Problem analysis and processing method and device, electronic equipment, medium and program product

Abstract

This invention provides a problem analysis and processing method, apparatus, electronic device, medium, and program product, applicable to the field of artificial intelligence technology. The method includes: acquiring target problem text to be analyzed; performing type identification and text extraction on the target problem text based on a preset set of problem type tags to obtain a first text fragment and a second text fragment; selecting target knowledge resources from different types of knowledge resources based on the first and second text fragments and performing retrieval to obtain corresponding first and second candidate information; and performing dynamic balancing calculations on the first and second candidate information to obtain processing suggestions. By integrating philosophical analysis methods with an artificial intelligence problem parsing framework, different problem types in mixed inputs can be accurately separated, and appropriate retrieval strategies and generation logics can be adopted for each, thereby avoiding knowledge mismatch and improving the accuracy and persuasiveness of the response.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and more specifically to a problem analysis and processing method, apparatus, device, medium, and program product. Background Technology

[0002] With the deepening development of instant messaging and intelligent dialogue products, existing dialogue processing systems mostly adopt a process of keyword recognition, intent classification, unified retrieval, and template or generative responses. The system segments and quantifies the user's input text to determine a single intent or sentiment, then retrieves information from a knowledge base and outputs a response.

[0003] However, as the application of artificial intelligence continues to expand in depth and breadth, user dialogues are increasingly touching upon the levels of thought and emotion. In real-world scenarios, user input is often no longer singular and pure, but rather a mixture of various components, with significant divergences, particularly at the emotional and intellectual levels. This presents a core challenge for existing technologies: the difficulty in distinguishing and extracting different types of question components (such as factual and normative content) from the same text, resulting in the entire text being processed uniformly. Consequently, the system struggles to adapt to different types of knowledge when retrieving and generating responses—evidence-based data (for fact-finding) and rule / case-based knowledge (for normative consultation) are often mismatched, for example, using rules to answer factual questions or substituting data for value judgments. Furthermore, even when multiple types of information are retrieved, they often rely on fixed weights or simple ranking, lacking a dynamic balancing mechanism, leading to conflicting, one-dimensional, and unstable generated suggestions. Summary of the Invention

[0004] In view of the above problems, embodiments of this application provide problem analysis and processing methods, apparatus, devices, media, and program products.

[0005] According to a first aspect of this application, a problem analysis and processing method is provided, the method comprising: acquiring target problem text to be analyzed; performing type identification and text extraction on the target problem text based on a preset problem type tag set to obtain a first text fragment and a second text fragment, wherein the problem type tag set includes at least factual branch tags and normative branch tags; selecting target knowledge resources from different types of knowledge resources and performing retrieval based on the first text fragment and the second text fragment to obtain corresponding first candidate information and second candidate information; and performing dynamic balancing calculation on the first candidate information and the second candidate information to obtain processing suggestions.

[0006] According to an embodiment of this application, the step of performing dynamic balancing calculations on the first candidate information and the second candidate information to obtain processing suggestions includes: generating a factual evaluation vector based on the first candidate information, wherein the factual evaluation vector includes at least a feasibility score, a risk measure, and an uncertainty index; generating a normative evaluation vector based on the second candidate information, wherein the normative evaluation vector includes at least a normative matching degree, a responsibility spillover measure, and a long-term deviation measure; generating a set of candidate processing schemes based on the first candidate information and the second candidate information; for each candidate processing scheme in the set of candidate processing schemes, calculating a risk cost using the factual evaluation vector, calculating a normative deviation cost using the normative evaluation vector, calculating a corresponding multi-objective balancing cost based on the risk cost, the normative deviation cost, and a preset subjectivity loss cost; and selecting a target processing scheme from the set of candidate processing schemes based on the multi-objective balancing cost, and outputting a corresponding processing suggestion.

[0007] According to an embodiment of this application, the method further includes: calculating a subjectivity state vector based on user historical interaction data; predicting the impact on the subjectivity state vector for each candidate processing scheme in the candidate processing scheme set to obtain a first predicted subjectivity state vector, wherein the first predicted subjectivity state vector includes at least predicted autonomy and predicted dependence; eliminating the corresponding candidate processing scheme when the predicted autonomy is lower than a preset lower threshold and / or the predicted dependence is higher than a preset upper threshold; selecting the target processing scheme based on the processed candidate processing scheme set and the multi-objective balance cost, and outputting a corresponding processing suggestion; or outputting a processing suggestion that satisfies the subjectivity constraint when the processed candidate processing scheme set is empty.

[0008] According to an embodiment of this application, the method further includes: collecting historical conversation text and corresponding behavioral data within a preset time window, and extracting risk features including at least absolute word frequency, emotional intensity index, and short-term high-frequency interaction index; calculating a first risk index and a second risk index based on the risk features, wherein the first risk index is used to characterize idealization bias, and the second risk index is used to characterize impulsive decision-making; and adjusting the risk cost based on the first risk index and / or the second risk index during the calculation of the multi-objective balance cost.

[0009] According to an embodiment of this application, the method further includes: obtaining a weight parameter vector used for dynamic balance calculation during historical sessions to determine a historical weight benchmark vector; generating a current candidate weight parameter vector based on the factual evaluation vector and normative evaluation vector of the current session, and smoothly updating the current candidate weight parameter vector based on the historical weight benchmark vector to obtain an updated weight parameter vector, wherein the smooth update process satisfies that the weight change amplitude of adjacent sessions does not exceed a preset threshold; and calculating the multi-objective balance cost based on the updated weight parameter vector.

[0010] According to an embodiment of this application, the step of performing type identification and text extraction on the target question text includes: calculating factual branch label scores and normative branch label scores for the target question text based on the question type label set, and comparing the factual branch label scores and the normative branch label scores with corresponding preset conditions respectively; and triggering a text extraction operation when both the factual branch label scores and the normative branch label scores meet the corresponding preset conditions to obtain the first text fragment and the second text fragment.

[0011] According to an embodiment of this application, the method further includes: obtaining a multidimensional divergence label set based on the target question text recognition, and matching the multidimensional divergence label set to a target node set in a sentiment divergence graph, wherein the sentiment divergence graph includes multiple nodes and multiple edges, the multiple nodes include at least divergence type nodes, triggering factor nodes, and resolution path nodes, and the edges include at least divergence association edges, the divergence association edges carrying association strength values; performing neighborhood expansion and / or shortest path calculation in the sentiment divergence graph based on the target node set to obtain the target evolution path and the corresponding cumulative weight; and reordering and / or eliminating the candidate processing scheme set based on the cumulative weight, and / or adjusting the multi-objective balance cost based on the cumulative weight.

[0012] According to an embodiment of this application, the method further includes: converting the processed set of candidate processing schemes into a set of candidate intervention actions, the set of candidate intervention actions including multiple candidate intervention actions; predicting a second predicted subject state vector corresponding to each candidate intervention action based on the first predicted subject state vector and the set of candidate intervention actions; and selecting a target intervention action based on the second predicted subject state vector, and outputting a processing suggestion corresponding to the target intervention action.

[0013] According to an embodiment of this application, the method further includes: performing type identification and text extraction on the target question text based on the preset question type tag set to obtain a third text fragment, wherein the question type tag set further includes emotional question tags.

[0014] According to an embodiment of this application, the method further includes: obtaining corresponding third candidate information based on the third text fragment. The step of performing a dynamic balance calculation on the first candidate information and the second candidate information to obtain a processing suggestion includes: performing a dynamic balance calculation on the first candidate information, the second candidate information, and the third candidate information to obtain the processing suggestion.

[0015] A second aspect of this application provides a problem analysis and processing apparatus, comprising: a data acquisition module for acquiring target problem text to be analyzed; a text extraction module for performing type identification and text extraction on the target problem text based on a preset problem type tag set to obtain a first text fragment and a second text fragment, wherein the problem type tag set includes at least factual branch tags and normative branch tags; a knowledge retrieval module for selecting target knowledge resources from different types of knowledge resources based on the first text fragment and the second text fragment, and performing retrieval to obtain corresponding first candidate information and second candidate information; and a problem processing module for performing dynamic balancing calculations on the first candidate information and the second candidate information to obtain processing suggestions.

[0016] A third aspect of this application provides an electronic device comprising: one or more processors; and a memory for storing one or more computer programs, wherein the one or more processors execute the one or more computer programs to implement the steps of the method described above.

[0017] A fourth aspect of this application also provides a computer-readable storage medium having a computer program or instructions stored thereon, which, when executed by a processor, implement the steps of the above-described method.

[0018] The fifth aspect of this application also provides a computer program product, including a computer program or instructions that, when executed by a processor, implement the steps of the above-described method.

[0019] According to embodiments of this application, by performing type identification on the same target question text and extracting first and second text fragments corresponding to factual branch labels and normative branch labels respectively, the system can identify and separate information requests of different natures in a single input, avoiding misunderstanding caused by treating complex questions with a single intent. Furthermore, by selecting different types of knowledge resources and performing retrieval based on different text fragments, factual content can be prioritized for matching objective data, statistics, or verifiable information, while normative content can be prioritized for matching rules, cases, or handling paths, thereby reducing knowledge call mismatches and improving the relevance and interpretability of candidate information. Moreover, by dynamically balancing the first and second candidate information to generate processing suggestions, adaptive fusion between evidence prompts and normative constraints can be achieved, reducing conflicts and inconsistencies caused by different information sources, improving the stability, executability, and overall consistency of output suggestions, and thus improving the processing quality of complex scenario problems and user interaction experience. Attached Figure Description

[0020] The above-mentioned contents, other objects, features and advantages of this application will become clearer from the following description of embodiments with reference to the accompanying drawings, in which:

[0021] Figure 1 The illustrations depict application scenarios of problem analysis and processing methods, apparatuses, devices, media, and program products according to embodiments of this application.

[0022] Figure 2 A flowchart illustrating a problem analysis and processing method according to an embodiment of this application is shown schematically;

[0023] Figure 3 A flowchart illustrating a method for filtering a set of candidate processing schemes according to some exemplary embodiments of this application is shown schematically;

[0024] Figure 4 The flowchart illustrates a method for type identification and text extraction of target question text according to some exemplary embodiments of this application;

[0025] Figure 5 This schematically illustrates a structural block diagram of a problem analysis and processing apparatus according to an embodiment of this application; and

[0026] Figure 6 A block diagram schematically illustrates an electronic device suitable for implementing a problem analysis and processing method according to an embodiment of this application. Detailed Implementation

[0027] The embodiments of this application will now be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of this application. In the following detailed description, numerous specific details are set forth to provide a thorough understanding of the embodiments of this application for ease of explanation. However, it will be apparent that one or more embodiments may be implemented without these specific details. Furthermore, descriptions of well-known structures and technologies are omitted in the following description to avoid unnecessarily obscuring the concepts of this application.

[0028] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of this application. The terms “comprising,” “including,” etc., as used herein indicate the presence of the stated features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0029] All terms used herein (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein are to be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0030] When using expressions such as "at least one of A, B and C", they should generally be interpreted in accordance with the meaning that is commonly understood by those skilled in the art (e.g., "a system having at least one of A, B and C" should include, but is not limited to, a system having A alone, a system having B alone, a system having C alone, a system having A and B, a system having A and C, a system having B and C, and / or a system having A, B and C, etc.).

[0031] With the development of instant messaging, intelligent customer service, and companion-style dialogue products, computer-based dialogue processing systems are widely used to understand, retrieve, and generate responses to user-input questions. Currently, existing systems typically employ intent recognition / keyword extraction, template-based, or generative response processing methods: at the front end, user text is segmented, vectorized, or encoded using sentence vectors; at the understanding layer, the text is classified into single intents or judged for sentiment; at the retrieval layer, candidate information is retrieved from a pre-built knowledge base or external search results; and at the output layer, the retrieved content is summarized, rewritten, or directly concatenated to form a response.

[0032] However, in real-world dialogue scenarios, user input often contains multiple cognitive needs simultaneously. For example, a single text may include inquiries about objective facts, probabilities, or verifiable data (such as feasibility, success rate, supporting data, and the feasibility of a particular communication arrangement), as well as consultations on behavioral norms, solution choices, or rule judgments (such as how to proceed, whether this is appropriate, available processing paths, how responsibilities should be allocated, and what boundary settings are more suitable). Existing technologies primarily rely on single-label intents, making it difficult to identify and differentiate these different types of components within the same target question text. Furthermore, they lack the ability to extract fine-grained fragments from the same text to obtain text segments corresponding to different types, resulting in the system only being able to process the entire text in a uniform manner.

[0033] In terms of knowledge retrieval, existing dialogue systems typically employ a single knowledge source or a unified retrieval strategy to retrieve information from multiple knowledge sources in a hybrid manner. Since factual questions rely more on verifiable evidence such as objective data, research conclusions, and statistical information, while normative questions depend more on interpretable decision-making frameworks such as rule sets, decision trees, and typical cases, a unified retrieval and generation approach can easily lead to mismatches, such as answering facts with rules or substituting value judgments with data. This results in one-sided, unverifiable, or unenforceable responses.

[0034] Furthermore, even if the system can recall multiple types of candidate information separately, existing technologies mostly employ simple relevance ranking or fixed-weight fusion, directly splicing information from different sources for output. This lacks a dynamic balancing mechanism to address conflicts or differences in emphasis between different types of candidate information. For example, factual candidate information may indicate high risk or low feasibility, while normative candidate information may emphasize responsibility, boundaries, or handling paths. Without dynamic balancing calculations, the system struggles to synthesize consistent processing recommendations, easily leading to contradictory suggestions, bias towards a single dimension, or instability due to fluctuations in retrieval noise.

[0035] Based on this, embodiments of this application provide a problem analysis and processing method, the method comprising: acquiring target problem text to be analyzed; performing type identification and text extraction on the target problem text based on a preset problem type tag set to obtain a first text fragment and a second text fragment, wherein the problem type tag set includes at least factual branch tags and normative branch tags; selecting target knowledge resources from different types of knowledge resources and performing retrieval based on the first text fragment and the second text fragment to obtain corresponding first candidate information and second candidate information; and performing dynamic balancing calculation on the first candidate information and the second candidate information to obtain processing suggestions. According to embodiments of this application, by performing type identification on the same target question text and extracting first and second text fragments corresponding to factual branch labels and normative branch labels respectively, the system can identify and separate information requests of different natures in a single input, avoiding misunderstanding caused by treating complex questions with a single intent. Furthermore, by selecting different types of knowledge resources and performing retrieval based on different text fragments, factual content can be prioritized for matching objective data, statistics, or verifiable information, while normative content can be prioritized for matching rules, cases, or handling paths, thereby reducing knowledge call mismatches and improving the relevance and interpretability of candidate information. Moreover, by dynamically balancing the first and second candidate information to generate processing suggestions, adaptive fusion between evidence prompts and normative constraints can be achieved, reducing conflicts and inconsistencies caused by different information sources, improving the stability, executability, and overall consistency of output suggestions, and thus improving the processing quality of complex scenario problems and user interaction experience.

[0036] It should be noted that the problem analysis and processing methods, apparatus, devices, media and program products provided in the embodiments of this application can be used in the fields of big data technology, artificial intelligence technology and philosophy, that is, the technical means in the fields of big data technology and artificial intelligence technology can be applied to the field of philosophy.

[0037] In the technical solution of this application, the user information (including but not limited to user personal information, user image information, user device information, such as location information) and data (including but not limited to data used for analysis, stored data, and displayed data) involved are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, storage, use, processing, transmission, provision, disclosure, and application of related data all comply with relevant laws, regulations, and standards, take necessary confidentiality measures, do not violate public order and good morals, and provide corresponding operation entry points for users to choose to authorize or refuse.

[0038] In scenarios where personal information is used for automated decision-making, the methods, devices, and systems provided in this application all offer users corresponding operation entry points, allowing them to choose to agree to or reject the automated decision results; if the user chooses to reject, the process proceeds to expert decision-making. Here, "expert decision-making" refers to the decision-making activities of personnel who specialize in a particular field, possess specialized experience, knowledge, and skills, and have reached a certain level of professional expertise.

[0039] Figure 1 The illustrations depict application scenarios of problem analysis and processing methods, apparatuses, devices, media, and program products according to embodiments of this application.

[0040] like Figure 1 As shown, application scenario 100 according to this embodiment may include a first terminal device 101, a second terminal device 102, a third terminal device 103, a network 104, and a server 105. The network 104 serves as a medium for providing a communication link between the first terminal device 101, the second terminal device 102, the third terminal device 103, and the server 105. The network 104 may include various connection types, such as wired or wireless communication links, or fiber optic cables, etc.

[0041] Users can use the first terminal device 101, the second terminal device 102, and the third terminal device 103 to interact with the server 105 via the network 104 to receive or send messages, etc. Various communication client applications can be installed on the first terminal device 101, the second terminal device 102, and the third terminal device 103, such as shopping applications, web browser applications, search applications, instant messaging tools, email clients, social media platform software, etc. (for example only).

[0042] In the embodiments of this application, the first terminal device 101 can be an example of the first device, and the second terminal device 102 and / or the third terminal device 103 can be an example of at least one second device. The first device and the second device can communicate collaboratively through an internal client mechanism to implement the data distribution and rendering logic described in the problem analysis and processing method.

[0043] In some embodiments, the first device and at least one second device may be different display modules, windows or screens on the same computing terminal (such as a host), or multiple physical devices that work together through a network, such as different client instances deployed on a desktop computer, tablet terminal or mobile device respectively.

[0044] The first terminal device 101, the second terminal device 102, and the third terminal device 103 can be various electronic devices with displays and support web browsing, including but not limited to smart mobile terminals, tablet computers, laptop computers, and desktop computers.

[0045] Server 105 can be a server that provides various services, such as a backend management server that supports websites browsed by users using the first terminal device 101, the second terminal device 102, and the third terminal device 103 (this is just an example). The backend management server can analyze and process data such as received user requests, and feed back the processing results (such as web pages, information, or data obtained or generated according to user requests) to the terminal devices.

[0046] It should be noted that the problem analysis and processing methods provided in the embodiments of this application can generally be executed by server 105. Correspondingly, the problem analysis and processing apparatus provided in the embodiments of this application can generally be located in server 105. The problem analysis and processing methods provided in the embodiments of this application can also be executed by a server or server cluster that is different from server 105 and capable of communicating with the first terminal device 101, the second terminal device 102, the third terminal device 103, and / or server 105. Correspondingly, the problem analysis and processing apparatus provided in the embodiments of this application can also be located in a server or server cluster that is different from server 105 and capable of communicating with the first terminal device 101, the second terminal device 102, the third terminal device 103, and / or server 105.

[0047] It should be understood that Figure 1 The number of terminal devices, networks, and servers shown is merely illustrative. Depending on implementation needs, any number of terminal devices, networks, and servers can be included.

[0048] The following will be based on Figure 1 The described scene, through Figures 2-4 The problem analysis and processing methods of the embodiments of the present invention will be described in detail.

[0049] Figure 2 A flowchart illustrating a problem analysis and processing method according to an embodiment of this application is shown schematically.

[0050] like Figure 2 As shown, the problem analysis and processing method 200 of this embodiment includes operations S210 to S240.

[0051] In operation S210, the target problem text to be analyzed is obtained.

[0052] In the embodiments of this application, the target question text can be directly typed and submitted by the user in the input box of the mobile application. After receiving the submission event, the system takes the input content as the text to be analyzed and generates a session identifier, submission timestamp and source channel identifier so that subsequent steps can associate, track and trace the same round of input.

[0053] In some embodiments, the target problem text may originate from an instant messaging session. The system receives multiple messages sent by the user in the chat window through the message access interface and merges consecutive messages that meet the preset time interval condition into a single piece of text to be analyzed. The original message boundary markers are retained during merging, so that when key statements need to be located in subsequent processing, the specific message fragment can be traced back.

[0054] In the embodiments of this application, the target question text received by the system can carry metadata to form a text object, wherein the metadata may include language identifier, character encoding and source type; when a non-default language or mixed language is detected, the system selects the corresponding word segmentation configuration, stop word list or encoder configuration according to the language identifier.

[0055] In some embodiments, the system can recover the text to be analyzed from local drafts or historical input cache under privacy constraints. For example, if a user fails to submit due to network interruption, the application can read the locally encrypted draft content as the text to be analyzed after restarting, along with the draft generation time, to determine whether it is necessary to prompt the user that the content may no longer be applicable to the current context.

[0056] In embodiments of this application, the system can jointly construct the text to be analyzed from the current round of input and the dialogue context. Specifically, the system can read the interaction content between the user and the system from the most recent rounds in the session cache, and combine the context with the current input according to a preset splicing template; the template may include speaker tags to support subsequent steps in processing pronouns, omissions, and succession relationships.

[0057] In some embodiments, the system can de-identify the input text before proceeding to subsequent steps: performing mask replacement on sensitive segments such as phone numbers, addresses, and account numbers, while retaining entity category placeholders to maintain semantic clues; the de-identified text is used as the text to be analyzed for subsequent type identification and information retrieval, thereby reducing information loss while protecting privacy.

[0058] In operation S220, based on a preset set of question type labels, the target question text is identified and extracted to obtain a first text fragment and a second text fragment. The set of question type labels includes at least factual branch labels and normative branch labels.

[0059] In the embodiments of this application, type recognition can be implemented using a multi-label discrimination method. Specifically, the system inputs the target question text into a multi-label classifier, outputs the confidence scores of factual branches and normative branches, and when both confidence scores reach the threshold, the entire text is simultaneously labeled as a factual branch label and a normative branch label to ensure that the two types of text fragments can be extracted in parallel.

[0060] In some embodiments, when the type identification result only meets the preset conditions of the factual branch label and does not meet the preset conditions of the normative branch label, the system can determine the entire target question text as the first text fragment and set the second text fragment as an empty fragment; in subsequent steps, only the first text fragment is used to perform retrieval and generate candidate information, while the normative branch contribution in the dynamic balance calculation is set to zero or the second candidate information is ignored.

[0061] In some embodiments, when the type identification result only meets the preset conditions of the normative branch label but not the preset conditions of the factual branch label, the system can determine the entire target question text as the second text fragment and set the first text fragment as an empty fragment; in subsequent steps, the target knowledge resource is selected and the retrieval is performed based only on the second text fragment, while the contribution of the factual branch in the dynamic balance calculation is set to zero or the first candidate information is ignored.

[0062] In the embodiments of this application, in order to avoid insufficient subsequent search results due to empty fragments, when only one type is identified, the system can generate placeholder fragments for another type. The placeholder fragments can be preset neutral query templates or domain keyword sets. The system performs supplementary searches in the corresponding type knowledge resources based on the placeholder fragments to obtain auxiliary candidate information for explaining or constraining the output, and applies low weight to the auxiliary candidate information during dynamic balancing calculation.

[0063] In the embodiments of this application, type identification and extraction can be completed in one step using an end-to-end sequence labeling model. Specifically, the model can output category labels in the label set for each character or word, and the system directly generates the first text segment and the second text segment based on the span boundary of the consecutive labels, thereby avoiding the accumulation of errors caused by classifying first and then extracting.

[0064] In some embodiments, the system can first obtain the attention weight distribution of factual and normative information, and then construct corresponding segments based on the several statements or clauses with the highest weights of each; when some statements have high attention weights of both types, they are copied to the two segments and a shared tag is attached so that the statements can be used to trigger the retrieval of different knowledge resources in subsequent searches.

[0065] In the embodiments of this application, the system can maintain different feature dictionaries and triggering patterns for factual and normative branches respectively. For example, the factual branch may focus more on patterns such as numerical units, probability words, verification words, and comparison structures; the normative branch may focus more on patterns such as obligation words, permission words, boundary words, and scheme selection structures; during extraction, the matched triggering pattern is used as an anchor point to expand the preset window to the left and right to generate fragments, thereby improving the stability of extraction.

[0066] In some embodiments, type recognition can be enhanced based on the dialogue context. Specifically, the system can concatenate and encode the current target question text with the previous system response or the previous user supplementary explanation to reduce ambiguity using context; for example, if the user only enters "Then what should I do?", the system can jointly determine that the current text belongs to two types of tags based on the ratio of factual and normative discussions in the previous round, and extract the factual conditions and normative constraints associated with "what should I do?" respectively.

[0067] Furthermore, based on a preset set of question type labels, the target question text can be type-identified and text extracted to obtain a first text fragment, a second text fragment, and a third text fragment. The question type label set includes at least factual question labels, normative question labels, and sentiment question labels.

[0068] In some embodiments, the extraction of the third text fragment can be enhanced based on emotion trigger signals. Specifically, the system can identify emotion words, intensity adverbs, negation structures, rhetorical questions, punctuation patterns such as exclamation marks / ellipses, and non-lexical features such as emoticons in the target question text, and use them as trigger anchors for emotional branches; the system can expand a preset window to the left and right with the trigger anchor as the center to generate a third text fragment containing emotional expression and relational appeal.

[0069] In some embodiments, when a user expresses multiple emotions or switches between emotion objects within the same target question text, the system can split the third text segment into multiple sub-segments and attach object identifiers. Specifically, the system identifies sub-segments corresponding to different emotion objects based on emotion object references (such as "I," "he," "you," "company," "doctor," etc.) and sentence transition words, and assigns object identifiers and priorities to each sub-segment to avoid confusing the object of reassurance and leading to inappropriate responses during subsequent empathy retrieval and generation.

[0070] In operation S230, based on the first text fragment and the second text fragment, target knowledge resources are selected from different types of knowledge resources and retrieval is performed to obtain the corresponding first candidate information and second candidate information.

[0071] In the embodiments of this application, the system can maintain different resource routing rules for the first text fragment and the second text fragment. Specifically, when the first text fragment is identified as a factual branch, the system routes it to a structured or semi-structured data source containing statistical data, research conclusions, or experimental results. For example, this data source may include a scientific database, which can be an internally built research data aggregation library, a publicly available scientific research data mirror library, or a set of third-party data interfaces, used to provide verifiable experimental conclusions, statistical results, indicator ranges, or evidence items. When the second text fragment is identified as a normative branch, the system routes it to a data source containing rule clauses, case summaries, or handling procedures, thereby achieving resource type differentiation during the retrieval stage. For example, this data source may include a moral philosophy theory library, which may contain normative theoretical frameworks, value conflict analysis templates, argumentative structures for responsibility attribution and boundary setting, and case-based argumentative fragments corresponding to different situations, used to support the interpretation and solution generation of normative issues.

[0072] In the embodiments of this application, a philosophical approach to moral divergence analysis is introduced to deconstruct user questions into factual inquiry and normative consultation. By integrating philosophical analysis methods with an artificial intelligence problem-solving framework, different question types in mixed inputs can be accurately identified, and appropriate retrieval strategies and generation logics can be applied to each type, thereby avoiding knowledge mismatch and improving the accuracy and persuasiveness of the responses.

[0073] In some embodiments, the system can configure different retrieval index structures for different types of knowledge resources. For example, knowledge resources corresponding to factual branches use numerical field indexes, time indexes, or entity-attribute indexes to support precise filtering; knowledge resources corresponding to normative branches use topic indexes or case tag indexes to support scenario matching, and the system calls the corresponding indexes to generate candidate information during retrieval.

[0074] In embodiments of this application, the system can generate a structured query request based on a first text fragment. Specifically, entity, condition, and constraint information is extracted from the first text fragment and mapped to query parameters submitted to a factual knowledge resource interface. Simultaneously, the system can generate a semantic query request based on a second text fragment and submit the query request to a rule base or case base to obtain normative candidate information.

[0075] In some embodiments, the system employs different similarity calculation methods for different types of knowledge resources. Specifically, for factual branches, the system uses exact matching or range matching to calculate the relevance between the query and the data entry; for normative branches, the system uses semantic similarity or topic consistency scoring to rank candidate rules or cases, and forms first candidate information and second candidate information respectively.

[0076] In the embodiments of this application, the system can select target knowledge resources based on the reliability level of the knowledge resources. For factual branches, the system prioritizes authoritative data sources or verified datasets; for normative branches, the system prioritizes rules or case resources that have been manually compiled or reviewed; when the preferred resource does not return results, the system calls secondary resources according to a preset priority order.

[0077] In some embodiments, the retrieval processes corresponding to the first and second text fragments can be processed in parallel. Specifically, the system allocates factual branch retrieval tasks and normative branch retrieval tasks to different execution threads or service instances. Once either branch retrieval is completed, the corresponding candidate information can be returned, thereby reducing the overall response latency.

[0078] In the embodiments of this application, when the number of candidate information returned by the knowledge resource corresponding to a certain text fragment is lower than a preset threshold, a supplementary retrieval strategy can be triggered. Specifically, for factual branches, the system expands the time range or condition tolerance; for normative branches, the system introduces rules or cases of adjacent scenarios or similar roles, and marks the supplemented candidate information as auxiliary candidate information.

[0079] In some embodiments, the system dynamically adjusts the knowledge resource selection strategy based on user configuration or contextual preferences. Specifically, when the system detects that the user prefers authoritative data, it prioritizes official data sources in the factual branch; when the system detects that the user prefers operational guidance, it prioritizes process-oriented rule bases in the prescriptive branch, thereby making the candidate information more aligned with the current interaction needs.

[0080] For embodiments that include a third text fragment, corresponding third candidate information can be obtained based on an empathy model. Specifically, the emotion category and intensity level can be obtained based on the third text fragment; then, candidate strategy entries that match the emotion category, intensity level, and scene label can be retrieved from the empathy model or empathy strategy library; finally, the expression framework in the strategy entries and the problem context are filled into slots to form third candidate information, thereby reducing the risk of tone drift caused by direct generation.

[0081] In this embodiment, the empathy model can complement knowledge resource retrieval. Specifically, when a third text fragment triggers strong emotions and there is insufficient factual / normative branch information, the system can prioritize outputting clarifying empathy strategy points from the empathy model, and use these points as supplementary query conditions to inversely constrain factual or normative retrieval (e.g., prioritizing the retrieval of rules and cases on "how to explain uncertainty, how to give boundaries"), thereby forming synergy at the candidate information level.

[0082] In other words, embodiments of this application propose an intelligent dialogue system architecture that integrates emotional reasoning and factual decision-making. It constructs a closed-loop framework encompassing in-depth problem analysis, dynamic adaptation of multi-source knowledge, and personalized emotional decision-making, effectively addressing the problems of traditional dialogue systems in complex scenarios such as psychological counseling and business ethics, where they prioritize facts over emotion or where logic and empathy are disconnected. Specifically, embodiments of this application utilize multi-dimensional intent decoupling technology to separate factual needs and emotional appeals in user questions, achieving accurate component identification; through a dynamic knowledge graph fusion mechanism, it optimizes response strategies by combining domain knowledge bases with real-time emotional context; and by employing an ethical conflict resolution engine, it introduces a value alignment module to ensure that decisions conform to ethical norms and social consensus.

[0083] It should be noted that the method according to the embodiments of this application has been piloted and verified in multiple scenarios such as psychological counseling and business ethics counseling. The verification results show that it can significantly improve the ability to deal with complex problems. Through actual testing, the method according to the embodiments of this application has at least the following technical effects: accurate component recognition: the accuracy of dialogue text parsing is improved by 42%; knowledge adaptation optimization: through feedback from more than 2,000 users, the comprehensive score of response relevance and emotional fit is improved by 65%; conflict resolution ability: in ethical dilemma dialogues, the effective guidance to reach consensus reaches 83%.

[0084] In operation S240, a dynamic balance calculation is performed on the first candidate information and the second candidate information to obtain processing suggestions.

[0085] In the embodiments of this application, dynamic equilibrium calculation can be achieved by constructing factual evaluation vectors, normative evaluation vectors, or optional affective evaluation vectors. The system summarizes and quantifies the features of the first candidate information, the second candidate information, or the optional third candidate information, forming a multi-dimensional evaluation representation for subsequent calculations. This allows candidate information from different sources and with different structures to be compared and integrated in a unified numerical space. When generating factual evaluation vectors, the system can comprehensively consider the elements in the first candidate information that reflect feasible conditions, potential risks, and the degree of certainty of information, and map them into multiple evaluation dimensions. Correspondingly, when generating normative evaluation vectors, the system quantifies the content in the second candidate information that reflects rule adaptability, the scope of responsibility, and long-term impact trends to support subsequent comprehensive evaluation.

[0086] In the embodiments of this application, dynamic balancing calculation can be achieved by evaluating the content completeness of the first candidate information, the second candidate information, and, optionally, the third candidate information. Specifically, the system calculates the semantic coverage of the first candidate information, the second candidate information, and, optionally, the third candidate information to the original target question text. When the coverage of any candidate information is significantly insufficient, its participation in the processing suggestion generation is reduced, thereby avoiding biased conclusions caused by incomplete information.

[0087] In some embodiments, dynamic balancing calculations can be performed based on information consistency constraints. Specifically, the system detects whether there are explicit conflicting statements among the first candidate information, the second candidate information, or the optional third candidate information. For example, factual candidate information may deny a certain condition while normative candidate information may presuppose that condition. When a conflict is detected, the system marks the conflicting part and prioritizes conditional or branching suggestions when generating processing recommendations to reduce the risk of being misled by direct conflicts.

[0088] In the embodiments of this application, the dynamic balancing calculation adopts a hierarchical fusion strategy. Specifically, the system first integrates factual candidate information to form a unified factual description result; then it integrates normative candidate information to form a unified rule or disposal path description; finally, it performs fusion between the two types of integrated results, thereby reducing noise interference caused by an excessive number of original candidate information.

[0089] In some embodiments, dynamic balancing calculations are adjusted in conjunction with user preference parameters. Specifically, the system dynamically adjusts the presentation ratio of the first candidate information, the second candidate information, or the optional third candidate information in the processing suggestions based on the user's clicks, pauses, or feedback behavior on data-explanatory or rule-guided content in historical interactions, making the output more in line with individual usage habits.

[0090] In embodiments of this application, customer types can also be introduced for dynamic balancing calculations. Customer types can be determined using a multi-label approach. Specifically, customer types can be determined based on customer profiles, interaction history, or real-time identification results. For example, customer types include rational customers, emotional customers, and balanced customers. The system can assign basic weights and upper limit constraints based on three branches: factual, normative, and emotional, respectively. The following table is an example.

[0091] Table 1 Customer Type Mapping Table

[0092]

[0093] In a further embodiment, a multi-channel processing mechanism can be employed to perform parallel parsing and handling of user questions. Specifically, after type identification and text extraction, the system allocates the same target question text to a factual processing channel, a normative processing channel, and an emotional processing channel, respectively, and maintains independent retrieval, evaluation, and generation processes for each channel. The factual processing channel focuses on summarizing objective conditions, verifiable conclusions, and risk factors; the normative processing channel focuses on deriving rule boundaries, responsibility constraints, and handling paths; and the emotional processing channel focuses on analyzing emotional states, psychological needs, and communication strategies. Through multi-channel processing, attention competition and information interference caused by mixed reasoning within the same channel can be avoided, making the recall of candidate information and subsequent dynamic balancing calculations more stable and controllable.

[0094] In some embodiments, the system can further employ a knowledge source isolation strategy for the three types of processing channels to achieve type adaptation of knowledge invocation. Specifically, the system routes factual channels to objective data sources and selects target knowledge resources based on credibility. For example, objective data sources may include authoritative statistical databases, experimental conclusion databases, or structured indicator datasets. The system routes normative channels to expert knowledge bases and selects target knowledge resources based on authority. For example, expert knowledge bases may include manually compiled rule and clause databases, case summary databases, or process-oriented handling databases. The system routes emotional channels to mental model databases and selects target knowledge resources based on empathy. For example, mental model databases may include empathy models, emotion-soothing strategy databases, or communication script template databases. By configuring appropriate knowledge sources and priority rules for different types of questions, the risk of knowledge mismatch can be reduced, the relevance of candidate information can be improved, and a more reliable input basis can be provided for subsequent fusion to generate processing suggestions that take into account facts, norms, and emotions.

[0095] According to embodiments of this application, by acquiring the target question text to be analyzed and performing type identification and text extraction based on a set of question type labels, the system can distinguish and retain information components of different natures in the same input, avoiding misunderstanding caused by treating complex questions with a single intent. Furthermore, by selecting different types of knowledge resources and performing retrieval based on the extracted first and second text fragments, it helps to prioritize matching factual content with objective data or verifiable information, and normative content with rules or processing paths, thereby reducing knowledge call mismatches and improving the relevance of candidate information. At the same time, by dynamically balancing the first and second candidate information to generate processing suggestions, adaptive fusion can be performed between multiple types of information, reducing information conflicts and single-dimensional bias, improving the consistency, stability, and executability of the output results, thereby improving the overall analysis and processing effect of complex target question texts and the user interaction experience.

[0096] The problem analysis and processing methods of the embodiments of this application will be specifically described below by way of preferred embodiments.

[0097] In the embodiments of this application, dynamic balancing calculation is used to form a consistent processing recommendation between factual candidate information and normative candidate information. Specifically, the system can generate a factual evaluation vector based on the first candidate information. The factual evaluation vector includes at least a feasibility score, a risk measure, and an uncertainty index, used to characterize the objective feasibility conditions, potential risk level, and degree of information certainty reflected by the first candidate information. The system can also generate a normative evaluation vector based on the second candidate information. The normative evaluation vector includes at least a normative matching degree, a responsibility spillover measure, and a long-term deviation measure, used to characterize the degree of adaptation of the second candidate information to the preset normative framework, the scope of its impact on the responsibilities of relevant parties, and the deviation trend from long-term behavioral norms.

[0098] In the embodiments of this application, a set of candidate processing schemes can be further generated based on the first candidate information and the second candidate information. Specifically, each candidate processing scheme in the set can be obtained by combining the factual conditions and limiting factors extracted from the first candidate information with the rule constraints and disposal paths extracted from the second candidate information, so that each candidate processing scheme has both executable factual basis and normative constraints. For each candidate processing scheme in the set, the system uses a factual evaluation vector to calculate the corresponding risk cost, uses a normative evaluation vector to calculate the corresponding normative deviation cost, and on this basis, combines a preset subjectivity loss cost to calculate a multi-objective balance cost, so that each candidate processing scheme is uniformly measured in three dimensions: factual risk, normative consistency, and subjectivity impact.

[0099] In the embodiments of this application, the system can select a target processing scheme from a set of candidate processing schemes based on a multi-objective balanced cost, and output a processing suggestion corresponding to the target processing scheme. Specifically, the system can determine the candidate processing scheme with the minimum multi-objective balanced cost as the target processing scheme, or select the candidate processing scheme with the optimal multi-objective balanced cost from a set of candidate processing schemes that meet preset constraints as the target processing scheme, and convert it into a user-oriented processing suggestion output, so as to ensure the executability of the suggestion while taking into account risk control and standardization consistency.

[0100] In the embodiments of this application, in order to further constrain and screen the candidate processing scheme set after generating the candidate processing scheme set and calculating the multi-objective balance cost, so as to reduce the risk of user state imbalance caused by output suggestions and improve the acceptability and stability of processing suggestions, the following method is also provided.

[0101] Figure 3 The flowchart illustrates a method for screening a set of candidate processing schemes according to some exemplary embodiments of this application.

[0102] like Figure 3 As shown, the method for screening the candidate processing scheme set includes operations S310 to S340.

[0103] In operation S310, the subjectivity state vector is calculated based on the user's historical interaction data.

[0104] Historical interaction data may include: historical conversation text, message sending timestamps, number of messages, interaction frequency, dialogue object identifiers (e.g., "conversations related to a specific object"), and user feedback on historical suggestions (e.g., whether they were adopted, duration of stay, likes / dislikes, etc.). The system can perform feature processing on historical interaction data to obtain feature vectors used to calculate the subjective state.

[0105] For example, the feature vector may include the following computable features: interaction intensity features, such as interaction frequency within a preset window (e.g., number of messages per hour, number of conversation rounds per day), interaction concentration (e.g., peak frequency that suddenly increases in a short period of time); content distribution features, such as the proportion of self-active topics and the proportion of topics related to the other party; request and confirmation features, such as the frequency of repeated confirmation / reassuring statements; suggestion adoption behavior features, such as historical suggestion adoption rate, adoption delay, and changes in interaction intensity after adoption.

[0106] In embodiments of this application, a subjectivity state vector can be calculated based on the aforementioned feature vectors. The subjectivity state vector includes autonomy and dependence, and can be normalized to map autonomy and dependence to a uniform range (e.g., 0–1 or 0–100). For example, the system can calculate autonomy and dependence using a preset mapping function or scoring model, respectively.

[0107] In operation S320, for each candidate processing scheme in the candidate processing scheme set, the impact on the subjectivity state vector is predicted to obtain the first predicted subjectivity state vector, which includes at least the predicted autonomy and the predicted dependence.

[0108] In some embodiments, candidate processing schemes can first be transformed into computable scheme feature vectors. The scheme feature vectors can be extracted from elements such as action type, constraint type, and guidance method in the scheme text. For example, whether it contains action elements such as "encouraging self-activity," "boundary setting," "delaying decision-making / reducing interaction intensity," or "overly accommodating / completely relying on the other party's feedback." Subsequently, the system inputs the current subjectivity state vector and the scheme feature vector into a preset state transition function or prediction model, outputting a predicted subjectivity state vector. The prediction model can be a linear mapping, a tree model, or a neural network, or a regularized incremental update function.

[0109] In operation S330, if the prediction autonomy is lower than the preset lower threshold and / or the prediction dependence is higher than the preset upper threshold, the corresponding candidate processing schemes are eliminated.

[0110] In operation S340, a target processing scheme is selected based on the processed candidate processing scheme set and the multi-objective balance cost, and the corresponding processing suggestion is output; or when the processed candidate processing scheme set is empty, a processing suggestion that satisfies the subjectivity constraint is output.

[0111] In the embodiments of this application, when candidate processing schemes still exist after elimination processing, a target processing scheme can be selected based on the processed candidate processing scheme set and the corresponding multi-objective balance cost, and a processing suggestion can be output. If the processed candidate processing scheme set is empty, the system outputs a processing suggestion that satisfies the subjectivity constraint, such as outputting a more conservative action suggestion, a suggestion to reduce the interaction intensity, or a suggestion to guide supplementary information, to ensure that the output does not violate the preset threshold constraint.

[0112] By converting historical interaction data into computable features and further calculating the subjectivity state vector and its predicted state under different candidate processing schemes, state-gated screening of candidate schemes is achieved. Compared with existing technologies that rely solely on relevance ranking or fixed-weight fusion, the embodiments of this application can eliminate schemes that may lead to a decrease in autonomy or an increase in dependence before output, and provide alternative suggestions that meet constraints when no feasible schemes are available. This improves the security, stability, and acceptability of processing suggestions and reduces the inconsistency risk caused by policy fluctuations in multi-turn dialogue scenarios.

[0113] In the embodiments of this application, in order to quantify the user's decision-making risk and incorporate it into dynamic balance calculations when generating processing suggestions, historical conversation text and corresponding behavioral data can be collected within a preset time window, and risk features can be extracted from them. Specifically, historical conversation text may include the user's input content in multi-turn dialogues, and behavioral data may include message sending timestamps, interaction frequency, continuous interaction duration, and interaction time distribution, etc. The system counts the frequency of occurrence of absolute words from historical conversation text. Absolute words can be words or phrases in a preset vocabulary list used to represent "uniqueness, inevitability, and extremeness." The system can also calculate emotion intensity indicators based on an emotion dictionary, emotion classification model, or emotion intensity regression model. At the same time, the system extracts short-term high-frequency interaction indicators based on behavioral data, such as the number of times or duration of message messages exceeding a preset threshold within a unit of time, thereby forming a set of risk features for assessing decision-making risk.

[0114] In embodiments of this application, the system can calculate a first risk index and a second risk index based on risk characteristics. The first risk index characterizes idealization bias and can be correlated with features such as the frequency of absolute terms, the proportion of overly positive evaluative expressions, and their stability within a time window. The second risk index characterizes impulsive decision-making and can be correlated with features such as emotion intensity indicators, short-term high-frequency interaction indicators, and abrupt changes in interaction rhythms. These risk indices can be generated through weighted summation, piecewise mapping, or a trained scoring model, and can be further normalized to a unified dimension so that they can participate in the calculation along with subsequent cost terms.

[0115] In the embodiments of this application, the system can use a first risk index and / or a second risk index to adjust the risk cost during the calculation of multi-objective balance costs. Specifically, the risk cost can be calculated from the aforementioned factual evaluation vector to reflect the objective risk level of candidate processing solutions. The system applies an adjustment coefficient to the risk cost and / or adjusts its weight in the multi-objective balance costs based on the first risk index and / or the second risk index, so that when the risk of idealization bias or impulsive decision-making increases, the proportion of the risk cost in the comprehensive evaluation increases, thereby prompting the target processing solutions to be more inclined to robust and controllable suggestion outputs. In this way, the system can introduce the risk state reflected by multi-turn dialogue and interaction behavior into dynamic balance calculation, reducing the probability of outputting radical or unstable suggestions under high-risk conditions.

[0116] In the embodiments of this application, to improve the consistency and stability of processing suggestions in multi-turn dialogue scenarios, the weight parameters used in dynamic balancing calculations can be subject to temporal continuity control. Specifically, the weight parameter vector used for dynamic balancing calculations can be obtained from historical session records, and a historical weight baseline vector can be determined accordingly. The historical weight baseline vector can be obtained statistically from historical weight parameter vectors within a preset time window, such as by using a weighted average, moving average, or by using the most recent weight parameter vector as the baseline vector, thereby reflecting the system's overall balance preference and output direction in historical interactions.

[0117] In the embodiments of this application, a current candidate weight parameter vector can be generated based on the factual evaluation vector and the normative evaluation vector of the current session. Subsequently, the current candidate weight parameter vector can be smoothly updated based on the historical weight benchmark vector to obtain the updated weight parameter vector. The smooth update can be implemented by linear interpolation, exponential moving average, or limit update, so that the updated weight parameter vector takes into account both the input characteristics of the current session and the output continuity of the historical sessions.

[0118] In the embodiments of this application, the smooth update process ensures that the weight change amplitude of adjacent sessions does not exceed a preset threshold. Specifically, the system can constrain the change of each component of the weight parameter vector before and after the update. When the change of a certain component exceeds the preset threshold, the change of that component is pruned to within the threshold range, thereby suppressing drastic weight changes caused by single input fluctuations or retrieval noise. After obtaining the updated weight parameter vector, the system calculates the multi-objective balance cost based on the updated weight parameter vector, and evaluates and ranks the candidate processing schemes accordingly, thereby maintaining the continuity of the processing suggestion output direction and improving the user experience during multiple rounds of interaction.

[0119] In the embodiments of this application, in order to solve the problem that the prior art has difficulty in accurately distinguishing and extracting the corresponding information when the same target problem text contains both factual claims and normative claims, the following method is also provided.

[0120] Figure 4 The flowchart illustrates a method for type identification and text extraction of target question text according to some exemplary embodiments of this application.

[0121] like Figure 4 As shown, the method for type identification and text extraction of target question text may include operations S410~S420.

[0122] In operation S410, based on the problem type label set, the factual branch label score and normative branch label score are calculated for the target problem text, and the factual branch label score and normative branch label score are compared with the corresponding preset conditions.

[0123] Specifically, the label score can be formed by the confidence level output by the text classification model, the weighted score obtained by rule matching, or a comprehensive score obtained by fusing the two. For example, the system can vectorize the target question text and input it into a multi-label classifier to obtain the probability values ​​of the two types of labels; or it can improve the factual branch label score for segments containing features such as "probability, data, feasibility, verification" and improve the normative branch label score for segments containing features such as "should, appropriate, how to handle, boundary, responsibility". Subsequently, the system compares the factual branch label score with a first preset condition and the normative branch label score with a second preset condition, where the preset conditions can be a score threshold, a confidence threshold, or a comprehensive judgment condition combining text length and the number of trigger word hits.

[0124] In operation S420, if both the factual branch label score and the normative branch label score meet the corresponding preset conditions, the text extraction operation is triggered to obtain the first text fragment and the second text fragment.

[0125] In the embodiments of this application, text extraction can be performed at the sentence level or clause level: the system can first segment the target question text into sentences, and calculate factual and normative scores for each sentence. Sentences with higher factual scores exceeding a preset threshold are assigned to the first text segment, and sentences with higher normative scores exceeding a preset threshold are assigned to the second text segment. For sentences containing both types of features, the system can further segment them into clauses based on conjunctions, punctuation, or dependency relationships before categorizing them, thereby improving the type purity of the segments. After extraction, the system obtains the first text segment corresponding to the factual branch and the second text segment corresponding to the normative branch, and can attach segment-level confidence labels to both segments for reference in subsequent steps.

[0126] In some embodiments, when neither the factual branch label score nor the normative branch label score meets the corresponding preset conditions, the system can determine the target question text as a low-confidence input and trigger the generation of a clarification prompt. When only the factual branch label score meets the corresponding preset conditions while the normative branch label score does not, the system can enter a single-branch processing mode. Specifically, the system determines the entire target question text or the portion with the higher factual score as the first text segment, and sets the second text segment as an empty segment or a preset placeholder segment. Similarly, when only the normative branch label score meets the corresponding preset conditions while the factual branch label score does not, the system also enters a single-branch processing mode. Specifically, the system determines the entire target question text or the portion with the higher normative score as the second text segment, and sets the first text segment as an empty segment or a preset placeholder segment.

[0127] In the embodiments of this application, in order to incorporate the utilization of the divergence evolution structure during the processing suggestion generation process, the system, after completing the generation of the candidate processing scheme set and the calculation of its multi-objective balance cost, can further regulate the candidate processing scheme set by combining the sentiment divergence map. Specifically, a multi-dimensional divergence label set can be obtained based on the target question text recognition. The multi-dimensional divergence label set is used to structurally annotate the divergence topics, triggering factors, and potential disposal directions described in the target question text, such as including at least one or more of the following: divergence type dimension label, triggering factor dimension label, and disposal path dimension label.

[0128] In embodiments of this application, a multidimensional set of divergence labels can be matched to a set of target nodes in a sentiment divergence graph. The sentiment divergence graph can be stored in a graph database and contains multiple nodes and edges. Nodes include at least divergence type nodes, triggering factor nodes, and resolution path nodes; edges include at least divergence association edges, used to represent the association or evolutionary relationship between different divergences. Divergence association edges carry association strength values, which can be used to characterize the tendency or co-occurrence strength of a particular divergence type in historical samples, thereby providing a quantifiable basis for subsequent graph calculations.

[0129] In embodiments of this application, the system can perform neighborhood expansion and / or shortest path calculation in the sentiment divergence graph based on the target node set to obtain the target evolutionary path and its corresponding cumulative weight. Specifically, neighborhood expansion can be used to search for neighboring nodes highly correlated with the current divergence around the target node to obtain potential deep-root cause nodes or associated divergence nodes; shortest path calculation can be used to find connection paths between different node types to obtain candidate evolutionary links from triggering factors to divergence types and then to resolution paths. The system can cumulatively calculate the path based on the correlation strength value of each divergence-related edge on the path to obtain the corresponding cumulative weight, which is used to measure the relative importance or probability of occurrence of different evolutionary paths.

[0130] In the embodiments of this application, the system can reorder and / or eliminate candidate processing schemes based on cumulative weights, and / or adjust the multi-objective balancing cost based on cumulative weights. Specifically, when a candidate processing scheme matches an objective evolution path or solution path node with a high cumulative weight, the system can increase the ranking priority of that candidate processing scheme in the candidate processing scheme set; when a candidate processing scheme conflicts with a low cumulative weight or a high-weight path, the system can lower its ranking or eliminate it from the candidate processing scheme set. Furthermore, the system can adjust the weight of the cost component in the multi-objective balancing cost based on cumulative weights, so that processing schemes consistent with high-weight evolution paths obtain more favorable cost value in the comprehensive evaluation, thereby making the final output processing recommendations more consistent with the divergence structure and common evolutionary patterns, improving the targeting and stability of the recommendations.

[0131] In the embodiments of this application, in order to further transform the generation process of processing suggestions from static text combination into evaluable action selection, the system, after obtaining the processed candidate processing scheme set, can convert the processed candidate processing scheme set into a candidate intervention action set. The candidate intervention action set includes multiple candidate intervention actions, each of which is used to characterize an executable interactive guidance method or disposal action unit. Specifically, the system can extract key behavioral phrases, constraint statements, or guidance statements in the candidate processing schemes in a structured manner according to preset scheme-action mapping rules, and encode them as candidate intervention actions; the candidate intervention actions can carry action type identifiers, target identifiers, and parameter fields. For example, the action type may include question clarification, delayed decision prompts, boundary setting guidance, information verification prompts, or task decomposition guidance, etc., and the target may include the user himself, the relationship arrangement between the user and the interaction object, or subsequent communication processes, etc., so that subsequent predictions can process different schemes with a unified data structure.

[0132] In embodiments of this application, the system can predict a second predictive subjective state vector corresponding to each candidate intervention action based on a first predictive subjective state vector and a set of candidate intervention actions. Specifically, the first predictive subjective state vector is used to characterize the user state prediction result obtained during the candidate processing scheme screening stage, and the candidate intervention action serves as an input variable that may affect the state. The system can generate an action feature representation for each candidate intervention action, and input the first predictive subjective state vector and the action feature representation into a preset state prediction function or prediction model, outputting the second predictive subjective state vector corresponding to the candidate intervention action. The second predictive subjective state vector may include at least predictive autonomy and predictive dependence, used to characterize the level that the user state may reach after executing the intervention action, thereby enabling a comparable evaluation of the effects of different intervention actions.

[0133] In the embodiments of this application, the system can select a target intervention action based on a second predictive subjectivity state vector and output processing suggestions corresponding to the target intervention action. Specifically, the system can filter and sort multiple candidate intervention actions according to preset selection criteria. The preset selection criteria may include at least ensuring that the predictive autonomy is not lower than a preset lower threshold, that the predictive dependence is not higher than a preset upper threshold, and that the state change amplitude meets stability constraints. When there are multiple candidate intervention actions that meet the criteria, the system can further combine the multi-objective balance cost of the corresponding candidate processing schemes to select the best one. Finally, the system can convert the target intervention action into processing suggestion text, prompts, or interactive guidance content output to the user, so that the processing suggestions not only have an interpretable source, but also meet the preset state constraints and stability requirements in a predictive sense.

[0134] Corresponding to the above-described problem analysis and processing methods, embodiments of this application also provide a problem analysis and processing apparatus.

[0135] Figure 5 A schematic block diagram of a problem analysis and processing apparatus according to an embodiment of this application is shown.

[0136] like Figure 5 As shown, the problem analysis and processing device 500 of this embodiment includes a data acquisition module 510, a text extraction module 520, a knowledge retrieval module 530, and a problem processing module 540.

[0137] The data acquisition module 510 can be used to acquire the target problem text to be analyzed. In one embodiment, the data acquisition module 510 can be used to perform the operation S210 described above, which will not be repeated here.

[0138] The text extraction module 520 can be used to perform type recognition and text extraction on the target question text based on a preset question type label set, to obtain a first text fragment and a second text fragment. The question type label set includes at least factual branch labels and normative branch labels. In one embodiment, the text extraction module 520 can be used to perform the operation S220 described above, which will not be repeated here.

[0139] The knowledge retrieval module 530 can be used to select target knowledge resources from different types of knowledge resources based on the first text fragment and the second text fragment, and perform a retrieval to obtain the corresponding first candidate information and second candidate information. In one embodiment, the knowledge retrieval module 530 can be used to perform the operation S230 described above, which will not be repeated here.

[0140] The problem processing module 540 can be used to perform dynamic balancing calculations on the first candidate information and the second candidate information to obtain processing suggestions. In one embodiment, the problem processing module 540 can be used to execute the operation S240 described above, which will not be repeated here.

[0141] According to an embodiment of this application, the text extraction module 520 can also be used to calculate factual branch label scores and normative branch label scores for the target question text based on the question type label set, and compare the factual branch label scores and normative branch label scores with the corresponding preset conditions respectively; and when both the factual branch label scores and normative branch label scores meet the corresponding preset conditions, trigger the text extraction operation to obtain the first text fragment and the second text fragment.

[0142] According to an embodiment of this application, the problem processing module 540 can also be used to generate a factual evaluation vector based on the first candidate information, the factual evaluation vector including at least a feasibility score, a risk measure, and an uncertainty index; generate a normative evaluation vector based on the second candidate information, the normative evaluation vector including at least a normative matching degree, a responsibility spillover measure, and a long-term deviation measure; generate a set of candidate processing schemes based on the first and second candidate information; for each candidate processing scheme in the set of candidate processing schemes, calculate the risk cost using the factual evaluation vector, calculate the normative deviation cost using the normative evaluation vector, calculate the corresponding multi-objective balance cost based on the risk cost, the normative deviation cost, and the preset subjectivity loss cost; and select a target processing scheme from the set of candidate processing schemes based on the multi-objective balance cost, and output the corresponding processing suggestion.

[0143] According to an embodiment of this application, the problem processing module 540 can also be used to calculate a subjectivity state vector based on user historical interaction data; predict the impact on the subjectivity state vector for each candidate processing scheme in the candidate processing scheme set to obtain a first predicted subjectivity state vector, the first predicted subjectivity state vector including at least predicted autonomy and predicted dependence; when the predicted autonomy is lower than a preset lower threshold and / or the predicted dependence is higher than a preset upper threshold, the corresponding candidate processing scheme is eliminated; and a target processing scheme is selected based on the processed candidate processing scheme set and the multi-objective balance cost, and the corresponding processing suggestion is output; or when the processed candidate processing scheme set is empty, a processing suggestion that satisfies the subjectivity constraint is output.

[0144] According to an embodiment of this application, the problem processing module 540 can also be used to collect historical conversation text and corresponding behavioral data within a preset time window, and extract risk features including at least absolute word frequency, emotional intensity index and short-term high-frequency interaction index; calculate a first risk index and a second risk index based on the risk features, wherein the first risk index is used to characterize idealization bias and the second risk index is used to characterize impulsive decision-making; and adjust the risk cost based on the first risk index and / or the second risk index during the calculation of multi-objective balance cost.

[0145] According to an embodiment of this application, the problem processing module 540 can also be used to obtain the weight parameter vector used for dynamic balance calculation during historical sessions to determine the historical weight benchmark vector; generate the current candidate weight parameter vector based on the factual evaluation vector and normative evaluation vector of the current session, and smoothly update the current candidate weight parameter vector based on the historical weight benchmark vector to obtain the updated weight parameter vector, wherein the smooth update process satisfies that the weight change amplitude of adjacent sessions does not exceed a preset threshold; and calculate the multi-objective balance cost based on the updated weight parameter vector.

[0146] According to embodiments of this application, the problem processing module 540 can also be used to obtain a multi-dimensional divergence label set based on target problem text recognition, and match the multi-dimensional divergence label set to a target node set in a sentiment divergence graph. The sentiment divergence graph includes multiple nodes and multiple edges. The multiple nodes include at least divergence type nodes, triggering factor nodes, and solution path nodes. The edges include at least divergence association edges, which carry association strength values. Based on the target node set, the module performs neighborhood expansion and / or shortest path calculation in the sentiment divergence graph to obtain the target evolution path and the corresponding cumulative weight. Based on the cumulative weight, the module reorders and / or eliminates the candidate processing scheme set, and / or adjusts the multi-objective balance cost based on the cumulative weight.

[0147] According to an embodiment of this application, the problem processing module 540 can also be used to convert the processed set of candidate processing solutions into a set of candidate intervention actions, the set of candidate intervention actions including multiple candidate intervention actions; predict a second predictive subject state vector corresponding to each candidate intervention action based on a first predictive subject state vector and the set of candidate intervention actions; and select a target intervention action based on the second predictive subject state vector and output a processing suggestion corresponding to the target intervention action.

[0148] According to embodiments of this application, the problem analysis and processing apparatus 500 further includes a third text fragment processing module. This third text fragment processing module can be used to perform type identification and text extraction on the target problem text based on a preset problem type tag set, obtaining a third text fragment. The problem type tag set also includes sentiment-related problem tags. Dynamic balancing calculations are performed on the first candidate information, the second candidate information, and the third candidate information to obtain processing suggestions.

[0149] According to embodiments of this application, any multiple modules among the data acquisition module 510, text extraction module 520, knowledge retrieval module 530, and problem processing module 540 can be combined into one module, or any one of these modules can be split into multiple modules. Alternatively, at least some of the functions of one or more of these modules can be combined with at least some of the functions of other modules and implemented in one module. According to embodiments of this application, at least one of the data acquisition module 510, text extraction module 520, knowledge retrieval module 530, and problem processing module 540 can be at least partially implemented as hardware circuitry, such as a field-programmable gate array (FPGA), a programmable logic array (PLA), a system-on-a-chip, a system-on-a-substrate, a system-on-package, an application-specific integrated circuit (ASIC), or any other reasonable means of integrating or packaging circuitry, or implemented in software, hardware, or firmware, or in any suitable combination of any of these three implementation methods. Alternatively, at least one of the data acquisition module 510, text extraction module 520, knowledge retrieval module 530, and problem processing module 540 may be implemented at least partially as a computer program module, which can perform corresponding functions when the computer program module is run.

[0150] Figure 6 A block diagram schematically illustrates an electronic device suitable for implementing a problem analysis and processing method according to an embodiment of this application.

[0151] like Figure 6As shown, an electronic device 600 according to an embodiment of this application includes a processor 601, which can perform various appropriate actions and processes according to a program stored in a read-only memory 602 or a program loaded from a storage portion 608 into a random access memory 603. The processor 601 may include, for example, a general-purpose microprocessor, an instruction set processor and / or an associated chipset and / or a dedicated microprocessor. The processor 601 may also include onboard memory for caching purposes. The processor 601 may include a single processing unit or multiple processing units for executing different steps of the method flow according to an embodiment of this application.

[0152] Random access memory 603 stores various programs and data required for the operation of electronic device 600. Processor 601, read-only memory 602, and random access memory 603 are interconnected via bus 604. Processor 601 executes various steps of the method flow according to embodiments of this application by executing programs in read-only memory 602 and / or random access memory 603. It should be noted that the programs may also be stored in one or more memories other than read-only memory 602 and random access memory 603. Processor 601 may also execute various steps of the method flow according to embodiments of this application by executing programs stored in said one or more memories.

[0153] According to embodiments of this application, the electronic device 600 may further include an input / output interface 605, which is also connected to a bus 604. The electronic device 600 may also include one or more of the following components connected to the input / output interface 605: an input section 606 including a keyboard, mouse, etc.; an output section 607 including a cathode ray tube, liquid crystal display, etc., and a speaker, etc.; a storage section 608 including a hard disk, etc.; and a communication section 609 including a network interface card, such as a local area network card, modem, etc. The communication section 609 performs communication processing via a network such as the Internet. A drive 610 is also connected to the input / output interface 605 as needed. A removable medium 611, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on the drive 610 as needed so that computer programs read from it can be installed into the storage section 608 as needed.

[0154] Embodiments of this application also provide a computer-readable storage medium, which may be included in the device / apparatus / system described in the above embodiments; or it may exist independently and not assembled into the device / apparatus / system. The computer-readable storage medium carries one or more programs, which, when executed, implement the method according to the embodiments of this application.

[0155] According to embodiments of this application, the computer-readable storage medium can be a non-volatile computer-readable storage medium, such as including but not limited to: portable computer disks, hard disks, random access memory, read-only memory, erasable programmable read-only memory, portable compact disk read-only memory, optical storage devices, magnetic storage devices, or any suitable combination thereof. In embodiments of this application, the computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. For example, according to embodiments of this application, the computer-readable storage medium may include the read-only memory 602 described above, and / or random access memory 603, and / or one or more memories other than read-only memory 602 and random access memory 603.

[0156] Embodiments of this application also include a computer program product comprising a computer program containing program code for performing the methods shown in the flowchart. When the computer program product is run on a computer system, the program code is used to cause the computer system to implement the methods provided in the embodiments of this application.

[0157] In one embodiment, the computer program may rely on a tangible storage medium such as an optical storage device or a magnetic storage device. In another embodiment, the computer program may also be transmitted and distributed in the form of signals over a network medium, and downloaded and installed via the communication section 609, and / or installed from the removable medium 611. The program code contained in the computer program can be transmitted using any suitable network medium, including but not limited to: wireless, wired, etc., or any suitable combination thereof.

[0158] In embodiments of this application, the computer program can be downloaded and installed from a network via communication section 609, and / or installed from removable medium 611. When the computer program is executed by processor 601, it performs the functions defined in the system of embodiments of this application. According to embodiments of this application, the systems, devices, apparatuses, modules, units, etc., described above can be implemented by computer program modules.

[0159] According to embodiments of this application, program code for executing the computer programs provided in the embodiments of this application can be written in any combination of one or more programming languages. Specifically, these computational programs can be implemented using high-level procedural and / or object-oriented programming languages, and / or assembly / machine languages. The program code can be executed entirely on the user's computing device, partially on the user's device, partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing device can be connected to the user's computing device via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computing device (e.g., via the Internet using an Internet service provider).

[0160] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0161] Those skilled in the art will understand that the features described in the various embodiments of this application can be combined and / or combined in various ways, even if such combinations or combinations are not explicitly described in this application. In particular, the features described in the various embodiments of this application can be combined and / or combined in various ways without departing from the spirit and teachings of this application. All such combinations and / or combinations fall within the scope of this application.

Claims

1. A problem analysis and processing method, characterized in that, The method includes: Obtain the target problem text to be analyzed; Based on a preset set of question type tags, the target question text is subjected to type recognition and text extraction to obtain a first text fragment and a second text fragment, wherein the set of question type tags includes at least factual branch tags and normative branch tags; Based on the first text fragment and the second text fragment, target knowledge resources are selected from different types of knowledge resources and a retrieval is performed to obtain corresponding first candidate information and second candidate information; and A dynamic balance calculation is performed on the first candidate information and the second candidate information to obtain processing suggestions.

2. The method according to claim 1, characterized in that, The step of performing a dynamic balance calculation on the first candidate information and the second candidate information to obtain processing suggestions includes: A factual evaluation vector is generated based on the first candidate information, and the factual evaluation vector includes at least a feasibility score, a risk measure, and an uncertainty index. A normative evaluation vector is generated based on the second candidate information. The normative evaluation vector includes at least normative matching degree, responsibility spillover measure, and long-term deviation measure. A set of candidate processing schemes is generated based on the first candidate information and the second candidate information. For each candidate processing scheme in the set, a risk cost is calculated using the factual evaluation vector, a normative deviation cost is calculated using the normative evaluation vector, and a corresponding multi-objective balance cost is calculated based on the risk cost, the normative deviation cost, and a preset subjectivity loss cost. Based on the multi-objective balance cost, a target processing scheme is selected from the set of candidate processing schemes, and the corresponding processing suggestion is output.

3. The method according to claim 2, characterized in that, The method further includes: Calculate the subjectivity state vector based on user historical interaction data; For each candidate processing scheme in the candidate processing scheme set, predict the impact on the subjectivity state vector to obtain a first predicted subjectivity state vector, which includes at least predicted autonomy and predicted dependence. If the predicted autonomy is lower than a preset lower threshold and / or the predicted dependence is higher than a preset upper threshold, the corresponding candidate processing schemes are eliminated; and The target processing scheme is selected based on the processed candidate processing scheme set and the multi-objective balance cost, and the corresponding processing suggestion is output; or when the processed candidate processing scheme set is empty, a processing suggestion that satisfies the subjectivity constraint is output.

4. The method according to claim 2, characterized in that, The method further includes: Collect historical conversation text and corresponding behavioral data within a preset time window, and extract risk features including at least absolute word frequency, emotional intensity index and short-term high-frequency interaction index; Based on the aforementioned risk characteristics, a first risk index and a second risk index are calculated, wherein the first risk index characterizes idealization bias, and the second risk index characterizes impulsive decision-making; and In calculating the multi-objective balance cost, the risk cost is adjusted based on the first risk index and / or the second risk index.

5. The method according to any one of claims 2 to 4, characterized in that, The method further includes: Obtain the weight parameter vector used for dynamic balancing calculations during historical sessions to determine the historical weight baseline vector; A current candidate weight parameter vector is generated based on the factual evaluation vector and the normative evaluation vector of the current session. This current candidate weight parameter vector is then smoothly updated based on the historical weight baseline vector to obtain an updated weight parameter vector. The smooth update process ensures that the weight change amplitude between adjacent sessions does not exceed a preset threshold. The multi-objective balance cost is calculated based on the updated weight parameter vector.

6. The method according to claim 1, characterized in that, The process of type identification and text extraction for the target question text includes: Based on the set of question type tags, factual branch tag scores and normative branch tag scores are calculated for the target question text, and the factual branch tag scores and normative branch tag scores are compared with corresponding preset conditions; and When both the factual branch label score and the normative branch label score meet the corresponding preset conditions, a text extraction operation is triggered to obtain the first text fragment and the second text fragment.

7. The method according to any one of claims 2 to 4, characterized in that, The method further includes: A multidimensional divergence label set is obtained based on the target question text recognition, and the multidimensional divergence label set is matched to the target node set in the sentiment divergence graph. The sentiment divergence graph includes multiple nodes and multiple edges. The multiple nodes include at least divergence type nodes, triggering factor nodes, and resolution path nodes. The edges include at least divergence association edges, and the divergence association edges carry association strength values. Based on the target node set, neighborhood expansion and / or shortest path calculation are performed in the sentiment divergence graph to obtain the target evolution path and corresponding cumulative weight; and The candidate processing scheme set is reordered and / or eliminated based on the cumulative weight, and / or the multi-objective balance cost is adjusted based on the cumulative weight.

8. The method according to claim 3, characterized in that, The method further includes: The processed set of candidate treatment schemes is converted into a set of candidate intervention actions, which includes multiple candidate intervention actions. Based on the first predicted subject state vector and the set of candidate intervention actions, predict the second predicted subject state vector corresponding to each candidate intervention action; and Based on the second predictive subjectivity state vector, a target intervention action is selected, and a processing suggestion corresponding to the target intervention action is output.

9. The method according to claim 3, characterized in that, The method further includes: based on the preset question type tag set, performing type identification and text extraction on the target question text to obtain a third text fragment, wherein the question type tag set also includes sentiment question tags.

10. The method according to claim 9, characterized in that, The method further includes: obtaining corresponding third candidate information based on the third text fragment; The step of performing dynamic balancing calculations on the first candidate information and the second candidate information to obtain processing suggestions includes: performing dynamic balancing calculations on the first candidate information, the second candidate information, and the third candidate information to obtain the processing suggestions.

11. A problem analysis and processing apparatus, characterized in that, The device includes: The data acquisition module is used to: acquire the target problem text to be analyzed; The text extraction module is used to: perform type identification and text extraction on the target question text based on a preset question type tag set to obtain a first text fragment and a second text fragment, wherein the question type tag set includes at least factual branch tags and normative branch tags; The knowledge retrieval module is configured to: select target knowledge resources from different types of knowledge resources based on the first text fragment and the second text fragment, and perform a retrieval to obtain corresponding first candidate information and second candidate information; and The problem processing module is used to: perform dynamic balance calculations on the first candidate information and the second candidate information to obtain processing suggestions.

12. An electronic device, comprising: One or more processors; Memory, used to store one or more computer programs. The characteristic feature is that the one or more processors execute the one or more computer programs to implement the steps of the method according to any one of claims 1 to 10.

13. A computer-readable storage medium having a computer program or instructions stored thereon, characterized in that, When the computer program or instructions are executed by a processor, they implement the steps of the method according to any one of claims 1 to 10.

14. A computer program product comprising a computer program or instructions, characterized in that, When the computer program or instructions are executed by a processor, they implement the steps of the method according to any one of claims 1 to 10.

Images