weights. As a result, the conflict identification results may contain outdated or low-credibility knowledge, affecting the reliability of subsequent arbitration decisions.

[0067] To address this, this application further proposes a method for constructing dynamic conflict maps with dual temporal markers, see [link to relevant documentation]. Figure 3 Specifically, it includes: 301. Using a cross-modal conflict graph builder, assign an event valid timestamp and a system entry timestamp to each knowledge unit in the standardized knowledge set as a dual temporal marker for the knowledge unit.

[0068] 302. Using a cross-modal conflict graph builder, features are extracted from each knowledge unit in the standardized knowledge set to obtain the initial feature vector of the corresponding modality. The associated compliance attributes of each knowledge unit are parsed from the compliance tag set. The initial feature vector is weighted by credibility based on the associated compliance attributes to generate a modal feature vector with compliance constraints.

[0069] 303. Using a cross-modal conflict graph builder, modal feature vectors with compliance constraints carrying dual temporal labels are projected into a shared semantic space, and cross-modal semantic similarity between different knowledge units is calculated in the shared semantic space to obtain a cross-modal conflict probability matrix.

[0070] 304. Using a cross-modal conflict graph builder, knowledge unit pairs exceeding the conflict detection threshold are identified as conflict node pairs based on the cross-modal conflict probability matrix. Conflict edges are constructed for each conflict node pair. The initial weights of the conflict edges are determined according to the corresponding probability values ​​in the cross-modal conflict probability matrix. The weights of the conflict edges are subject to a time-decay constraint based on the system entry timestamps of the knowledge units in the conflict node pairs. At the same time, knowledge units that exceed the effective time range are marked as invalid based on the effective event timestamps, thus generating a dynamic conflict graph.

[0071] For example, when assigning an event validity timestamp and a system entry timestamp to each knowledge unit in a standardized knowledge set as dual-temporal markers, the event validity timestamp refers to the time range within which the knowledge content remains valid in the real world. The system entry timestamp refers to the point in time when the knowledge unit is captured, processed, and stored in the standardized knowledge set by the system. These two timestamps together constitute the dual-temporal markers of the knowledge unit, used to accurately track the lifecycle of knowledge and its existence time in the system.

[0072] When extracting features from each knowledge unit in a standardized knowledge set to obtain the initial feature vector of the corresponding modality, the purpose of feature extraction is to convert the knowledge content of different modalities (such as text, images, audio, etc.) into a unified numerical representation, i.e., the initial feature vector, so as to facilitate subsequent calculation and analysis.

[0073] When parsing the associated compliance attributes of each knowledge unit from the compliance tag set, and then weighting the initial feature vector with credibility based on these attributes to generate a modal feature vector with compliance constraints, the compliance tag set contains compliance information related to the knowledge unit, such as the reliability of the data source, authorization status, and timeliness. Parsing the associated compliance attributes refers to extracting specific compliance indicators from these tags. Credibility weighting involves adjusting the initial feature vector based on these compliance attributes to reflect the reliability of the knowledge unit.

[0074] When projecting modal feature vectors with compliance constraints and dual temporal labels onto a shared semantic space, and calculating cross-modal semantic similarity between different knowledge units in this shared semantic space to obtain a cross-modal conflict probability matrix, the shared semantic space is an abstract, unified vector space where knowledge units from different modalities can be directly compared after projection. The projection process is typically implemented using a cross-modal alignment network, which learns how to map feature vectors from different modalities into the same space, ensuring that semantically similar knowledge units are close together and semantically dissimilar knowledge units are far apart. Within the shared semantic space, semantic similarity between different knowledge units can be quantified using metrics such as Euclidean distance and cosine similarity. Based on these similarities and predefined conflict judgment rules, the probability of conflict between any two knowledge units can be further calculated, thus forming a cross-modal conflict probability matrix.

[0075] In this dynamic conflict graph generation process, knowledge unit pairs exceeding a conflict detection threshold are identified as conflict node pairs based on a cross-modal conflict probability matrix. A conflict edge is constructed for each conflict node pair. The initial weight of the conflict edge is determined based on the corresponding probability value in the cross-modal conflict probability matrix. A time-decay constraint is applied to the weight of the conflict edge based on the system entry timestamp of the knowledge units in the conflict node pair. Simultaneously, knowledge units exceeding the valid time range are marked as invalid based on the valid event timestamp. The conflict detection threshold is a preset value used to determine whether the similarity or conflict probability between two knowledge units reaches a level sufficient to constitute a conflict. When the conflict probability between two knowledge units exceeds this threshold, they are identified as a conflict node pair, and a conflict edge is established between them. The initial weight of the conflict edge directly comes from the corresponding probability value in the cross-modal conflict probability matrix, reflecting the initial intensity of the conflict.

[0076] By employing the aforementioned technical solutions, assigning both an event validity timestamp and a system entry timestamp as dual-temporal markers to each knowledge unit, the system can accurately distinguish between the actual validity period of knowledge and the time it entered the system. This prevents outdated knowledge from participating in conflict detection, ensuring the timeliness of knowledge. Simultaneously, by parsing associated compliance attributes from the compliance tag set and weighting the initial feature vector with credibility, a modal feature vector with compliance constraints is generated. This ensures that the feature representation of knowledge fully reflects its compliance reliability, thereby improving the accuracy of conflict detection. Furthermore, projecting the modal feature vector carrying dual-temporal markers and compliance constraints onto the shared semantic space and calculating cross-modal semantic similarity unifies the knowledge representation of different modalities, more effectively identifying implicit conflicts. Conflict node pairs are identified based on the conflict probability matrix, and conflict edges are constructed. A time-decaying constraint is applied to the weights of the conflict edges based on the system entry timestamp, avoiding the continued influence of old conflicts and enabling the conflict graph to dynamically reflect the latest state of knowledge.

[0077] In some of the embodiments described above in this application, it is proposed to parse the associated compliance attributes from the compliance tag set and weight the initial feature vector with credibility based on them to generate a modal feature vector with compliance constraints. However, in this process, the mapping and weighting of compliance attributes lack systematic processing, which results in the failure to reasonably quantify and integrate multi-dimensional information such as authorization status, source credibility and timeliness. This makes the feature vector unable to accurately reflect the comprehensive credibility of the knowledge unit, thereby affecting the accuracy and reliability of the subsequent construction of the dynamic conflict graph.

[0078] To address this, this application further proposes a method for parsing the associated compliance attributes of each knowledge unit from a set of compliance tags, and then weighting the initial feature vector based on the credibility of these associated compliance attributes to generate a modal feature vector with compliance constraints. Specifically, this method includes: extracting the associated compliance attributes of each knowledge unit from the set of compliance tags, whereby the associated compliance attributes include authorization status identifier, source credibility identifier, and timeliness status identifier. The authorization status identifier is mapped to an authorization credibility score, the source credibility identifier to a source authority score, and the timeliness status identifier to a timeliness freshness score. The authorization credibility score, source authority score, and timeliness freshness score are weighted and summed based on a preset weight fusion rule to obtain the initial node weight for each knowledge unit. The initial node weights are then used to adjust the weights of each dimension of the initial feature vector to generate a modal feature vector with compliance constraints.

[0079] The system extracts associated compliance attributes for each knowledge unit from the compliance tag set. These attributes include authorization status identifiers, source trust identifiers, and expiration status identifiers, aiming to obtain compliance metadata related to each knowledge unit. These attributes are key information for assessing the credibility and validity of knowledge units. In practical implementation, several methods can be used: One method is to organize the compliance tag set as structured data, such as key-value pairs or database tables, where the unique identifier of each knowledge unit serves as the key, and its corresponding associated compliance attribute serves as the value. The system can directly extract the authorization status identifier, source trust identifier, and expiration status identifier by querying this identifier. Another method is, if the compliance tag set exists in semi-structured or unstructured text form, the system can utilize Natural Language Processing (NLP) technology or predefined parsing rules to identify and extract the authorization status identifier, source trust identifier, and expiration status identifier from the text.

[0080] Mapping authorization status identifiers to authorization credibility scores, source credibility identifiers to source authority scores, and timeliness status identifiers to timeliness freshness scores aims to convert qualitative or categorical compliance identifiers into quantifiable numerical scores. This numerical representation allows for mathematical operations on different compliance dimensions, integrating them into a unified credibility metric. In practical implementation, the following methods can be used: One method is to predefine a mapping table, mapping specific authorization status identifiers (e.g., "authorized," "authorized expired," "unauthorized") to specific authorization credibility scores (e.g., 1.0, 0.5, 0.0). Source credibility identifiers (e.g., "officially released," "third-party report," "user-generated") and timeliness status identifiers (e.g., "latest," "expired," "soon to expire") are also mapped to source authority scores and timeliness freshness scores in a similar manner. Another method is to establish a rule engine that dynamically calculates the corresponding scores based on the complex combination logic of authorization status identifiers, source credibility identifiers, and timeliness status identifiers. For example, knowledge that is "officially released" and "latest" will have higher source authority scores and timeliness freshness scores.

[0081] The weighted sum of the authorization credibility score, source authority score, and timeliness freshness score, based on a preset weight fusion rule, yields the initial node weight for each knowledge unit. This aims to aggregate individual compliance scores into a single initial node weight. The preset weight fusion rule allows for flexible adjustment of the priority of different compliance aspects according to actual application needs. In practical implementation, the following methods can be used: One method is to preset a fixed set of weights (e.g., authorization credibility score weight w1, source authority score weight w2, and timeliness freshness score weight w3), and then calculate the initial node weight using the formula: Initial Node Weight = w1 × Authorization Credibility Score + w2 × Source Authority Score + w3 × Timeliness Freshness Score. These weights can be determined based on domain expert experience or historical data analysis. Another method is to dynamically adjust the weight fusion rule based on the current dialogue context features or user cognitive state vector. For example, in urgent task scenarios, the weight of the timeliness freshness score can be increased. In scenarios requiring high authority, the weight of the source authority score can be increased. This can be achieved using a small neural network or decision tree model, which takes dialogue context features and the user cognitive state vector as input and outputs dynamically adjusted weights.

[0082] This approach leverages initial node weights to weight and adjust the dimensions of the initial feature vector, generating a modal feature vector with compliance constraints. The aim is to directly integrate the calculated initial node weights into the initial feature vector of the knowledge unit. By adjusting the dimensions of the feature vector according to these weights, the semantic representation of the knowledge unit is "constrained" by its compliance status, making it more representative of its credibility within the shared semantic space. In practical implementation, the following methods can be used: One approach is to treat the initial node weights as a scalar and multiply them by each dimension of the initial feature vector. For example, the modal feature vector with compliance constraints = initial node weights × initial feature vector. This method scales the overall magnitude of the feature vector, making its "influence" in the semantic space proportional to its credibility. Another approach is to further decompose or map the initial node weights into weighting factors for different dimensions of the initial feature vector. For example, if some dimensions of the initial feature vector represent the "factuality" or "authority" of the knowledge, then the source authority score portion of the initial node weights can exert a greater influence on these dimensions. This can be achieved through a small attention network or gating mechanism that dynamically generates dimensional weighting coefficients based on the initial node weights.

[0083] The above technical solution extracts the associated compliance attributes corresponding to each knowledge unit from the compliance tag set, including authorization status identifier, source credibility identifier, and timeliness status identifier. This step clearly covers key dimensions such as authorization compliance, source reliability, and timeliness validity, providing a comprehensive foundation for subsequent quantification and avoiding credibility bias caused by attribute omissions. The authorization status identifier is mapped to an authorization credibility score, the source credibility identifier to a source authority score, and the timeliness status identifier to a timeliness freshness score. This numerical conversion transforms abstract compliance identifiers into calculable scores, solving the deficiency that attributes cannot be directly used in mathematical models and enabling consistent processing of compliance information across different dimensions. Based on preset weight fusion rules, the authorization credibility score, source authority score, and timeliness freshness score are weighted and summed to obtain the initial node weight for each knowledge unit. This step dynamically adjusts the weight ratio of each score according to the rules, achieving a balanced fusion of multi-dimensional information and ensuring that the weights reflect the relative importance of compliance attributes, rather than a single dominant factor. By using the initial weights of nodes to adjust the dimensions of the initial feature vector, a modal feature vector with compliance constraints is generated. By directly adjusting the dimensions of the feature vector through weights, the features are accurately embedded with credibility constraints in the semantic space, thereby more realistically representing the semantic and compliance status of knowledge units in subsequent conflict detection.

[0084] When the aforementioned technical solution is used in conjunction with the solution based on the standardized knowledge set and compliance tag set, which uses a cross-modal conflict graph builder to perform unified cross-modal semantic mapping and conflict identification, constructing a dynamic conflict graph with dual temporal labels, its technical effect is even more significant. By directly integrating compliance constraints into the modal feature vector, the similarity calculation results between different knowledge units in the shared semantic space will naturally consider the compliance and credibility of the knowledge units. This enables the cross-modal conflict graph builder to more accurately identify potential conflicts and avoid misjudgments caused by untrustworthy or non-compliant knowledge. In addition, the knowledge node weights of this dynamic conflict graph are constrained by the compliance tag set. The initial node weights obtained through this solution are directly used to adjust the feature vector, making this constraint more refined and quantifiable. This ensures that the weight of each knowledge node in the graph can truly reflect its multi-dimensional compliance status, thereby providing a more reliable decision-making basis for the subsequent dynamic arbitration engine.

[0085] In some of the solutions mentioned above in this application, a modal feature vector with compliance constraints carrying dual temporal tags is proposed to be projected into a shared semantic space and cross-modal semantic similarity between different knowledge units is calculated to construct a dynamic conflict graph. However, in this process, the timeliness difference of knowledge units (such as the difference of system input timestamps) is not considered when calculating similarity, and the problem of whether the effective timestamps of events overlap is not handled. This may lead to an inaccurate conflict probability matrix, which cannot accurately reflect the potential conflicts between knowledge units, thereby affecting the quality of the dynamic conflict graph and the reliability of the subsequent arbitration process.

[0086] To address this, this application further proposes a method that projects a modal feature vector with compliance constraints and dual temporal tags onto a shared semantic space, and calculates the cross-modal semantic similarity between different knowledge units in the shared semantic space to obtain a cross-modal conflict probability matrix. The method specifically includes the following steps: Modal feature vectors with compliance constraints and dual temporal labels are input into a cross-modal alignment network and projected onto a shared semantic space to obtain the semantic coordinate position of each knowledge unit in the shared semantic space. The cross-modal alignment network is a deep learning model designed to learn a unified representation of different modal data (such as text, images, and audio) in a shared semantic space. Its role is to map feature vectors from different modalities to a common low-dimensional vector space, such that semantically related data points from different modalities are close to each other, while semantically unrelated points are far apart. Possible implementations include, but are not limited to: a multimodal encoder based on the Transformer architecture, fusing information from different modalities through self-attention and cross-attention mechanisms; or a contrastive learning framework, learning alignment by maximizing the similarity of positive sample pairs (semantically consistent data from different modalities) in the shared space while minimizing the similarity of negative sample pairs (semantically inconsistent data from different modalities). The shared semantic space is an abstract vector space where knowledge units from different modalities are represented as semantically meaningful vectors, i.e., semantic coordinate positions. In this space, the distance or angle between vectors can reflect their semantic similarity or correlation. This space can be formed by the embedding vector space output by a single layer of a neural network. Alternatively, it can be a low-dimensional space mapped from high-dimensional features using dimensionality reduction techniques such as Principal Component Analysis (PCA) or t-SNE. Semantic coordinate position refers to the vector representation of a knowledge unit in the shared semantic space. Each knowledge unit, regardless of its original modality, is transformed into a point or vector in this space. Its function is to quantify the semantic information of the knowledge unit and serve as the basis for subsequent calculations of the similarity between knowledge units. It can be represented by a fixed-dimensional floating-point vector directly generated by the output layer of a cross-modal alignment network, or by embedding vectors obtained through multi-layer nonlinear transformations and dimensionality reduction of the original modality features.

[0087] The initial semantic similarity matrix is ​​obtained by calculating the Euclidean distance or cosine similarity between any two knowledge units based on their semantic coordinate positions. Euclidean distance is a commonly used metric to measure the "straight-line distance" between two points in a multidimensional space. In semantic space, the smaller the Euclidean distance, the closer the semantics of the two knowledge units. Its purpose is to provide an intuitive, spatial distance-based measure of semantic similarity, which can be achieved by calculating the square root of the sum of the squares of the differences in the corresponding dimensions of two vectors, or by batch calculating the Euclidean distance between all pairs of knowledge units through matrix operations. Cosine similarity is a metric to measure the consistency of the directions of two vectors, with values ​​between -1 and 1. In semantic space, the higher the cosine similarity, the more consistent the semantic directions of the two knowledge units, i.e., the more semantically similar they are. Its purpose is to provide a similarity metric that focuses only on semantic direction without considering vector magnitude (i.e., semantic strength). It can be achieved by calculating the dot product of two vectors divided by the product of their respective magnitudes, or by calculating the dot product after vector normalization. The initial semantic similarity matrix is ​​a symmetric matrix, where each element represents the initial semantic similarity value of a pair of knowledge units in a shared semantic space. Its function is to initially quantify the semantic correlation between all knowledge unit pairs, providing basic data for subsequent time-factor correction. It can be an N×N matrix, where N is the number of knowledge units, and the (i,j) elements of the matrix store the similarity between knowledge unit i and knowledge unit j. Alternatively, it can be stored in the form of a sparse matrix, recording only knowledge unit pairs with similarity higher than a certain threshold.

[0088] Based on the system entry timestamp corresponding to each knowledge unit, a timeliness difference coefficient is calculated between each pair of knowledge units. This timeliness difference coefficient is then used to attenuate and correct the corresponding similarity values ​​in the initial semantic similarity matrix. The system entry timestamp refers to the precise time record of when a knowledge unit is received, processed, and stored by the system. Its function is to reflect the "newness" of the knowledge unit within the system, i.e., its information freshness, which can be represented by the creation time recorded in the database, or the file creation or modification time in the file system. The timeliness difference coefficient is a numerical value used to quantify the impact of the difference in system entry timestamps between a pair of knowledge units on semantic similarity. Its function is to dynamically adjust the semantic similarity between knowledge units based on their "age" difference, so that even if knowledge units with large time differences are semantically similar, their similarity will be appropriately attenuated. Possible implementation methods include, but are not limited to: an exponential decay function based on the difference in system entry timestamps of two knowledge units, such as exp(-k×|t1-t2|), where k is the decay constant; or a piecewise linear function, dividing the coefficient into different intervals based on the time difference. Attenuation correction refers to reducing the overestimation of similarity due to significant time differences by multiplying the initial semantic similarity value by a time-sensitivity difference coefficient. Its purpose is to make similarity calculations more realistic and avoid over-reliance on outdated but semantically similar knowledge. This can be achieved by directly multiplying the initial similarity value by the time-sensitivity difference coefficient, or by using more complex fusion functions such as weighted averaging or nonlinear combinations.

[0089] The attenuated similarity value is converted into a conflict probability value. Based on the effective timestamps of each pair of knowledge units, it is determined whether there is overlap in the effective timestamps. If there is no overlap, the corresponding conflict probability value is increased, generating a cross-modal conflict probability matrix. The conflict probability value is a quantitative indicator measuring the possibility of conflict between a pair of knowledge units. Its function is to transform semantic similarity (after time correction) into a basis for conflict detection; the higher the value, the greater the possibility of conflict. Possible implementation methods include, but are not limited to: mapping the corrected similarity value to a probability range of 0 to 1 using the Sigmoid function; or converting the similarity value into a binary conflict indicator (0 or 1) using a preset threshold function, and further refining it into a probability. The effective timestamp of an event refers to the actual occurrence or effective time range of the event described by the knowledge unit (e.g., a time period or a specific point in time). Its function is to identify the timeliness boundary of the information carried by the knowledge unit. Unlike the system-entered timestamp, which focuses on "when the system knows about" the information, the effective timestamp of an event focuses on "when the information is real." Possible implementation methods include, but are not limited to: a time interval (start time - end time); or a single point in time representing the time when the event occurred. The cross-modal conflict probability matrix is ​​a symmetric matrix where each element represents the probability of a cross-modal conflict between a pair of knowledge units. This matrix is ​​a key input for constructing a dynamic conflict graph, comprehensively considering semantic similarity, the timeliness of system-entered timestamps, and the overlap of valid event timestamps. Its purpose is to comprehensively and accurately reflect the potential conflict relationships between knowledge units. Possible implementations include, but are not limited to: an N×N matrix, where N is the number of knowledge units, and the (i,j) elements of the matrix store the conflict probability between knowledge unit i and knowledge unit j; or storing it as a sparse matrix, only recording knowledge unit pairs with conflict probabilities higher than a certain threshold.

[0090] The above technical solution involves inputting modal feature vectors with compliance constraints and dual temporal tags into a cross-modal alignment network and projecting them onto a shared semantic space to obtain the semantic coordinates of each knowledge unit. This step ensures the alignment of features from different modalities within a unified semantic space, providing a basic framework for subsequent similarity calculations and avoiding errors caused by cross-modal semantic inconsistencies. Based on the semantic coordinates, the Euclidean distance or cosine similarity between any two knowledge units is calculated to obtain an initial semantic similarity matrix. This step establishes a basic metric for the semantic relationships between knowledge units. Then, based on the system entry timestamp corresponding to each knowledge unit, a timeliness difference coefficient is calculated between each pair of knowledge units. This coefficient is then used to attenuate and correct the corresponding similarity values ​​in the initial semantic similarity matrix. This step considers the differences in the novelty of knowledge and applies attenuation constraints to similarity through the timeliness difference coefficient, solving the problem of overestimation of similarity caused by time lag and improving the timeliness sensitivity of similarity calculations. The similarity value after attenuation correction is converted into a conflict probability value. Then, it is determined whether there is overlap based on the effective timestamp of each pair of knowledge units. If there is no overlap, the corresponding conflict probability value is increased to generate a cross-modal conflict probability matrix. This step handles the potential for conflict omissions caused by non-overlapping effective time ranges of events. By increasing the conflict probability value, the conflict identification when the time ranges do not match is strengthened, ensuring that the conflict probability matrix fully reflects the actual conflict relationship.

[0091] In some of the embodiments described above in this application, a dynamic arbitration engine is proposed to encapsulate the arbitration strategy into a visual decision path graph to provide a basis for decision-making. However, in its implementation, there is a lack of detailed feature extraction of the conflict graph, dynamic strategy generation, intermediate state recording, and effective visual encapsulation, which leads to an opaque decision-making process, difficulty for users to understand, and low intervention efficiency.

[0092] To address this, this application further proposes a method based on dynamic conflict graphs, dialogue context features, user cognitive state vectors, and compliance tag sets. This method encapsulates the decision logic, knowledge source weights, and conflict resolution paths corresponding to arbitration strategies into a visualized decision path graph using a dynamic arbitration engine. (See [link to relevant documentation]). Figure 4 The method includes: 401. Using the dynamic arbitration engine, the conflict distribution features of the dynamic conflict map are extracted to obtain the conflict type distribution vector, conflict intensity distribution vector, and conflict node density distribution vector.

[0093] The dynamic arbitration engine is the core processing unit of this method, responsible for receiving conflict information, generating arbitration strategies, and coordinating knowledge sources. Its role is to analyze, decide on, and visualize complex knowledge conflicts to support the decision-making process of human-computer dialogue systems when dealing with complex problems. The dynamic conflict graph is the core data structure used in this method to represent knowledge conflicts and their evolution. It organizes knowledge units and their conflict relationships in the form of a graph, where knowledge units are nodes and conflict relationships are edges. The "dynamic" nature of this graph is reflected in the fact that the attributes of its nodes and edges (such as weight and timeliness) are updated and adjusted in real time with factors such as time, user interaction, and compliance tags. The purpose of conflict distribution feature extraction is to quantify and abstract key conflict patterns and characteristics from the complex dynamic conflict graph, providing structured input for subsequent arbitration strategy generation. This process analyzes the graph's topology, node attributes, and edge attributes, transforming discrete conflict information into continuous or discrete feature vectors. The conflict type distribution vector is used to quantify the frequency or proportion of different types of conflicts in the dynamic conflict graph. Its function is to identify the main contradictions in the current conflict scenario, such as inconsistencies in factual information, differences in sentiment, or issues with the timeliness of information. This vector can be a multi-dimensional vector, with each dimension corresponding to a predefined conflict type, and its value representing the number or proportion of conflicts of that type. The conflict intensity distribution vector is used to describe the distribution of conflict severity in the dynamic conflict graph. Its function is to assess the overall risk level of the current conflict and guide the strength and priority of arbitration strategies. This vector can be represented as the distribution of the number of conflict edges within different intensity intervals, or it can summarize the overall characteristics of conflict intensity through statistical methods (such as mean, variance, median). The conflict node density distribution vector is used to reflect the density of conflicts around knowledge nodes in the dynamic conflict graph. Its function is to identify "conflict hotspots" in the graph, i.e., knowledge units that involve a large number of conflicts or are located at the core of the conflict. This vector can be obtained by calculating the number of conflict edges or conflict nodes in the local neighborhood of each knowledge node, and then performing statistical analysis or dimensionality reduction on these local density values. For example, the degree centrality or betweenness centrality of each node can be calculated, and these centrality indicators can be statistically distributed.

[0094] 402. Through the dynamic arbitration engine, the conflict type distribution vector, the conflict intensity distribution vector, the conflict node density distribution vector, the dialogue context features, the user cognitive state vector, and the compliance tag set input strategy orchestration model are used to generate an arbitration strategy sequence for the current conflict graph. The arbitration strategy sequence includes multiple strategy operators and their calling order.

[0095] The strategy orchestration model is the core intelligent module in this method, used to dynamically generate arbitration strategy sequences based on multi-dimensional inputs. Its role is to transform abstract conflict characteristics, dialogue contexts, user states, and compliance requirements into concrete, executable arbitration steps. This model can be implemented using machine learning techniques. The arbitration strategy sequence is a series of ordered arbitration operation instructions output by the strategy orchestration model. Its role is to guide the dynamic arbitration engine on how to progressively process and resolve conflicts. This sequence consists of multiple independent "strategy operators," and the execution order of these operators is clearly defined. A strategy operator is the basic operational unit constituting the arbitration strategy sequence; each operator represents a specific arbitration action or decision logic. Its role is to perform specific knowledge source weight adjustments, conflict resolution, or information presentation operations. The invocation order refers to the order in which the various strategy operators in the arbitration strategy sequence are executed. Its role is to ensure the logicality and effectiveness of the arbitration process, as the execution effects of different operators may be interdependent or influence each other.

[0096] 403. Through this dynamic arbitration engine, based on the calling order and parameter configuration of each strategy operator in the arbitration strategy sequence, the process of adjusting the weight of the knowledge source by each strategy operator and the process of selecting the conflict resolution path are simulated, and intermediate state records of each step are generated.

[0097] The simulation aims to pre-analyze the potential impact of the strategy operators before their actual execution, thereby assessing their effectiveness and generating detailed decision-making process records. Its purpose is to improve the transparency and interpretability of decision-making, enabling users to understand the arbitration engine's thought process. Simulations can be performed by running strategy operators in a virtual environment or by using pre-defined causal models to predict their impact on knowledge source weights and conflict graph states. The knowledge source weight adjustment process refers to dynamically changing the credibility or importance of different knowledge sources during arbitration, according to the instructions of the strategy operators. Its purpose is to prioritize more reliable, authoritative, or context-appropriate knowledge when conflicts exist. The conflict resolution path selection process refers to determining how to handle and resolve identified knowledge conflicts during arbitration, according to the instructions of the strategy operators. Its purpose is to provide users with clear conflict solutions or explanations. Intermediate state records are detailed logs of each decision and state change during the simulation. Their purpose is to comprehensively track the decision-making evolution of the arbitration engine, providing foundational data for visualization and user intervention. This record may include information such as the currently executed policy operator, the knowledge source weights after the policy operator is executed, local changes in the conflict graph, and the conflict resolution path selected in the current step and the corresponding decision basis.

[0098] 404. Through the dynamic arbitration engine, the arbitration strategy sequence, the intermediate state record, the determined knowledge source weight allocation scheme and conflict resolution path are visualized and encapsulated according to the graph structure to generate a visualized decision path graph containing nodes, edges and decision labels. The nodes in the visualized decision path graph represent decision states or intermediate results, the edges represent strategy operators or decision flows, and the decision labels are used to mark the decision basis for each step.

[0099] The purpose of visualization encapsulation is to present the complex arbitration decision-making process to users in an intuitive and easy-to-understand graphical way. Its function is to enhance the system's transparency, interpretability, and user intervention. The encapsulation process involves mapping abstract strategy sequences, intermediate states, and results to graphical elements (such as nodes, edges, and labels). A graph structure is a non-linear data organization form composed of nodes (or vertices) and edges connecting nodes. Its function is to naturally represent the state transitions and logical flow in the decision-making process, allowing users to clearly see each step from the initial conflict to the solution. The visualized decision path diagram is the graphical interface presented to users in this method. Its function is to enable users to intuitively understand how the arbitration engine draws conclusions from multi-source conflicts and to provide an entry point for user intervention. This diagram can contain nodes and edges of different colors, shapes, or sizes to distinguish different decision states, strategy operators, or conflict types. Nodes in the visualized decision path diagram represent key states or intermediate results in the decision-making process. Their function is to mark important milestones or data snapshots in the decision-making process. Edges in the visualized decision path diagram represent strategy operators or decision flows connecting different nodes. Its purpose is to demonstrate the dynamic evolution and logical relationships of the decision-making process. Decision labels are text descriptions attached to nodes or edges of the visualized decision-making path diagram. Their purpose is to provide detailed justification or explanation for each step of the decision, helping users understand why a particular decision was made.

[0100] Through the aforementioned technical solution, a dynamic arbitration engine extracts conflict distribution features from the dynamic conflict graph, obtaining conflict type distribution vectors, conflict intensity distribution vectors, and conflict node density distribution vectors. This step comprehensively captures the multi-dimensional features of the conflict graph, providing fundamental data support for subsequent strategy generation and avoiding the lack of decision-making basis due to insufficient feature extraction. The dynamic arbitration engine inputs the conflict type distribution vector, conflict intensity distribution vector, conflict node density distribution vector, dialogue context features, user cognitive state vector, and compliance tag set into the strategy orchestration model to generate an arbitration strategy sequence for the current conflict graph. The arbitration strategy sequence includes multiple strategy operators and their invocation order. This process integrates multi-source features and user state, dynamically generating the strategy sequence, ensuring the adaptability and real-time nature of the strategy, and solving the problem of non-dynamic strategy generation. Then, the dynamic arbitration engine simulates the adjustment process of knowledge source weights and the selection process of conflict resolution paths for each strategy operator according to the invocation order and parameter configuration of each strategy operator in the arbitration strategy sequence, generating intermediate state records for each step. This simulation and recording mechanism tracks the intermediate steps of the decision-making process in detail, enhancing process transparency and facilitating user understanding of the adjustment logic.

[0101] In some of the solutions mentioned above in this application, feature extraction of dynamic conflict graphs is proposed to support arbitration decisions. However, in this process, due to incomplete or unquantified conflict feature extraction, the generation of arbitration strategies is not accurate and efficient enough, and the detailed distribution of the conflict graph cannot be effectively captured, thereby affecting the decision accuracy and efficiency of the dynamic arbitration engine.

[0102] To address this, this application further proposes extracting conflict distribution features from the dynamic conflict map to obtain conflict type distribution vectors, conflict intensity distribution vectors, and conflict node density distribution vectors, including: Traverse all conflict edges in the dynamic conflict graph, and perform statistical counting based on the conflict type label carried by each conflict edge to generate a conflict type distribution vector. This conflict type distribution vector includes the number of factual conflicts, the number of emotional conflicts, and the number of time-sensitive conflicts.

[0103] Obtain the edge weight corresponding to each conflict edge as the conflict intensity value, divide all conflict intensity values ​​into intervals or calculate statistics to generate a conflict intensity distribution vector. This conflict intensity distribution vector contains the distribution of the number of conflict edges or the mean and variance of the intensity within different intensity intervals.

[0104] The number of conflicting edges or conflicting nodes within a preset neighborhood is calculated for each knowledge node as the center, and the local conflict density value of each knowledge node is obtained. All local conflict density values ​​are statistically analyzed or dimensionality reduced to generate a conflict node density distribution vector.

[0105] Conflict distribution feature extraction aims to quantify and structure the intrinsic attributes and spatial distribution patterns of conflicts from complex dynamic conflict graphs. Its role is to transform unstructured graph information into computable and analyzable numerical vectors, providing refined input for subsequent arbitration strategy formulation. Possible implementation methods include, but are not limited to: using graph neural networks (GNNs) to learn features from the graph and extract high-dimensional conflict representations; or employing traditional graph theory algorithms, such as centrality measures and clustering coefficients, to quantify the local and global characteristics of conflicts.

[0106] The conflict type distribution vector represents the quantity distribution of different types of conflicts in a dynamic conflict graph. This vector is generated by traversing all conflict edges in the dynamic conflict graph and statistically counting them based on the pre-assigned conflict type labels (e.g., factual conflict, emotional conflict, time-sensitive conflict, etc.) of each edge. This vector clearly shows the proportion of each type of conflict in the current graph. For example, the content of knowledge units can be analyzed using predefined rule matching or a classification model based on Natural Language Processing (NLP) to automatically assign type labels to conflict edges. Alternatively, when knowledge units are added to the database, a manual or semi-automatic annotation system can perform preliminary conflict type prediction and labeling.

[0107] The conflict intensity distribution vector describes the distribution of conflict severity in a dynamic conflict map. This vector is generated by obtaining the edge weight corresponding to each conflict edge as a conflict intensity value, and then dividing these intensity values ​​into intervals or calculating statistics. For example, conflict intensity values ​​can be divided into preset intensity intervals such as "low," "medium," and "high," and the number of conflict edges within each interval can be counted. Alternatively, statistics such as the mean, variance, and median of all conflict intensity values ​​can be calculated to reflect the central tendency and dispersion of the overall conflict intensity. This statistical information helps the arbitration engine assess the urgency and priority of conflicts.

[0108] The conflict node density distribution vector is used to reveal regions of concentrated conflict in a dynamic conflict graph. It is generated by calculating the number of conflicting edges or nodes within a predefined neighborhood of each knowledge node, thus obtaining the local conflict density value for each knowledge node. This vector is formed by statistically analyzing or reducing the dimensionality of all local conflict density values. For example, the predefined neighborhood can be k-hop neighbor nodes, i.e., all nodes within k steps of the center node. Alternatively, it can be defined as nodes within a specific semantic distance threshold. Statistical analysis of local conflict density values ​​can be performed using histogram distribution, while dimensionality reduction can utilize methods such as Principal Component Analysis (PCA) or t-distributed random neighborhood embedding (t-SNE) to map high-dimensional density information to a low-dimensional space for easier analysis and visualization.

[0109] Through the aforementioned technical solutions, the introduction of conflict type distribution vectors enables the arbitration engine to identify and differentiate conflicts of different natures, thereby formulating targeted and differentiated arbitration strategies. For example, it can trace the source of evidence for factual conflicts, provide emotional reassurance for emotional conflicts, and update information for time-sensitive conflicts, thus improving the accuracy of arbitration. The generation of conflict intensity distribution vectors allows the arbitration engine to assess the severity of conflicts, prioritize high-intensity conflicts, optimize the allocation of arbitration resources, and improve arbitration efficiency. Furthermore, the construction of conflict node density distribution vectors helps the arbitration engine locate "hotspot" areas where conflicts are concentrated, guiding arbitration resources to focus on these high-density conflict areas, thereby avoiding ineffective global searches and further improving the efficiency and accuracy of conflict resolution. These multi-dimensional conflict characteristics collectively provide the dynamic arbitration engine with accurate and comprehensive input, enabling it to generate more intelligent, efficient, and interpretable arbitration strategies, enhancing the human-computer dialogue system's ability to handle complex problems and improving the user experience.

[0110] In some of the solutions mentioned above in this application, conflict type distribution vector, conflict intensity distribution vector, conflict node density distribution vector, dialogue context features, user cognitive state vector, and compliance tag set are input into the policy orchestration model to generate arbitration policy sequences. However, in this process, how to effectively integrate the features from these different sources to generate dynamic and optimized policy sequences in order to improve the accuracy and adaptability of policy generation is a challenge.

[0111] To address this, this application further proposes concatenating the conflict type distribution vector, conflict intensity distribution vector, and conflict node density distribution vector to obtain a conflict graph feature vector. The conflict graph feature vector, dialogue context features, user cognitive state vector, and compliance tag set are then fused to generate a joint state representation. This joint state representation is input into a policy orchestration model, which employs a reinforcement learning network architecture to calculate the invocation probability of each candidate policy operator based on the joint state representation. Multiple policy operators are then sequentially selected from the policy operator library according to the invocation probabilities, and the execution order of the policy operators is determined to generate an arbitration policy sequence.

[0112] For example, concatenating the conflict type distribution vector, conflict intensity distribution vector, and conflict node density distribution vector yields a conflict graph feature vector. The conflict type distribution vector characterizes the nature or category distribution of conflicts in a dynamic conflict graph. For instance, it can statistically represent the number or proportion of factual, emotional, or time-sensitive conflicts, providing qualitative information to help the system understand the nature of the conflict. In practical applications, this can be obtained through pre-classification labeling of conflict edges or through semantic analysis and classification of conflict content using natural language processing techniques. The conflict intensity distribution vector quantifies the severity or impact distribution of conflicts in a dynamic conflict graph. For example, it can represent the number of conflict edges within different intensity ranges, or the mean and variance of conflict intensity, providing quantitative information to help the system assess the urgency and potential impact of the conflict. In practical applications, this can be obtained through statistical analysis of the weights of conflict edges or by using clustering algorithms to classify conflict intensity into different levels. The conflict node density distribution vector describes the local clustering of conflict nodes in a dynamic conflict graph. For example, it can represent the statistical distribution of the number of conflict edges or conflict nodes around each knowledge node, revealing conflict hotspots and helping the system identify key conflict points. In practical applications, it can be obtained by calculating the local centrality or clustering coefficient of each knowledge node, or by aggregating node neighborhood information through a graph neural network. The concatenation operation connects multiple independent feature vectors dimensionally to form a longer single vector containing all the original feature information. Its purpose is to integrate conflict information of different dimensions and granularities into a unified representation space, laying the foundation for subsequent feature fusion and avoiding information dispersion. For example, the conflict type distribution vector, conflict intensity distribution vector, and conflict node density distribution vector can be directly concatenated in sequence, or a simple fully connected layer can be used for preliminary feature combination. The conflict graph feature vector, obtained after the concatenation operation, comprehensively reflects the overall conflict state and structural characteristics of the dynamic conflict graph, providing a comprehensive and compact conflict situation awareness for the strategy orchestration model.

[0113] This paper fuses conflict graph feature vectors, dialogue context features, user cognitive state vectors, and compliance tag sets to generate a joint state representation. Dialogue context features capture the background information and historical interaction states of the current human-computer dialogue. For example, they can include previous user questions, system responses, dialogue rounds, and dialogue topics. Their role is to provide a dialogue context for strategy orchestration, making the generated strategies more aligned with the actual needs of the current dialogue. In practice, this can be obtained by encoding the dialogue history (e.g., using a Transformer model) or extracting key dialogue entities and intentions. User cognitive state vectors represent the current user's cognitive level, comprehension ability, emotional state, and task urgency. For example, they can include dimensions of user expertise and task urgency. Their role is to enable personalized adjustments to strategy orchestration based on individual user differences, improving user experience and decision-making efficiency. In practice, this can be constructed using user profiling, historical interaction data analysis, or real-time sentiment recognition models. The compliance tag set contains compliance information associated with a standardized knowledge set, such as the authorization status, credibility, and timeliness of the data source. Its role is to provide ethical and compliance constraints for strategy orchestration, ensuring that the generated decisions comply with preset compliance standards. In practical applications, this label set can be a binary vector representing whether various compliance checks have passed, or a multi-dimensional score vector quantifying various compliance attributes. Feature fusion is the process of effectively integrating feature information from different sources and modalities, aiming to generate a more expressive and comprehensive single representation. Its role is to combine the internal state of the conflict graph with the external environment (dialogue context, user cognition, compliance requirements) to form a complete understanding of the current decision-making scenario. In practical applications, this can be achieved by simply concatenating features and then adding a fully connected layer, attention mechanisms (such as multi-head attention), or gating mechanisms (such as gated recurrent units) to learn the complex interaction relationships between different features. The joint state representation, obtained after feature fusion, comprehensively reflects the internal state of the conflict graph, the external environment of the dialogue, the user's cognitive characteristics, and compliance requirements in the current human-computer dialogue scenario. Its role is to provide a high-dimensional, semantically rich input for the policy orchestration model, enabling it to make more accurate decisions that better meet actual needs.

[0114] The joint state representation is input into the policy orchestration model, which employs a reinforcement learning network architecture to calculate the invocation probability of each candidate policy operator based on the joint state representation. The policy orchestration model is the core decision-making unit of this method, responsible for generating the optimal arbitration policy sequence based on the current state. Its role is to automate and intelligently manage the complex decision-making process, achieving dynamic and adaptive policy generation. In practical applications, this model can be a deep learning-based sequence generation model, such as Transformer or recurrent neural networks. Reinforcement learning is a machine learning paradigm that learns optimal behavioral policies through interaction with the environment. Employing a reinforcement learning network architecture means that the policy orchestration model can learn how to generate the optimal arbitration policy through trial and error and reward mechanisms, continuously interacting with the environment. Its role is to enable the model to have self-learning and optimization capabilities, adapting to constantly changing conflict scenarios and user needs. Common reinforcement learning architectures include Actor-Critic models, Q-learning, or Proximal Policy Optimization (PPO). Candidate policy operators are the basic operational units that constitute the arbitration policy sequence. They represent various atomic decisions or actions that the system can execute, such as "adjusting knowledge source weights," "selecting conflict resolution paths," and "requesting user clarification." Their role is to decompose complex arbitration tasks into a series of manageable, discrete actions. The invocation probability is output by the policy orchestration model after evaluating each candidate policy operator based on the current joint state representation. It represents the likelihood of selecting that operator in the current state and guides the selection of policy operators, enabling the model to prioritize operators that are more likely to lead to good results. In practical applications, this is usually the result of the neural network output layer after passing through the softmax activation function.

[0115] Based on the invocation probabilities, multiple policy operators are sequentially selected from a policy operator library, and the execution order of these operators is determined to generate an arbitration policy sequence. The policy operator library stores all predefined or learnable candidate policy operators. Each operator encapsulates specific functions and parameters, providing a set of selectable actions for the policy orchestration model. This library can be a database or a collection of objects in memory. Sequential selection refers to picking policy operators one by one from the policy operator library according to a certain strategy (such as greedy selection, sampling, or beam search) based on the calculated invocation probabilities to construct an ordered sequence. Its purpose is to combine discrete policy operators into a coherent and executable arbitration process. For example, the operator with the highest probability can be selected each time, or random sampling can be performed based on the probability distribution. The execution order refers to the order in which the various policy operators are invoked in the arbitration policy sequence. Its purpose is to ensure the logicality and effectiveness of the arbitration process, as the execution of different operators may have dependencies or timing requirements. The execution order can be directly output by the reinforcement learning model or determined through preset rules or heuristic algorithms. The arbitration strategy sequence consists of multiple strategy operators and their determined execution order, representing a complete arbitration scheme for the current complex problem and conflict graph. Its role is to guide the dynamic arbitration engine in encapsulating decision logic, adjusting knowledge source weights, and selecting conflict resolution paths.

[0116] The aforementioned technical solution concatenates the conflict type distribution vector, conflict intensity distribution vector, and conflict node density distribution vector to form a comprehensive conflict graph feature vector. This avoids the dispersion of conflict information and provides centralized conflict situation awareness for subsequent decision-making. By fusing the conflict graph feature vector with dialogue context features, user cognitive state vectors, and compliance tag sets, a joint state representation is generated. This representation not only captures the internal state of the conflict graph but also incorporates the external environment, user personalized needs, and ethical compliance constraints, thus providing a comprehensive and rich decision-making basis for the policy orchestration model. The policy orchestration model adopts a reinforcement learning network architecture, which can dynamically calculate the invocation probability of each candidate policy operator based on this joint state representation. Through learning and optimization, it intelligently selects and determines the execution order of policy operators from the policy operator library, generating an arbitration policy sequence adapted to the current complex scenario. This dynamic policy generation mechanism based on reinforcement learning enables the system to flexibly adjust arbitration strategies according to real-time changes in conflict situations, user cognitive states, and compliance requirements. It avoids the limitations of traditional fixed rules or static models, greatly improving the intelligence, adaptability, and effectiveness of arbitration decisions. Thus, in human-computer dialogue scenarios, it can solve complex problems more accurately and compliantly, and provide users with a more interpretable and interventionist decision-making process.

[0117] In some of the embodiments described above in this application, a method for generating a visual decision path diagram using a dynamic arbitration engine is proposed to improve the transparency and operability of decision-making. However, in its implementation, there is a lack of step-by-step simulation of the execution process of the arbitration strategy sequence and detailed recording of intermediate states. This makes it impossible for users to intuitively understand the step-by-step evolution logic of the decision, and the system also has difficulty capturing the adjustment details and decision basis of each step in real time. This limits the traceability, interpretability, and dynamic optimization capability based on user feedback of the decision-making process.

[0118] To this end, this application further proposes to simulate the adjustment process of knowledge source weights and the selection process of conflict resolution paths for each strategy operator based on the calling order and parameter configuration of each strategy operator in the arbitration strategy sequence, generating intermediate state records for each step, specifically including: Based on the parameter configuration of the first strategy operator in the arbitration strategy sequence, the current knowledge source weights in the dynamic conflict graph are used as input to execute the first strategy operator, resulting in the adjusted first knowledge source weight set and the first intermediate conflict graph.

[0119] Based on the remaining conflict distribution in the first intermediate conflict map, the initial conflict resolution path selected after the execution of the first strategy operator is simulated, and the first decision basis corresponding to the initial conflict resolution path is recorded.

[0120] The first knowledge source weight set and the first intermediate conflict graph are used as inputs to the next strategy operator. The subsequent strategy operators in the arbitration strategy sequence are executed in sequence to obtain the adjusted knowledge source weight set and the corresponding intermediate conflict graph at each step.

[0121] Record the conflict resolution path selected at each step, the decision basis for each step, and the state change of the conflict graph after each step is executed. Encapsulate the records of all steps according to the execution sequence to generate this intermediate state record.

[0122] To better understand the above technical solution, the technical features involved are explained in detail below: Simulating the adjustment process of knowledge source weights and the selection process of conflict resolution paths for each policy operator refers to the system's pre-analysis of the potential effects of each policy operator under the current dynamic conflict graph state before the actual execution of the arbitration policy. The arbitration policy sequence consists of a series of policy operators, each designed to resolve a specific type of conflict or optimize knowledge source weights. This simulation process can be completed by a simulation module that receives the current state of the dynamic conflict graph (including knowledge source weights, conflict distribution, etc.) and the policy operators to be executed along with their parameters. It then calculates and outputs the new knowledge source weights and conflict resolution paths after execution based on the predefined logic of the operators. Alternatively, a rule-based reasoning simulator can be used. This simulator predicts the impact of policy operators on knowledge source weights and conflict resolution paths based on the type of policy operator (such as weight adjustment operator or conflict resolution operator) and its parameters, combined with a pre-defined knowledge source weight adjustment model or conflict resolution model.

[0123] Generating intermediate state records for each step refers to a detailed log of each decision and state change during the simulation. It captures key information before and after the execution of policy operators for subsequent analysis, backtracking, and visualization. These records can be stored in memory or a database in structured data formats (such as JSON or XML), linked with timestamps or step numbers, and include data such as the set of knowledge source weights after each simulation step, intermediate conflict graphs, selected conflict resolution paths, and corresponding decision-making criteria. Alternatively, linked lists or tree structures can be used to organize these intermediate state records, with each node representing a step, containing the state information of that step and a pointer to the next step, thus forming a traceable decision path.

[0124] Using the current knowledge source weights in the dynamic conflict graph as input means that when simulating the first policy operator, it's necessary to obtain the current system's evaluation status of each knowledge source to ensure the accuracy of the simulation. Knowledge source weights reflect the credibility, authority, or importance of different knowledge sources in solving complex problems. This input can be directly read from the knowledge node attributes of the dynamic conflict graph; each knowledge node may be associated with the weight information of its source. Alternatively, it can be obtained through a separate knowledge source weight management module. Alternatively, it can be obtained by querying a real-time updated knowledge source weight database or cache, which stores the weights of each knowledge source dynamically calculated by the system based on historical performance and compliance tag sets.

[0125] Obtaining the adjusted first knowledge source weight set and the first intermediate conflict graph refers to the direct result after the execution of the first policy operator, representing the first stage of the decision-making process. The adjusted knowledge source weight set can be a list of key-value pairs, where the key is the knowledge source identifier and the value is the adjusted weight. The first intermediate conflict graph is the graph structure after updating some knowledge node weights, conflict edge weights, or even removing some conflict edges, based on the original dynamic conflict graph and according to the effect of the first policy operator. Alternatively, the first intermediate conflict graph can also be realized by generating a graph snapshot, which contains the latest state of all knowledge nodes, conflict edges, and their attributes after the execution of the first policy operator.

[0126] Based on the distribution of remaining conflicts in the first intermediate conflict graph, simulating the initial conflict resolution path selected after the execution of the first policy operator refers to the types, intensities, and locations of conflicts that still exist in the dynamic conflict graph after the execution of the first policy operator. A conflict resolution path refers to a series of operations or decision steps taken to resolve these remaining conflicts. This can be accomplished using a conflict resolution path generator, which analyzes the conflict distribution in the first intermediate conflict graph and recommends or selects an optimal resolution path based on preset conflict resolution rules (e.g., prioritizing the resolution of high-intensity conflicts, prioritizing the resolution of factual conflicts, etc.) or based on a machine learning model (such as reinforcement learning). Alternatively, it can be achieved using heuristic algorithms, such as greedy algorithms or... The search algorithm seeks the shortest or optimal path from the current conflict state to the target conflict-free state in the first intermediate conflict graph.

[0127] Recording the primary decision criterion for the initial conflict resolution path explains why this particular path was chosen, providing transparency and interpretability to the decision-making process. This criterion can be the rule ID that triggered the path selection, the confidence level of the model's prediction, key dimensions in the user's cognitive state vector (e.g., a higher level of user expertise leads to a more aggressive resolution path), or specific keywords from the dialogue context. Alternatively, it can be the feature vector or decision tree path used by the policy orchestration model when selecting the path, stored in text or encoded form.

[0128] Executing subsequent policy operators in the arbitration policy sequence sequentially means that the policy operators are executed step by step in a predetermined order, with each step based on the result of the previous step. This can be achieved using a loop structure that iterates through each policy operator in the arbitration policy sequence, using the output of the previous operator as the input of the current operator. Alternatively, it can be achieved using an event-driven architecture, where the execution of the next policy operator is triggered after one policy operator has completed execution and generated an intermediate state.

[0129] Recording the conflict resolution path chosen at each step, the corresponding decision basis for each step, and the state changes of the conflict graph after each step is a comprehensive record of the entire simulation process, ensuring the integrity and traceability of the decision-making process. This information can be encapsulated in a data structure, such as a log object containing fields like "Step ID," "Strategy Operator Name," "Input Status," "Output Status," "Selected Path," and "Decision Basis," and then these log objects can be arranged chronologically. Alternatively, a version control system can be used to take snapshots of each state change in the conflict graph and associate them with the corresponding strategy operators, resolution paths, and decision basis.

[0130] Encapsulating all step records according to their execution sequence to generate this intermediate state record means organizing scattered records into a unified, ordered whole, facilitating subsequent processing and display. This can be achieved by storing all step records in an ordered list, where each element represents a complete record of a step. Alternatively, these records can be serialized into a single file or data stream, containing explicit step separators and timing information.

[0131] By employing the aforementioned technical solution, this application addresses the issues of opaque decision-making processes and the inability to optimize in real-time by simulating the execution of an arbitration strategy sequence and generating intermediate state records for each step. For instance, this solution makes complex decision-making processes highly transparent and explainable. Users are no longer merely passive recipients of results but can intuitively understand how the system makes decisions step by step, including the logic for adjusting knowledge source weights and the basis for selecting conflict resolution paths. This step-by-step simulation and detailed recording provide a complete audit trail capability for the decision-making process, making system behavior traceable and verifiable, thus greatly enhancing the system's credibility.

[0132] Furthermore, by generating intermediate state records for each step, including the adjusted set of knowledge source weights, intermediate conflict graphs, selected conflict resolution paths, and corresponding decision-making criteria, a refined entry point for user intervention is provided. Users can review and provide feedback on any intermediate step in the decision-making process, such as adjusting the weight of a knowledge source, modifying the conflict resolution path, or even injecting new knowledge. This fine-grained user feedback can be captured and parsed by the system in real time into a standardized intervention paradigm, thereby triggering the recalculation of the arbitration strategy and achieving dynamic optimization of the decision-making process. This not only enhances the system's interveneability but also enables the system to learn and evolve from every user interaction, continuously optimizing its arbitration strategy generation logic. Simultaneously, since each simulation is based on the real-time state of the current dynamic conflict graph, the contextual relevance and logical coherence of strategy adjustments are ensured, avoiding the discontinuity caused by leapfrog decisions, thus improving the accuracy and robustness of arbitration decisions.

[0133] In some of the solutions mentioned above in this application, a visualization encapsulation based on a graph structure is proposed to generate a decision path diagram. However, in this process, the lack of specific mapping rules and structured processing results in an unclear and unsystematic visualization of the decision process. Users find it difficult to intuitively track decision steps, understand the decision basis and conflict resolution path, thereby reducing interpretability and effective intervention capability.

[0134] In response, this application further proposes a method for visually encapsulating the arbitration strategy sequence, the intermediate state record, the determined knowledge source weight allocation scheme, and the conflict resolution path according to a graph structure, generating a visualized decision path graph containing nodes, edges, and decision labels, specifically including: Each strategy operator in the arbitration strategy sequence is mapped to a set of candidate edges in the decision path graph, and each intermediate state in the intermediate state record is mapped to an intermediate node in the decision path graph.

[0135] Based on the conflict resolution path selected at each step in the intermediate state record, the actual path edge is determined from the candidate edge, and a decision label is generated for the actual path edge according to the decision criteria corresponding to each step in the intermediate state record.

[0136] The determined knowledge source weight allocation scheme is mapped to the attribute information of the intermediate node or node, and the determined conflict resolution path is mapped to the complete path from the starting node to the ending node.

[0137] The intermediate node, the actual path edge, the decision label, and the attribute information are arranged and connected according to the execution sequence to generate the visual decision path diagram.

[0138] For example, mapping each strategy operator in the arbitration strategy sequence to a set of candidate edges in the decision path graph aims to concretize abstract strategy operations into visual elements. Strategy operators are the basic building blocks of arbitration strategies, representing specific decision operations or rules, such as adjusting knowledge source weights or selecting conflict resolution algorithms. Mapping them to candidate edges displays all possible decision branches in the visual path graph, allowing users to anticipate the possible outcomes of different strategy choices. For instance, a mapping table between strategy operator types and edge types can be predefined, mapping the "knowledge source weight adjustment" operator to an edge with a specific color or style, while mapping the "conflict resolution rule selection" operator to another edge with a different style. Alternatively, a set of candidate edges can be dynamically generated based on the function of the strategy operator or its potential impact range. For example, an operator may correspond to multiple parameter configurations, with each configuration corresponding to a candidate edge, representing different execution paths for the operator under different parameters.

[0139] Simultaneously, each intermediate state in the intermediate state record is mapped to an intermediate node in the decision path graph. The intermediate state records the system state at each step during the execution of the arbitration strategy sequence, such as changes in knowledge source weights and conflict graphs. Mapping it to intermediate nodes visualizes the abstract intermediate calculation results, allowing users to clearly track each step of the decision-making process. For example, key indicators from the intermediate state record (such as the current number of conflicts and a summary of the knowledge source weight distribution) can be extracted and displayed as tags or attributes on the nodes. Alternatively, the summary information of the conflict graph snapshot after each step (such as conflict type distribution and intensity) can be used as node content, allowing users to view further details by clicking on the node.

[0140] Based on the conflict resolution path selected at each step in the intermediate state record, the actual path edge is determined from the candidate edges. After the arbitration strategy sequence is executed, an actual conflict resolution path is generated. This step aims to highlight or select the actual decision path being executed from all possible candidate edges, thereby avoiding information redundancy and allowing the user to focus on the actual decision-making process. For example, the actual path edge can be highlighted by changing its color, thickness, or adding animation effects. Alternatively, unselected candidate edges can be weakened (e.g., reduced transparency, displayed as dashed lines) or hidden directly, displaying only the actual path.

[0141] Furthermore, decision labels are generated for each actual path edge based on the decision-making criteria corresponding to each step in the intermediate state record. The decision-making criteria explain why a particular path is chosen or a certain strategy operator is executed. Adding decision labels to actual path edges enhances the interpretability of the decision-making process, helping users understand the deep logic behind the system's specific choices. For example, a decision label could be a brief text description, such as "Based on the user's high level of expertise, an aggressive resolution strategy is chosen." Alternatively, a decision label could be a link to a detailed explanatory document or log, allowing users to view a more detailed decision-making reasoning process.

[0142] The determined knowledge source weighting scheme is mapped to the attribute information of the intermediate node or node. The knowledge source weighting scheme is one of the core outputs of the arbitration process, determining the importance of different knowledge sources in answer generation. Using it as a node attribute allows users to intuitively assess the contribution of each knowledge source. For example, weight values ​​can be displayed directly next to the node as numerical values ​​or percentages, or weight can be indicated by changes in node color or size. Alternatively, the weight distribution of each knowledge source can be displayed in a list or pie chart format within the node attribute panel.

[0143] Simultaneously, the determined conflict resolution path is mapped as a complete path from the starting node to the ending node. The complete conflict resolution path is an end-to-end view of the entire arbitration process, showing the complete evolution from the initial conflict state to the resolved state. For example, this can be represented by highlighting all actual path edges and intermediate nodes from the starting node to the ending node. Alternatively, a "path replay" function can be provided to dynamically display each step of the conflict resolution process.

[0144] The intermediate node, the actual path edge, the decision label, and the attribute information are arranged and connected according to the execution sequence to generate the visualized decision path diagram. Arranging and connecting according to the execution sequence ensures the logical coherence and readability of the visualized decision path diagram, enabling users to clearly understand the temporal order of the decision-making process. For example, a directed graph layout from left to right or from top to bottom can be used, with connecting lines between nodes representing the decision flow. Alternatively, a hierarchical layout algorithm can be used, placing nodes at the same time step in the same level, with edges connecting different levels.

[0145] The above technical solution maps each strategy operator in the arbitration strategy sequence to a set of candidate edges in the decision path graph, representing each decision step as a potential path option, allowing users to preview all possible decision branches. Simultaneously, each intermediate state in the intermediate state record is mapped to an intermediate node in the decision path graph, transforming abstract intermediate results into visual elements to help users track the decision progress in real time. Based on the conflict resolution path selected at each step in the intermediate state record, the actual path edge is determined from the candidate edges, ensuring that the visualization only reflects the actual execution path and avoiding redundant information interference. Decision labels are generated for the actual path edges based on the decision basis corresponding to each step in the intermediate state record, and explanatory labels are added to clarify the reasons for the selection, enhancing users' understanding of the decision logic. The determined knowledge source weight allocation scheme is mapped to intermediate nodes or node attribute information, intuitively displaying the importance weight of knowledge sources, allowing users to evaluate the weight allocation basis. The determined conflict resolution path is mapped to a complete path from the starting node to the ending node, providing an end-to-end decision view, facilitating users' grasp of the overall process. By arranging and connecting intermediate nodes, actual path edges, decision labels, and attribute information according to the execution sequence, all elements are arranged in chronological order to generate a logically coherent and visual decision path diagram, thereby improving the interpretability of the decision-making process and the efficiency of user intervention.

[0146] In some of the embodiments described above in this application, user intervention results are obtained based on a visualized decision path diagram to optimize decision-making and conflict resolution. However, in this process, user intervention is not standardized and not fed back in real time, which makes it impossible to dynamically adjust the arbitration strategy. User feedback is disconnected from the system decision-making process and cannot be systematically utilized, thereby reducing the system's real-time responsiveness and knowledge reuse efficiency.

[0147] To address this, this application further proposes a method for obtaining user intervention results based on a visualized decision path diagram, parsing these results into a standardized intervention paradigm, and transmitting them back in real time to trigger a recalculation of the arbitration strategy. See [link to relevant documentation]. Figure 5 It includes the following steps, 501. Through the human-computer interaction interface provided by the visualized decision path diagram, capture the operation trajectory performed by the user on the visualized decision path diagram. The operation trajectory includes node weight adjustment operation, path selection operation, and knowledge injection operation.

[0148] The human-computer interaction interface (HCI) is a graphical interface for users to interact with the system. It can be implemented using web front-end technologies (e.g., frameworks like React or Vue), rendering the graph structure using SVG or Canvas libraries (e.g., D3.js or ECharts), and integrating event listeners to capture user interactions. Alternatively, it can be implemented using desktop application frameworks (e.g., Qt or Electron), displaying the graph using graphics rendering components and communicating with the backend system via API interfaces. Capturing the user's operational trajectory on this visualized decision-making path graph aims to record the user's specific behaviors on the interface, serving as a basis for subsequent analysis and system adjustments. This can be achieved through front-end event listeners (such as click, drag, and input events), capturing user interactions on nodes and edges, and recording information such as operation type, target element ID, and modified values. Alternatively, a behavior log recording module can be used to serialize and record a series of user operations on the interface, such as mouse movements, keyboard input, and button clicks, forming a timestamped sequence of operations. This operational trajectory covers the specific intervention types that users can perform, including modifications to the decision-making process, decision results, and knowledge base. For example, the node weight adjustment operation allows users to modify the weight values ​​of visual nodes (representing knowledge sources or decision factors) by dragging sliders, entering values, or clicking add / delete buttons. The path selection operation allows users to select or modify edges (representing strategy operators or decision flows) in the decision path graph by clicking or dragging, such as selecting different conflict resolution paths. The knowledge injection operation allows users to add new knowledge content by entering it in a text box, uploading files, or selecting predefined templates; this content will be transformed into new knowledge nodes.

[0149] 502. Perform semantic parsing on the operation trajectory, extract the operation type, operation object, and operation parameters corresponding to each operation, and structurally encode the operation type, operation object, and operation parameters according to the preset paradigm template to generate the standardized intervention paradigm.

[0150] The purpose of semantic parsing of the operation trajectory is to transform raw, unstructured user actions into system-understandable instructions with clear meaning. This can be achieved by identifying the operation type, operation object (such as node ID, edge ID), and operation parameters (such as new weight values, new knowledge text) from the operation log based on predefined syntax rules and regular expressions. Alternatively, Natural Language Processing (NLP) techniques can be used to perform intent recognition and entity extraction on the free text description input by the user, mapping it to predefined operation types and parameters. Extracting the operation type, operation object, and operation parameters corresponding to each operation aims to clarify the core elements of user intervention, providing a foundation for subsequent structured coding. For example, if the operation trajectory is in JSON format {"type":"weight_adjust","target_node_id":"K1","new_weight":0.7}, these fields can be directly extracted. The operation type, operation object, and operation parameters are then structured and encoded according to a preset paradigm template to generate this standardized intervention paradigm, aiming to unify the format of user intervention data and make it easier for the system to process and store. This can be achieved by defining a JSON Schema or XML Schema as a paradigm template, filling the parsed operation information into the corresponding fields to form a standardized data structure. Alternatively, serialization frameworks such as Protocol Buffers or Thrift can be used to define cross-language data structures and encode the operation information into standardized messages in binary or text format.

[0151] 503. The standardized intervention paradigm is fed back to the dynamic arbitration engine in real time. The dynamic arbitration engine then adjusts the corresponding strategy operator parameters in the current arbitration strategy sequence online based on the operation object and operation parameters in the standardized intervention paradigm, triggering the recalculation of the arbitration strategy.

[0152] The purpose of transmitting this standardized intervention paradigm back to the dynamic arbitration engine in real time is to ensure that user intervention can immediately influence the arbitration decision-making process. Through this dynamic arbitration engine, based on the operation objects and parameters in the standardized intervention paradigm, the corresponding strategy operator parameters in the current arbitration strategy sequence are adjusted online, triggering the recalculation of the arbitration strategy. This aims to dynamically adjust the arbitration strategy to adapt to real-time user feedback.

[0153] 504. The standardized intervention paradigm is synchronously sent to the cross-modal conflict graph builder. The cross-modal conflict graph builder extracts new knowledge content based on the knowledge injection operation in the standardized intervention paradigm, converts the new knowledge content into new knowledge nodes, and updates the dynamic conflict graph.

[0154] The purpose of simultaneously sending the standardized intervention paradigm to the cross-modal conflict graph builder is to ensure that user modifications or injections of knowledge are promptly reflected in the conflict graph. Through this cross-modal conflict graph builder, new knowledge content is extracted based on the knowledge injection operations within the standardized intervention paradigm. This new knowledge content is then converted into new knowledge nodes and used to update the dynamic conflict graph. This allows users to directly add new information to the knowledge base, thereby enriching and refining the conflict graph.

[0155] The above technical solution captures operation trajectories, including node weight adjustment, path selection, and knowledge injection, through a human-computer interaction interface provided by a visualized decision path graph. This feature allows users to directly perform diverse interventions on the visualized interface, solving the problem of limited user feedback in traditional methods and making interventions more intuitive and comprehensive. Semantic parsing of the operation trajectory extracts the operation type, object, and parameters, and structures and encodes them into a standardized intervention paradigm. This transforms fuzzy user input into a unified format that can be processed by the machine, addressing the limitation that intervention results are difficult for the system to directly utilize, and ensuring the accuracy and efficiency of subsequent processing. The standardized intervention paradigm is fed back to the dynamic arbitration engine in real time. Based on the object and parameters of the operation, the strategy operator parameters are adjusted and recalculation is triggered. This feature updates the decision logic in real time based on the specific content of the intervention, solving the problem of static arbitration strategies and achieving dynamic optimization and immediate response in the decision-making process.

[0156] In some of the embodiments described above in this application, it is proposed to adjust the corresponding strategy operator parameters in the current arbitration strategy sequence online according to the operation object and operation parameters in the standardized intervention paradigm to trigger the recalculation of the arbitration strategy. However, in its implementation, there may be problems such as insufficient adjustment, inaccurate positioning, untimely updates or invalid recalculation, which leads to system response delay after user intervention, chaotic decision logic and failure of knowledge source weight allocation scheme, reducing the real-time performance and reliability of the dynamic arbitration engine.

[0157] To address this, this application further proposes an online adjustment method for the corresponding strategy operator parameters in the current arbitration strategy sequence based on the operational objects and parameters in the aforementioned standardized intervention paradigm, triggering a recalculation of the arbitration strategy. This method includes: parsing the standardized intervention paradigm to extract the operational objects, operational parameters, and operational type identifiers, whereby the operational type identifiers distinguish between weight adjustment, path selection, and knowledge injection. The method involves locating the target strategy operator in the current arbitration strategy sequence based on the operational object and updating the internal configuration parameters of the target strategy operator in real time according to the operational parameters. Based on the updated target strategy operator and other unchanged strategy operators in the current arbitration strategy sequence, a re-adjusted arbitration strategy sequence is generated. The adjusted arbitration strategy sequence is then input into a strategy orchestration model for recalculation, outputting updated decision logic and an updated knowledge source weight allocation scheme.

[0158] For example, parsing the standardized intervention paradigm extracts the operation object, operation parameters, and operation type identifier. The operation type identifier is used to distinguish between weight adjustment, path selection, and knowledge injection. This step aims to achieve a structured understanding of the intervention behavior performed by users through the human-computer interaction interface provided by the visual decision path diagram. The standardized intervention paradigm is a unified representation of user intervention behavior. By parsing it, the system can accurately identify the user's intent, including the object the user wants to operate on (such as a knowledge node or a decision path), the specific operation content (such as the adjusted weight value or the selected path branch), and the type of operation (whether it is adjusting existing weights, selecting a new path, or injecting new knowledge). For example, a rule-based parser can be used, predefining a set of grammar rules and keywords, to identify the operation type based on specific patterns contained in the paradigm (such as "weight: [value]", "select path: [path ID]", "inject knowledge: [text content]"), and extract the corresponding operation object and operation parameters. Alternatively, a machine learning-based semantic parsing model can be used to train a sequence-to-sequence (Seq2Seq) model or a Transformer-based model. The input is standardized intervention paradigm text, and the output is structured JSON or XML data, which includes explicit operation object fields, operation parameter fields, and operation type identifier fields, thereby handling more complex natural language expressions.

[0159] The system locates the target strategy operator in the current arbitration strategy sequence based on the operation object and updates its internal configuration parameters in real time according to the operation parameters. This step ensures that user intervention can accurately target specific aspects of the arbitration strategy sequence. The arbitration strategy sequence consists of a series of strategy operators, each responsible for specific decision logic or parameter adjustments. Through the operation object, the system can quickly identify which strategy operator needs modification and adjust its internal configuration instantly using the operation parameters, thereby achieving fine-grained control over the arbitration process. For example, a strategy operator registry can be maintained, recording the unique identifier of each strategy operator, the type of operation object it handles, and the structure of its internal configuration parameters. After parsing the operation object, the system queries the registry to find the strategy operator matching the operation object and directly modifies the operator's configuration object in memory according to the operation parameters (e.g., new weight values, new thresholds). Alternatively, a metadata-driven location and update mechanism can be used, where each strategy operator is accompanied by metadata describing its function, configurable parameters, and its association with specific objects in the knowledge graph (such as knowledge nodes and conflict edges). The system dynamically locates the strategy operator responsible for adjusting the weight of a knowledge node by querying the metadata based on the operation object (e.g., a knowledge node ID) and operation type (e.g., weight adjustment), and calls the API interface provided by the operator, passing in the operation parameters to update its internal configuration.

[0160] Based on the updated target policy operator and other unchanged policy operators in the current arbitration policy sequence, a revised arbitration policy sequence is generated. This step aims to ensure the integrity and logical consistency of the entire arbitration policy sequence after user intervention causes updates to some policy operators. By recombinating the updated target policy operator with other unaffected policy operators in the sequence, a new arbitration policy sequence reflecting the user's intervention intent can be quickly constructed, providing accurate input for subsequent recalculation. For example, the arbitration policy sequence can be represented as an ordered list of policy operators. When a target policy operator is updated, the system finds the position of that operator in the list, replaces the original operator instance with the updated operator instance, and obtains a new list of policy operators, i.e., the revised arbitration policy sequence. Alternatively, the arbitration policy sequence can also be modeled as a directed acyclic graph (DAG), where nodes are policy operators and edges represent the execution order. When a policy operator is updated, the system only needs to update the content of the corresponding node in the DAG, while keeping the connections of other nodes and edges unchanged, and then re-traverses the DAG to generate a new policy operator execution sequence.

[0161] The adjusted arbitration strategy sequence is input into the strategy orchestration model for recalculation, outputting updated decision logic and updated knowledge source weight allocation schemes. This step is crucial for achieving effective user intervention, translating user adjustments to strategy operators into updated decision logic and knowledge source weight allocation schemes. By re-inputting the new arbitration strategy sequence into the strategy orchestration model, the system can quickly derive the optimal decision path and knowledge source weights based on the latest strategy configuration, thus achieving real-time, dynamic responses to complex problem arbitration processes. For example, the strategy orchestration model can be a rule-based reasoning engine that receives the adjusted arbitration strategy sequence and executes these operators one by one according to the defined strategy operators and their invocation order. During the execution of each operator, the knowledge source weights are adjusted based on its internal logic and parameters, and a conflict resolution path is selected. The engine outputs the decision logic and knowledge source weight allocation schemes after processing by all operators. Alternatively, the strategy orchestration model can be a reinforcement learning (RL) model. Upon receiving the adjusted arbitration strategy sequence, this RL model uses it as new policy input, performing rapid iteration in a simulation environment or using a pre-trained model for reasoning. The model will re-evaluate the rewards of different decision paths based on the new policy sequence, and output the optimal decision logic (i.e. conflict resolution path) and knowledge source weight allocation scheme under the current policy.

[0162] Through the aforementioned technical solution, by accurately analyzing the standardized intervention paradigm and extracting the operation object, operation parameters, and operation type identifiers, the system ensures accurate identification and structured input of user intervention intentions, avoiding misjudgments and processing errors caused by information ambiguity. The system precisely locates the corresponding target strategy operator in the arbitration strategy sequence based on the operation object and updates its internal configuration parameters in real time according to the operation parameters. This ensures that user interventions can be applied to key stages of the decision-making process promptly and accurately, avoiding adjustment lags and positioning deviations. Based on the updated target strategy operator and the unchanged strategy operator, the system can efficiently regenerate the adjusted arbitration strategy sequence, ensuring the coherence and integrity of the arbitration logic while reducing unnecessary computational overhead. Inputting this adjusted arbitration strategy sequence into the strategy orchestration model for recalculation allows for the rapid output of updated decision logic and knowledge source weight allocation schemes, thereby achieving real-time response to user interventions. This ensures that the system always makes decisions based on the latest user intentions and knowledge status, greatly enhancing the adaptability, transparency, and user controllability of the dynamic arbitration engine.

[0163] In some embodiments of this application, a method for updating a dynamic conflict graph based on user intervention is proposed. However, in this process, the injection of new knowledge may not fully consider its modality type, compliance attributes, and potential conflict relationships, leading to low update efficiency and inaccurate conflict detection. This results in the inability to integrate new knowledge from user intervention in real time, exacerbating the problems of a knowledge fusion black box and insufficient conflict detection. To address this, this application further proposes a method for extracting new knowledge content based on knowledge injection operations in a standardized intervention paradigm, converting the new knowledge content into new knowledge nodes, and updating the dynamic conflict graph.

[0164] For example, this method includes: parsing the knowledge injection operation in the standardized intervention paradigm, extracting new knowledge content and its corresponding modality type and metadata information; extracting features from the new knowledge content based on the modality type to generate a new knowledge feature vector, and assigning corresponding associated compliance attributes to the new knowledge feature vector according to the compliance tag set; calculating the semantic similarity between the new knowledge feature vector and knowledge nodes in the dynamic conflict graph, and identifying potential conflict relationships between the new knowledge content and knowledge nodes in the dynamic conflict graph based on semantic similarity; assigning an effective event timestamp and a system entry timestamp as dual temporal markers to the new knowledge content, converting the new knowledge content into new knowledge nodes, and constructing new conflict edges between the new knowledge nodes and knowledge nodes in the dynamic conflict graph based on potential conflict relationships, while incrementally updating the dynamic conflict graph based on the new knowledge nodes and new conflict edges.

[0165] The process involves analyzing the knowledge injection operation within the standardized intervention paradigm to extract the new knowledge content, its corresponding modality type, and metadata. This step aims to accurately identify and extract the original new knowledge injected by the user and its related attributes from the standardized intervention paradigm generated by the user intervention. The analysis process can employ Natural Language Processing (NLP) technology to perform semantic analysis on the text-based intervention paradigm, identifying structured information such as the subject, predicate, and object of the knowledge content. Alternatively, if several pre-paradigms contain links or identifiers pointing to external knowledge sources, these identifiers can be used to access and extract the corresponding multimodal data. The extraction of modality types (such as text, image, audio, and video) and metadata information (such as knowledge source, creation time, author, and topic tags) ensures that the new knowledge can be correctly classified, encoded, and traced in subsequent processing. For example, for textual knowledge, keywords and summaries can be extracted. For image knowledge, descriptive tags and shooting location can be extracted.

[0166] Based on the modality type, features are extracted from the new knowledge content to generate a new knowledge feature vector. Then, corresponding associated compliance attributes are assigned to this new knowledge feature vector according to the compliance tag set. After obtaining the new knowledge content and its modality type, it needs to be transformed into a unified vector representation that the system can understand and process. The feature extraction process selects an appropriate deep learning model or feature engineering method based on the modality type of the new knowledge. For example, for text modality, pre-trained language models (such as BERT and GPT series) can be used to generate high-dimensional semantic vectors. For image modality, convolutional neural networks (such as ResNet and Vision Transformer) can be used to extract visual feature vectors. For audio modality, acoustic models (such as WaveNet) can be used to extract acoustic feature vectors. Simultaneously, based on the compliance tag set, associated compliance attributes are assigned to the generated new knowledge feature vector. This may involve incorporating information such as authorization status identifiers, source trust identifiers, and timeliness status identifiers from the compliance tag set into the feature vector as weighted coefficients or additional dimensions, or binding it as independent metadata to the feature vector, ensuring that the new knowledge has compliance considerations before entering the graph.

[0167] The semantic similarity between the new knowledge feature vector and the knowledge nodes in the dynamic conflict graph is calculated. Based on this semantic similarity, potential conflict relationships between the new knowledge content and the knowledge nodes in the dynamic conflict graph are identified. This step aims to evaluate the semantic association between the newly injected knowledge and the existing knowledge in the dynamic conflict graph, thereby discovering potential conflicts. Semantic similarity calculation is usually performed in a shared semantic space by comparing the distance or angle (such as cosine similarity or Euclidean distance) between the new knowledge feature vector and the feature vectors of all knowledge nodes in the graph. For example, if the cosine similarity of two knowledge units is lower than a preset threshold, or the Euclidean distance is greater than a preset threshold, semantic inconsistency may exist. The identification of potential conflict relationships can be based on a similarity threshold; that is, when the similarity is lower than a certain value, a potential conflict is considered to exist. Alternatively, clustering algorithms can be used to cluster the new knowledge with existing knowledge. If the new knowledge falls into a region close to an existing conflict cluster, it is marked as a potential conflict.

[0168] The new knowledge content is assigned a valid event timestamp and a system entry timestamp as dual-temporal markers. This new knowledge content is then converted into a new knowledge node. Based on the potential conflict relationship, new conflict edges are constructed between this new knowledge node and knowledge nodes in the dynamic conflict graph. Simultaneously, the dynamic conflict graph is incrementally updated based on the new knowledge node and these new conflict edges. To maintain the timeliness and historical evolution of knowledge, the new knowledge content is assigned a valid event timestamp and a system entry timestamp as dual-temporal markers when converted into a new knowledge node. The valid event timestamp represents the actual occurrence or effective time range of the event described by the knowledge content, while the system entry timestamp records the time when the knowledge is received and processed by the system. The new knowledge node will contain a new knowledge feature vector, associated compliance attributes, and dual-temporal markers. Based on the identified potential conflict relationship, new conflict edges are constructed between the new knowledge node and semantically conflicting knowledge nodes in the dynamic conflict graph. These conflict edges can carry attributes such as conflict type and conflict intensity. By adding the new knowledge node and new conflict edges to the existing graph, the dynamic conflict graph is incrementally updated, ensuring that the graph can reflect the latest state of knowledge and conflict situations in real time.

[0169] Through the aforementioned technical solution, by meticulously analyzing the knowledge injection operation in the standardized intervention paradigm, new knowledge content, modality type, and metadata information are extracted. This ensures the complete capture of the original data and formal characteristics of new knowledge from user intervention, providing accurate input for subsequent feature extraction and avoiding processing deviations caused by missing information. Feature extraction is performed based on modality type, and associated compliance attributes are assigned, ensuring that new knowledge meets ethical and compliance requirements before integration, preventing conflicts arising from adding compliance afterward, and transforming it into a vector form that the system can process. The semantic similarity between the new knowledge feature vector and knowledge nodes in the dynamic conflict graph is calculated, and potential conflict relationships are identified. Semantic similarity calculations are used to directly compare the semantic consistency between new knowledge and existing knowledge, automatically detecting potential conflicts and ensuring that inconsistencies are not introduced when new knowledge is integrated, thus improving the accuracy of conflict identification. New knowledge content is assigned an effective event timestamp and a system entry timestamp as dual temporal markers, which are then converted into new knowledge nodes. New conflict edges are constructed based on potential conflict relationships, and the dynamic conflict graph is incrementally updated to maintain the timeliness of knowledge and enable the graph to dynamically reflect changes. The incremental update mechanism ensures that the system responds to the injection of new knowledge in real time, thereby improving update efficiency.

[0170] In some of the solutions mentioned above in this application, dynamic conflict maps and user cognitive state vectors are updated based on intervention results to optimize the system decision state in real time. However, in this process, there is a lack of a refined analysis mechanism for intervention results, a dynamic integration method for new knowledge content, proactive maintenance measures for knowledge timeliness, and real-time adaptability to user behavior patterns. This may result in the update process failing to accurately process user operations, effectively integrate new knowledge, promptly eliminate outdated information, or accurately reflect changes in user cognition, thereby reducing the efficiency and accuracy of system updates and affecting the overall credibility of decision-making.

[0171] To address this, this application further proposes a method for updating the dynamic conflict graph and user cognitive state vector based on intervention results. This method includes: parsing the operation object and operation parameters from the standardized intervention paradigm; locating the corresponding target knowledge node in the dynamic conflict graph based on the operation object; and adjusting the node weight of the target knowledge node and the conflict edge weight associated with the target knowledge node based on the operation parameters. New knowledge content is extracted from the standardized intervention paradigm, and features are extracted from the new knowledge content to generate a new knowledge feature vector. An effective event timestamp and a system entry timestamp are assigned to the new knowledge content as dual-temporal markers. The new knowledge content is converted into new knowledge nodes. The semantic similarity between the new knowledge feature vector and the knowledge nodes in the dynamic conflict graph is calculated. Potential conflict relationships are identified based on semantic similarity, and new conflict edges are constructed between the new knowledge nodes and the knowledge nodes in the dynamic conflict graph. Based on the updated dual-temporal markers of all knowledge nodes, knowledge nodes exceeding the effective event timestamp are marked as invalid. A decay constraint is applied to the weights of all conflict edges based on the system entry timestamp, generating the updated dynamic conflict graph. The user behavior patterns corresponding to the operation trajectory are analyzed from the standardized intervention paradigm. Based on the user behavior patterns, the user's professional level dimension or task urgency dimension in the user cognitive state vector are incrementally corrected to generate an updated user cognitive state vector.

[0172] For example, when parsing the operation object and operation parameters from the standardized intervention paradigm, locating the corresponding target knowledge node in the dynamic conflict graph based on the operation object, and adjusting the node weight of the target knowledge node and the weight of the conflict edge associated with the target knowledge node based on the operation parameters, this step aims to accurately understand the user's intervention intention and translate it into a quantitative adjustment of specific knowledge nodes and their associated conflict edges in the dynamic conflict graph. By refining the parsing of the standardized intervention paradigm, the system can identify the specific knowledge unit (operation object) that the user wants to modify and how to modify it (operation parameters), thereby avoiding errors that may be caused by generalization and ensuring the targeted application of intervention results. For example, the standardized intervention paradigm can be parsed using a predefined parser or a rule-based engine. If the operation object is a knowledge node ID and the operation parameter is a new weight value, the system directly updates the weight attribute of the knowledge node. If the operation object is a conflict edge ID and the operation parameter is a weight adjustment ratio, the system multiplicatively or additively corrects the current weight of the conflict edge according to the ratio. Alternatively, a machine learning-based approach can be used to train a small neural network model that takes the standardized intervention paradigm as input and outputs specific adjustment instructions for the weights of the target knowledge node and the conflict edge. The model is able to learn complex patterns of user intervention and map them onto fine-tuning of the graph.

[0173] This process, which involves extracting new knowledge content from standardized intervention paradigms, generating new knowledge feature vectors through feature extraction, assigning valid event timestamps and system entry timestamps as dual-temporal markers to the new knowledge content, converting the new knowledge content into new knowledge nodes, calculating the semantic similarity between the new knowledge feature vectors and knowledge nodes in the dynamic conflict graph, identifying potential conflict relationships based on semantic similarity, and constructing new conflict edges between the new knowledge nodes and knowledge nodes in the dynamic conflict graph, addresses the problem of dynamically integrating new knowledge content into the system. This ensures that user-injected new information is effectively integrated and proactively detects potential conflicts with existing knowledge. By assigning dual-temporal markers to new knowledge, the system can better manage the lifecycle and effectiveness of knowledge. For example, feature extraction of new knowledge content can utilize a pre-trained multimodal embedding model to uniformly map new knowledge from different modalities such as text, images, and audio to a shared semantic space, generating new knowledge feature vectors. Semantic similarity calculation can use cosine similarity or Euclidean distance. Potential conflict relationship identification can be performed by setting a similarity threshold or using a specially trained conflict detection classifier. New knowledge nodes and conflict edges are created and connected via the graph database's API. Alternatively, a knowledge graph embedding approach can be used, encoding new knowledge content and its metadata into entities and relations, and embedding them into the vector space of the existing knowledge graph. Potential conflicts are identified by calculating the distance between new knowledge entities and existing entities, and new conflict edges are automatically generated. Bitemporal tagging can be extracted from user input or knowledge source metadata, or inferred from new knowledge content using natural language processing techniques.

[0174] By marking knowledge nodes that have exceeded their valid event timestamps as invalid based on the dual-temporal tags of all updated knowledge nodes, and applying decay constraints to the weights of all conflict edges according to the system's input timestamps to generate an updated dynamic conflict graph, this step aims to dynamically maintain the timeliness and relevance of knowledge, avoid interference from outdated information in system decision-making, and reflect potential changes in conflicts over time. By marking knowledge nodes that have exceeded their validity period as invalid, the system can ensure that only currently valid knowledge is used. By applying decay constraints to the weights of conflict edges, the system can reflect the dynamic evolution of conflicts; for example, older conflicts may have been resolved or become less important.

[0175] By analyzing user behavior patterns corresponding to operational trajectories from standardized intervention paradigms, and incrementally adjusting the user's professional level or task urgency dimensions in the user's cognitive state vector based on these patterns to generate an updated user cognitive state vector, the system can capture user behavior characteristics in real time during interactions and dynamically adjust its understanding of user cognitive states accordingly. This enhances the system's adaptability and personalized service capabilities. By analyzing operational trajectories and identifying user behavior patterns, the system can more accurately reflect the user's professional level and task urgency, thereby optimizing subsequent interactions and decision-making processes. For example, operational trajectory analysis can be achieved by recording user events such as clicks, drags, and inputs on a visualized decision-making path. User behavior pattern identification can be based on a pre-defined set of rules, such as, "If a user frequently adjusts the weight of high-weight knowledge nodes, increase their professional level score," or "If a user performs a large number of operations in a short period without sufficient thought, increase their task urgency score." Incremental adjustments are achieved by weighted summation or linear superposition of the identified adjustment coefficients with the existing user cognitive state vector. Alternatively, a behavior pattern recognition model can be trained, taking user action trajectories as input and outputting adjustments to the user's expertise level and task urgency. This model can learn complex correlations between behavioral patterns and changes in cognitive state from historical user behavior data. Incremental adjustments can be made by using the model's output as the update gradient of the user's cognitive state vector, and then iteratively updating it using optimization algorithms.

[0176] By analyzing the operational objects and parameters from the standardized intervention paradigm and precisely adjusting the weights of target knowledge nodes and their associated conflict edges in the dynamic conflict graph, this application can directly and accurately apply the user's intervention intent to specific parts of the graph, avoiding errors caused by generalization in traditional methods and ensuring the targeted application and efficient execution of intervention results. Simultaneously, this application can extract new knowledge content injected by the user from the standardized intervention paradigm and perform feature extraction to generate new knowledge feature vectors. By assigning valid event timestamps and system entry timestamps as dual-temporal markers to the new knowledge content and converting it into new knowledge nodes, the system can not only seamlessly integrate new information into the dynamic conflict graph but also proactively identify and construct potential conflict relationships by calculating the semantic similarity between the new knowledge feature vectors and existing knowledge nodes. This effectively prevents inconsistencies that may arise from newly introduced knowledge and ensures the internal coordination of the knowledge base.

[0177] In some of the solutions mentioned above in this application, incremental correction of the user's cognitive state vector based on user behavior patterns is proposed to update the user's cognitive state. However, in its implementation, there is a lack of specific methods to extract effective behavioral features from the user's intervention operation trajectory and convert them into adjustment coefficients, resulting in inaccurate correction and inability to dynamically adapt to changes in user behavior.

[0178] To address this, this application further proposes an incremental correction of the user's professional level dimension or task urgency dimension in the user's cognitive state vector based on user behavior patterns, generating an updated user cognitive state vector. This process includes: parsing the operational trajectory from a standardized intervention paradigm, extracting the intervention frequency, intervention depth, and intervention accuracy corresponding to the operational trajectory as user behavior features; inputting the user behavior features into a preset behavior pattern recognition model, outputting the professional level adjustment coefficient or urgency adjustment coefficient corresponding to the user behavior pattern; obtaining the current user professional level dimension and task urgency dimension in the user cognitive state vector, and incrementally correcting the user professional level dimension based on the professional level adjustment coefficient, or incrementally correcting the user task urgency dimension based on the urgency adjustment coefficient; and combining the corrected user professional level dimension and task urgency dimension with other dimensions in the user cognitive state vector to generate an updated user cognitive state vector.

[0179] For example, analyzing operational trajectories from standardized intervention paradigms and extracting the corresponding intervention frequency, depth, and accuracy as user behavior features aims to quantify user behavior patterns from interactions with the visualized decision path graph. An operational trajectory is a sequence of all user actions on the interface, while the standardized intervention paradigm is a structured representation of these actions. Intervention frequency, depth, and accuracy are key indicators for measuring user behavior. For instance, intervention frequency can be calculated by counting the total number of intervention operations (such as node weight adjustment, path selection, and knowledge injection) performed by the user within a unit of time (e.g., one minute or one dialogue session). Intervention depth can be quantified by assessing the impact of user intervention operations on the decision path graph or knowledge source weights; for example, for node weight adjustment, the absolute or relative value of the weight change before and after adjustment can be calculated. For knowledge injection, the complexity of the injected knowledge or its impact on the existing knowledge graph structure can be assessed. Intervention accuracy can be evaluated by comparing the consistency between the system's decision results after user intervention and the preset expert decisions or subsequent user satisfaction feedback; for example, if the system shows higher accuracy or user satisfaction in subsequent dialogues after user adjustment, the intervention accuracy is considered high.

[0180] User behavior characteristics are input into a pre-defined behavior pattern recognition model, which outputs an adjustment coefficient for either professional level or urgency level corresponding to the user behavior pattern. This step aims to transform quantified user behavior characteristics into correction coefficients for specific dimensions of the user's cognitive state vector. The pre-defined behavior pattern recognition model is the core, capable of learning the mapping relationship between user behavior characteristics and cognitive state dimension adjustments. This model can employ machine learning models, such as support vector machines, decision trees, random forests, or neural networks (such as multilayer perceptrons), and is trained using historical user behavior data and corresponding cognitive state changes. For example, training data could include user behaviors with "high frequency, high depth, and high accuracy" corresponding to a "professional level adjustment coefficient +0.2". The adjustment coefficients output by the model can be continuous values ​​(e.g., between -0.5 and 0.5), representing the incremental correction to the professional level or urgency level dimension. Alternatively, they can be discrete values, for example, divided into several levels such as "increase," "slight increase," "no change," "slight decrease," and "decrease," with each level corresponding to a pre-defined adjustment value.

[0181] The process involves obtaining the current user expertise level and task urgency dimensions from the user's cognitive state vector. Then, it incrementally adjusts either the expertise level dimension based on an adjustment coefficient, or the task urgency dimension based on an adjustment coefficient. This step is the actual execution of the user cognitive state vector adjustment, employing incremental adjustment rather than complete replacement to maintain state continuity and stability. Incremental adjustment can be achieved by adding the current dimension value to the adjustment coefficient; for example, the updated expertise level equals the current expertise level plus the expertise level adjustment coefficient. To prevent dimension values ​​from exceeding a reasonable range, upper and lower limits can be set for truncation. Alternatively, a weighted average method can be used, fusing the current dimension value and the adjustment coefficient according to certain weights.

[0182] The revised user expertise level and task urgency dimensions are combined with other dimensions in the user cognitive state vector to generate an updated user cognitive state vector. This step ensures the integrity of the user cognitive state vector by reintegrating the revised specific dimensions with other dimensions unaffected by this intervention, forming a comprehensively updated cognitive state representation. The most direct approach is to replace the corresponding positions in the original user cognitive state vector with the revised expertise level and task urgency dimensions, while leaving other dimensions unchanged. Alternatively, the revised two dimension values ​​can be concatenated or combined with other unrevised dimension values ​​to form a new vector, ensuring that the vector's structure and dimensions remain consistent.

[0183] Through the above technical solution, this application achieves the extraction of effective behavioral features from user intervention operation trajectories and their conversion into adjustment coefficients, thereby accurately correcting the user's cognitive state vector and solving the problems of inaccurate correction and inability to adapt to dynamic changes. For example, by quantifying user behavioral features such as intervention frequency, intervention depth, and intervention accuracy, accurate basic data is provided for subsequent adjustments, avoiding ambiguity in feature extraction. A preset behavioral pattern recognition model is used to map behavioral features into adjustment coefficients, ensuring that coefficient generation is based on user behavioral patterns and dynamically adapts to different scenarios. Incremental correction, rather than a full reset, is performed on the user's cognitive state vector, maintaining state continuity and improving correction accuracy. By combining the corrected dimensions with other dimensions, the consistency of the entire user's cognitive state vector update is ensured, achieving dynamic optimization of the overall cognitive state.

[0184] In some of the embodiments described above in this application, a dynamic conflict graph is proposed to dynamically adjust the conflict detection sensitivity based on the user's cognitive state vector, in order to personalize conflict detection to adapt to different user needs. However, in its implementation, the sensitivity adjustment only depends on the user's state and does not fully consider the real-time conflict distribution state of the graph itself, historical detection performance, and compliance attributes of knowledge units. This may cause the detection sensitivity to deviate from the actual scenario requirements, affecting the accuracy and efficiency of conflict identification, and thus reducing the reliability and adaptability of the system's decision-making.

[0185] To address this, this application further proposes a method for determining the conflict detection sensitivity of a dynamic conflict map. This method can more comprehensively and accurately determine the conflict detection sensitivity, thereby improving the accuracy of conflict identification and the system's adaptability. For example, the method includes the following steps: This step involves analyzing the user's professional level and task urgency dimensions from the user's cognitive state vector to obtain a professional level score and a task urgency score. The goal is to quantify the user's cognitive state in the current human-computer dialogue scenario. The user's cognitive state vector is a multi-dimensional representation of their current cognitive state, including professional knowledge level and task urgency. The professional level dimension reflects the user's knowledge and experience in a specific domain, while the task urgency dimension indicates the time sensitivity of the current task. In practice, various methods can be used for analysis. For example, a pre-trained natural language processing model can be used to perform in-depth analysis of the user's historical dialogue records and user profile information to extract keywords, phrases, or semantic patterns related to professional level and task urgency, and map them to a preset scoring range to assess the user's professional level and task urgency. Alternatively, analyzing user interaction data within the system, such as the user's decision-making speed on complex problems and the accuracy of their corrections to system suggestions, combined with machine learning models for real-time inference, can yield more dynamic and accurate professional level and task urgency scores.

[0186] This step involves obtaining the distribution density of current conflict nodes and historical conflict detection accuracy statistics in a dynamic conflict graph to generate a graph state complexity metric. This step aims to evaluate the real-time state and historical performance of the dynamic conflict graph itself. The dynamic conflict graph is a dynamic representation of knowledge units and their identified conflict relationships. The distribution density of conflict nodes reflects the concentration of conflicts in the graph, while historical conflict detection accuracy statistics provide a performance baseline for the detection mechanism. The graph state complexity metric is a comprehensive measure reflecting the current conflict situation and the reliability of past detections. Specifically, the distribution density of conflict nodes can be obtained by calculating the number of conflict edges or conflict nodes within a specific region of the graph (e.g., a subgraph partitioned based on semantic relevance). Historical conflict detection accuracy statistics can be obtained from system logs, such as recording user feedback (e.g., confirmation, rejection, correction) or subsequent arbitration results after each conflict detection, and periodically calculating their accuracy. These data can be used to comprehensively generate the graph state complexity metric through weighted averaging, cluster analysis, or regression models. Another approach is to identify conflict hotspots using graph analysis algorithms (such as community detection algorithms) and quantify the conflict intensity in these areas as distribution density. Historical accuracy can then be periodically evaluated through A / B testing or expert annotation.

[0187] The professional level score, the task urgency score, and the graph state complexity index are input into the sensitivity adaptive adjustment model, which then outputs an initial sensitivity threshold. This step aims to integrate user state and graph state information to determine the initial conflict detection sensitivity. The sensitivity adaptive adjustment model is an intelligent model capable of learning the relationship between input factors (user state, graph state) and the optimal conflict detection sensitivity. The initial sensitivity threshold is a preliminary value that determines the strictness of conflict detection. This model can be implemented using a rule-based expert system or decision tree model, for example, by pre-setting a series of rules to match the corresponding threshold based on the input parameters. Alternatively, machine learning models, such as support vector machines, neural networks, or reinforcement learning models, can be used, trained on historical data (including conflict detection performance under different user states and graph states, and user feedback), to learn how to predict the optimal initial sensitivity threshold based on input features.

[0188] Based on the distribution of associated compliance attributes of each knowledge unit in the compliance tag set, the initial sensitivity threshold is adjusted to generate a conflict detection sensitivity. This sensitivity is used to determine whether a conflict exists between knowledge unit pairs during cross-modal conflict identification. This step aims to refine the initial sensitivity based on ethical and compliance considerations. The compliance tag set contains compliance metadata for each knowledge unit, such as data source, authorization status, and timeliness. The distribution of associated compliance attributes refers to the statistical characteristics of these compliance attributes in the considered knowledge units. The conflict detection sensitivity is a refined threshold used for actual conflict identification. During the adjustment process, a set of compliance adjustment rules can be preset. For example, if a knowledge unit contains a large amount of sensitive data or originates from a low-trust source, the initial sensitivity threshold should be adjusted upwards to more strictly detect potential conflicts and prevent the spread of non-compliant information. Conversely, if the compliance of a knowledge unit is extremely high, the sensitivity can be appropriately relaxed. Another correction method is to use an independent compliance risk assessment module to calculate the compliance risk coefficient based on the distribution of attributes such as authorization status identifier, source credibility identifier, and timeliness status identifier in the compliance label set, and then perform a weighted summation or multiplication correction with the initial sensitivity threshold to obtain the conflict detection sensitivity.

[0189] Through the aforementioned technical solution, this application comprehensively considers the user's cognitive state, the complexity of the knowledge graph, and the compliance attributes of the knowledge units. This allows the conflict detection sensitivity to more comprehensively and accurately reflect the actual needs and risk conditions of the current human-computer dialogue scenario, avoiding biases caused by adjustments in a single dimension and improving the accuracy of conflict identification. Simultaneously, the sensitivity can be personalized according to the user's professional level and the urgency of the task, responding to real-time changes in the conflict distribution in the knowledge graph and historical detection performance. This ensures that the system maintains efficient and appropriate conflict detection capabilities under scenarios of varying complexity and urgency, thereby enhancing dynamic adaptability.

[0190] In some of the solutions mentioned above in this application, the presentation complexity of the visual decision-making path diagram is adjusted based on the user's cognitive state vector to improve decision transparency and user intervention capabilities. However, in this process, there is a lack of specific mechanisms to dynamically determine how the presentation details should adapt to different users' professional levels and task urgency. This may result in the presentation content being too complex or too simplified, which is not suitable for the user's current cognitive state, affecting the user's understanding of the decision-making logic and interaction efficiency, thereby reducing the system's credibility and real-time intervention effect.

[0191] To address this, this application further proposes a method for determining the presentation complexity of a visualized decision path map. This method includes: parsing the user's professional level dimension and task urgency dimension from the user's cognitive state vector to obtain a professional level score and a task urgency score. Based on the professional level score, determining the number of expansion levels of knowledge source weight nodes and the level of detail of decision labels in the visualized decision path map using preset hierarchical mapping rules. Based on the task urgency score, determining the rendering refresh frequency level and the response sensitivity level of user interaction operations in the visualized decision path map using preset response mapping rules. Combining and encoding the number of expansion levels, the level of detail, the rendering refresh frequency level, and the response sensitivity level generates the presentation complexity of the visualized decision path map.

[0192] For example, by analyzing the user's professional level and task urgency dimensions from the user's cognitive state vector, a professional level score and a task urgency score can be obtained. The aim is to extract key indicators of the user's current cognitive state as a basis for subsequent adjustments to the presentation complexity. The user's cognitive state vector can be obtained in various ways, such as based on the user's historical operation records in the system, dialogue context information, explicit user settings, or through comprehensive prediction using machine learning models (such as neural networks). During analysis, the professional level score and task urgency score can be extracted directly from specific dimensions of the user's cognitive state vector, or obtained through predefined linear or nonlinear transformations. For example, the professional level dimension can be evaluated based on the user's depth of knowledge and operational proficiency in a specific field, while the task urgency dimension can be determined based on keywords appearing in the dialogue (such as "urgent" or "immediate") or the task priority set by the system. The scores can be discrete levels (such as beginner, intermediate, and advanced) or continuous numerical values.

[0193] Based on professional level scores, a pre-defined hierarchical mapping rule determines the number of expansion levels for knowledge source weight nodes and the level of detail for decision labels in the visualized decision path graph. Its purpose is to dynamically adjust the information density and granularity of the decision path graph according to the user's professional level, avoiding information overload or under-information. This pre-defined hierarchical mapping rule can be a lookup table, a piecewise function, or a rule-based expert system. For example, when the professional level score is "beginner," the system can set the number of expansion levels to a smaller value (e.g., displaying only core decision nodes) and the level of detail for decision labels to "brief description" to provide a high-level overview. When the professional level score is "advanced," the system can set the number of expansion levels to a larger value (e.g., displaying all child nodes) and the level of detail for decision labels to "detailed explanation" to provide more in-depth technical details. Knowledge source weight nodes can be stratified according to their importance or scope of influence in the decision-making process. The number of expansion levels can refer to the visible depth of nodes in the graph, and the level of detail can refer to the length of the text description of nodes or edges and the degree of use of professional terminology.

[0194] Based on task urgency scores, a preset response mapping rule is used to determine the rendering refresh frequency level of the visualized decision path map and the response sensitivity level of user interactions. This aims to optimize the system's real-time performance and interactive experience according to the task's urgency, ensuring rapid response at critical moments. This preset response mapping rule can define refresh frequencies and sensitivities for different urgency levels. For example, when the task urgency score is "high," the rendering refresh frequency level can be set to "real-time" (e.g., millisecond-level refresh), and the response sensitivity level for user interactions can be set to "extremely high" (e.g., imperceptible delay response), ensuring users can instantly access the latest information and intervene quickly. When the task urgency score is "low," the refresh frequency can be appropriately reduced (e.g., second-level refresh), and the response sensitivity can be set to "medium" to balance system resource consumption. The rendering refresh frequency level refers to the interval between interface updates after map data updates, and the response sensitivity level refers to the system's response delay after user clicks, drags, or other operations.

[0195] The presentation complexity of the visualization decision path map is generated by combining and encoding the number of expansion levels, level of detail, level of rendering refresh rate, and level of response sensitivity. The aim is to integrate multiple independent adjustment parameters into a unified complexity index, facilitating overall system control and management and ensuring consistent presentation. This combined encoding can be implemented in various ways; for example, the four level values ​​(after normalization) can be combined into a multi-dimensional vector, or mapped to a single complexity score using a pre-trained encoder. The encoding result can be an enumeration type (such as "minimalist," "standard," "detailed," "expert") or a continuous complexity index. This encoding result is then received by the visualization rendering module, which adjusts its rendering parameters, such as layout algorithms, animation effects, and information prompts, accordingly to achieve the desired presentation complexity.

[0196] Through the aforementioned technical solution, this application can accurately analyze the user's professional level and task urgency dimensions from the user's cognitive state vector, thereby obtaining a precise profile of the user's current cognitive state and providing a solid data foundation for subsequent dynamic adjustments. Based on the professional level score, the system can intelligently adjust the number of expansion levels of knowledge source weight nodes and the level of detail of decision labels in the visualized decision-making path diagram through preset hierarchical mapping rules. This ensures that professional users can delve into every detail of the decision-making logic, while ordinary users can obtain a concise and clear overview, avoiding the problems of information overload or insufficient information and improving the user's efficiency in understanding the decision-making process.

[0197] All of the above-mentioned optional technical solutions can be combined in any way to form the optional embodiments of this application, and will not be described in detail here.

[0198] Figure 6 This is a schematic diagram of the structure of a complex problem knowledge externalization modeling system based on human-computer dialogue, provided in an embodiment of this application. See also... Figure 6 The system includes: The acquisition module 601 is used to acquire the target complex problem, multi-source cross-modal knowledge data, dialogue context features, user cognitive state features and data traceability authorization information in the human-computer dialogue scenario. After compliance screening and structured coding, a standardized knowledge set, user cognitive state vector and compliance label set are obtained.

[0199] The identification module 602 is used to perform cross-modal semantic unified mapping and conflict identification based on the standardized knowledge set and the compliance tag set through the cross-modal conflict graph builder, and construct a dynamic conflict graph with dual temporal labels. The knowledge node weights of the dynamic conflict graph are constrained by the compliance tag set, and the conflict detection sensitivity of the dynamic conflict graph is dynamically adjusted based on the user's cognitive state vector.

[0200] The encapsulation module 603 is used to encapsulate the decision logic, knowledge source weights and conflict resolution paths corresponding to the arbitration strategy into a visual decision path diagram based on the dynamic conflict graph, the dialogue context features, the user cognitive state vector and the compliance tag set, through the dynamic arbitration engine. The presentation complexity of the visual decision path diagram is adjusted based on the user cognitive state vector, and the arbitration strategy needs to undergo ethical pre-checking.

[0201] Storage module 604 is used to obtain user intervention results based on the visualized decision path graph, parse the intervention results into a standardized intervention paradigm, and transmit it back in real time to trigger the recalculation of the arbitration strategy. After ethical pre-checking, the intervention paradigm is stored in the example library, which is used to optimize the arbitration strategy generation logic for similar conflict scenarios in the future, and to update the dynamic conflict graph and the user cognitive state vector simultaneously based on the intervention results.

[0202] It should be noted that the above-described embodiments of the complex problem knowledge externalization modeling system based on human-computer dialogue are only illustrative examples of the functional module divisions. In practical applications, the functions described above can be assigned to different functional modules as needed, that is, the internal structure of the computer device can be divided into different functional modules to complete all or part of the functions described above. Furthermore, the complex problem knowledge externalization modeling system based on human-computer dialogue and the method embodiment based on human-computer dialogue share the same concept; the specific implementation process is detailed in the method embodiment and will not be repeated here.

[0203] Those skilled in the art will understand that all or part of the steps of the above embodiments can be implemented by hardware or by a program instructing related hardware. The program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk.

[0204] The above are merely optional embodiments of this application and are not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for knowledge externalization modeling of complex problems based on human-machine dialogue, characterized in that, The method includes: The system acquires complex target questions, multi-source cross-modal knowledge data, dialogue context features, user cognitive state features, and data traceability authorization information in human-computer dialogue scenarios. After compliance screening and structured coding, it obtains a standardized knowledge set, a user cognitive state vector, and a compliance tag set. Based on the standardized knowledge set and the compliance tag set, a cross-modal semantic unified mapping and conflict identification are performed through a cross-modal conflict graph builder to construct a dynamic conflict graph with dual temporal labels. The knowledge node weights of the dynamic conflict graph are constrained by the compliance tag set, and the conflict detection sensitivity of the dynamic conflict graph is dynamically adjusted based on the user cognitive state vector. Based on the dynamic conflict graph, the dialogue context features, the user cognitive state vector, and the compliance tag set, the decision logic, knowledge source weights, and conflict resolution paths corresponding to the arbitration strategy are encapsulated into a visual decision path graph through the dynamic arbitration engine. The presentation complexity of the visual decision path graph is adjusted based on the user cognitive state vector, and the arbitration strategy needs to undergo ethical pre-checking. Based on the visualized decision path graph, user intervention results are obtained, and the intervention results are parsed into standardized intervention paradigms and transmitted back in real time to trigger recalculation of arbitration strategies. The intervention paradigms are stored in the example library after ethical pre-checking. The example library is used to optimize the arbitration strategy generation logic for similar conflict scenarios in the future, and the dynamic conflict graph and the user cognitive state vector are updated simultaneously based on the intervention results.

2. The method of claim 1, wherein, The process of acquiring complex target questions, multi-source cross-modal knowledge data, dialogue context features, user cognitive state features, and data source authorization information in human-computer dialogue scenarios, after compliance screening and structured coding, yields a standardized knowledge set, a user cognitive state vector, and a compliance tag set, including: Semantic parsing is performed on the target complex problem to extract the problem intent and key entities, thereby obtaining the core elements of the problem; Based on the core elements of the problem, modal alignment and feature unification are performed on the multi-source cross-modal knowledge data to obtain aligned knowledge items that are semantically associated with the core elements of the problem; The dialogue context features and the user cognitive state features are fused and encoded to generate a user cognitive state vector for guiding knowledge selection. Based on the core elements of the problem and the user cognitive state vector, candidate knowledge items are selected from the aligned knowledge items, and the legality of the data traceability authorization information is verified to generate a data source compliance identifier for each candidate knowledge item. Then, the candidate knowledge items are verified for compliance based on the data source compliance identifier. The verified knowledge items are encoded into the standardized knowledge set, and the user cognitive state vector is output. The user cognitive state vector is used to dynamically adjust the conflict detection sensitivity in subsequent steps. At the same time, the data source compliance identifier is aggregated into the compliance tag set associated with the standardized knowledge set according to preset rules.

3. The method of claim 2, wherein, The step of performing compliance verification on the candidate knowledge items based on the compliance identifier of the data source, and encoding the verified knowledge items into the standardized knowledge set, includes: Parse the data source compliance identifier corresponding to each candidate knowledge item, and extract the authorization status identifier, source trust identifier, and timeliness status identifier; Based on compliance policy rules, a comprehensive judgment is made on the authorization status identifier, the source trust identifier, and the timeliness status identifier, and candidate knowledge items whose judgment results meet the preset compliance threshold are determined as knowledge items that have passed the verification. The verified knowledge items are structured and encoded to generate standardized knowledge units that include knowledge content ontology and associated compliance attributes. The associated compliance attributes are mapped from the compliance identifier of the data source. All the standardized knowledge units constitute the standardized knowledge set.

4. The method of claim 1, wherein, Based on the standardized knowledge set and the compliance tag set, a cross-modal semantic unified mapping and conflict identification are performed through a cross-modal conflict graph builder to construct a dynamic conflict graph with dual temporal labels, including: The cross-modal conflict graph builder assigns an event validity timestamp and a system entry timestamp to each knowledge unit in the standardized knowledge set, serving as a dual-temporal marker for the knowledge unit. The cross-modal conflict graph builder extracts features from each knowledge unit in the standardized knowledge set to obtain an initial feature vector for the corresponding modality. It also parses the associated compliance attributes of each knowledge unit from the compliance tag set and weights the initial feature vectors based on the associated compliance attributes to generate a modal feature vector with compliance constraints. The cross-modal conflict graph builder projects the modal feature vector with compliance constraints carrying the dual temporal labels into a shared semantic space, and calculates the cross-modal semantic similarity between different knowledge units in the shared semantic space to obtain the cross-modal conflict probability matrix. The cross-modal conflict graph builder identifies knowledge unit pairs exceeding the conflict detection threshold as conflict node pairs based on the cross-modal conflict probability matrix. A conflict edge is constructed for each conflict node pair. The initial weight of the conflict edge is determined according to the corresponding probability value in the cross-modal conflict probability matrix. A time-decay constraint is applied to the weight of the conflict edge based on the system entry timestamp of the knowledge unit in the conflict node pair. Simultaneously, knowledge units exceeding the valid time range are marked as invalid based on the valid event timestamp, thus generating the dynamic conflict graph.

5. The method of claim 1, wherein, The process, based on the dynamic conflict graph, the dialogue context features, the user cognitive state vector, and the compliance tag set, encapsulates the decision logic, knowledge source weights, and conflict resolution paths corresponding to the arbitration strategy into a visual decision path graph through a dynamic arbitration engine, including: The dynamic arbitration engine extracts conflict distribution features from the dynamic conflict map to obtain conflict type distribution vector, conflict intensity distribution vector, and conflict node density distribution vector. The dynamic arbitration engine generates an arbitration strategy sequence for the current conflict graph by combining the conflict type distribution vector, the conflict intensity distribution vector, the conflict node density distribution vector, the dialogue context features, the user cognitive state vector, and the compliance tag set input strategy orchestration model. The arbitration strategy sequence includes multiple strategy operators and their calling order. The dynamic arbitration engine simulates the adjustment process of knowledge source weights and the selection process of conflict resolution paths for each strategy operator according to the calling order and parameter configuration of each strategy operator in the arbitration strategy sequence, and generates intermediate state records for each step. The dynamic arbitration engine visualizes and encapsulates the arbitration strategy sequence, intermediate state records, and the final determined knowledge source weight allocation scheme and conflict resolution path according to a graph structure, generating a visualized decision path graph containing nodes, edges, and decision labels. Nodes in the visualized decision path graph represent decision states or intermediate results, edges represent strategy operators or decision flows, and decision labels are used to mark the decision basis for each step.

6. The method of claim 5, wherein, The process of simulating the adjustment of knowledge source weights and the selection of conflict resolution paths for each strategy operator based on the calling order and parameter configuration of each strategy operator in the arbitration strategy sequence, generating intermediate state records for each step, includes: Based on the parameter configuration of the first strategy operator in the arbitration strategy sequence, the current knowledge source weight in the dynamic conflict graph is used as input, and the first strategy operator is executed to obtain the adjusted first knowledge source weight set and the first intermediate conflict graph. Based on the remaining conflict distribution in the first intermediate conflict map, the initial conflict resolution path selected after the execution of the first strategy operator is simulated, and the first decision basis corresponding to the initial conflict resolution path is recorded. Using the first knowledge source weight set and the first intermediate conflict graph as inputs to the next strategy operator, the subsequent strategy operators in the arbitration strategy sequence are executed sequentially to obtain the knowledge source weight set and the corresponding intermediate conflict graph after each step. Record the conflict resolution path selected at each step, the decision basis for each step, and the state change of the conflict graph after each step is executed. Encapsulate all the records of the steps according to the execution sequence to generate the intermediate state record.

7. The method of claim 1, wherein, The process of obtaining user intervention results based on the visualized decision path diagram, parsing the intervention results into a standardized intervention paradigm, and transmitting them back in real time to trigger a recalculation of the arbitration strategy includes: The operation trajectory performed by the user on the visual decision path map is captured through the human-computer interaction interface provided by the visual decision path map. The operation trajectory includes node weight adjustment operation, path selection operation and knowledge injection operation. The operation trajectory is semantically parsed to extract the operation type, operation object, and operation parameters corresponding to each operation. The operation type, operation object, and operation parameters are then structured and encoded according to a preset paradigm template to generate the standardized intervention paradigm. The standardized intervention paradigm is fed back to the dynamic arbitration engine in real time. The dynamic arbitration engine then adjusts the corresponding strategy operator parameters in the current arbitration strategy sequence online according to the operation object and operation parameters in the standardized intervention paradigm, thereby triggering the recalculation of the arbitration strategy. The standardized intervention paradigm is synchronously sent to the cross-modal conflict graph builder. The cross-modal conflict graph builder extracts new knowledge content based on the knowledge injection operation in the standardized intervention paradigm, converts the new knowledge content into new knowledge nodes, and updates the dynamic conflict graph.

8. The method of claim 1, wherein, The step of updating the dynamic conflict map and the user cognitive state vector based on the intervention results includes: The operation object and operation parameters are parsed from the standardized intervention paradigm. The target knowledge node in the dynamic conflict graph is located according to the operation object. The node weight of the target knowledge node and the conflict edge weight associated with the target knowledge node are adjusted according to the operation parameters. New knowledge content is extracted from the standardized intervention paradigm, and features are extracted from the new knowledge content to generate new knowledge feature vectors. Event valid timestamps and system entry timestamps are assigned to the new knowledge content as dual temporal markers. The new knowledge content is converted into new knowledge nodes. The semantic similarity between the new knowledge feature vectors and knowledge nodes in the dynamic conflict graph is calculated. Potential conflict relationships are identified based on the semantic similarity, and new conflict edges are constructed between the new knowledge nodes and knowledge nodes in the dynamic conflict graph. Based on the dual temporal markers of all knowledge nodes after the update, knowledge nodes that exceed the valid timestamp of the event are marked as invalid, and the weights of all conflict edges are subject to attenuation constraints based on the timestamps entered by the system, thereby generating the updated dynamic conflict graph. The user behavior pattern corresponding to the operation trajectory is parsed from the standardized intervention paradigm. Based on the user behavior pattern, the user professional level dimension or task urgency dimension in the user cognitive state vector is incrementally corrected to generate an updated user cognitive state vector.

9. The method of claim 1, wherein, The method for determining the conflict detection sensitivity of the dynamic conflict map includes: The user's professional level dimension and task urgency dimension are analyzed from the user's cognitive state vector to obtain the professional level score and task urgency score; Obtain the distribution density of current conflict nodes and the accuracy statistics of historical conflict detection in the dynamic conflict map, and generate a map state complexity index. The professional level score, the task urgency score, and the graph state complexity index are input into the sensitivity adaptive adjustment model, and the initial sensitivity threshold is output through the sensitivity adaptive adjustment model. Based on the distribution of associated compliance attributes of each knowledge unit in the compliance tag set, the initial sensitivity threshold is corrected to generate the conflict detection sensitivity. The conflict detection sensitivity is used to determine whether there is a conflict relationship between knowledge unit pairs during the cross-modal conflict identification process.

10. A complex problem knowledge externalization modeling system based on human-computer dialogue, characterized in that, The system: The acquisition module is used to acquire complex target questions, multi-source cross-modal knowledge data, dialogue context features, user cognitive state features, and data traceability authorization information in human-computer dialogue scenarios. After compliance screening and structured coding, standardized knowledge sets, user cognitive state vectors, and compliance tag sets are obtained. The identification module is used to perform cross-modal semantic unified mapping and conflict identification based on the standardized knowledge set and the compliance tag set through a cross-modal conflict graph builder, and construct a dynamic conflict graph with dual temporal labels. The knowledge node weights of the dynamic conflict graph are constrained by the compliance tag set, and the conflict detection sensitivity of the dynamic conflict graph is dynamically adjusted based on the user cognitive state vector. An encapsulation module is used to encapsulate the decision logic, knowledge source weights, and conflict resolution paths corresponding to the arbitration strategy into a visual decision path graph based on the dynamic conflict graph, the dialogue context features, the user cognitive state vector, and the compliance tag set, through a dynamic arbitration engine. The presentation complexity of the visual decision path graph is adjusted based on the user cognitive state vector, and the arbitration strategy needs to undergo ethical pre-checking. The storage module is used to obtain user intervention results based on the visualized decision path diagram, parse the intervention results into a standardized intervention paradigm, and transmit them back in real time to trigger the recalculation of the arbitration strategy. The intervention paradigm is stored in the example library after ethical pre-check. The example library is used to optimize the arbitration strategy generation logic for similar future conflict scenarios, and to update the dynamic conflict map and the user cognitive state vector simultaneously based on the intervention results.

Images