An intent recognition method based on multi-source evidence fusion reasoning

Abstract

The application discloses an intention recognition method based on multi-source evidence fusion reasoning, and belongs to the technical field of artificial intelligence, and comprises the following steps: step S1: acquiring multi-source heterogeneous evidence including at least text, behavior sequence and external knowledge data, converting into standardized evidence units containing evidence content, source reliability score and intention correlation degree, and constructing an initial evidence set based on an evidence reasoning graph model; step S2: inputting the initial evidence set into a reasoning engine based on a neural network, performing multi-hop reasoning through multiple rounds of active query: the first hop analyzes evidence consistency and generates a preliminary intention hypothesis and a follow-up question, subsequent hops dynamically generate new questions based on new evidence to acquire supplementary information in a directional manner, and the process is continued until a termination condition is met; step S3: dynamic evidence fusion calculation; and step S4: outputting an intention conclusion with the highest confidence according to a global fusion result, and generating an explainable evidence chain report containing a complete multi-hop reasoning path and key evidence support.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to an intent recognition method based on multi-source evidence fusion reasoning. Background Technology

[0002] Intent recognition, a core topic in the field of artificial intelligence, has significant application value in key scenarios such as financial risk control, cybersecurity, and intelligent auditing. Its goal is to infer the potential behavioral intent or purpose of a subject by analyzing various types of data generated by that subject. Currently, mainstream intent recognition methods mainly rely on the following two technical paths: The first is a judgment method based on single-point evidence and static rules. This type of method typically relies on evidence of a single dimension or limited dimensions (such as transaction amount thresholds, specific keyword triggers) combined with expert-preset "IF-THEN" type rules for logical judgment. First, it highly depends on complete and accurate prior knowledge, making it difficult to cope with new and complex intent patterns, resulting in poor flexibility and adaptability; second, rule systems can usually only handle deterministic, either-or logic, failing to effectively quantify the uncertainty of evidence and potential conflicts between different pieces of evidence, leading to a high misjudgment rate when information is ambiguous or contradictory; it lacks a clear explanation of the logic behind the conclusion, making it difficult to meet the stringent requirements of auditable and explainable decisions in fields such as finance and the judiciary. The second type is an end-to-end classification method based on machine learning models (especially deep learning models). These methods model intent recognition as a classification task, automatically learning the mapping relationship between features and intents from massive amounts of data by training neural network models (such as CNNs, RNNs, and Transformers). While these methods outperform rule-based systems in pattern capture, their limitations are also significant: First, the model is essentially a passive "pattern matcher" based on fixed input features, lacking the ability to actively question and explore information. When initial evidence is insufficient or ambiguous, the model cannot formulate hypotheses and actively seek new key evidence for verification like human experts, thus limiting its ability to handle complex and hidden intents. Second, the model typically inputs a simple concatenation of heterogeneous evidence from multiple sources, failing to explicitly model and utilize the deep logical, temporal, or causal relationships between the evidence, nor does it differentiate and integrate the credibility of evidence from different sources and of different quality, affecting the robustness of reasoning. Finally, the decision-making process of deep models is usually not interpretable, and their conclusions are difficult to trace back to specific chains of evidence, which is a major obstacle in risk control and compliance review.

[0003] Existing intent identification methods cannot make judgments based on a combination of multi-source evidence, and their use of evidence is not precise enough. Current intent identification methods passively rely on an initial fixed set of evidence. When information is ambiguous, fragmented, or conflicting, they lack the ability to actively explore and verify, resulting in inaccurate identification and poor interpretability of complex and hidden intents. Summary of the Invention

[0004] To address the aforementioned problems, this invention aims to solve the problems described above. One object of this invention is to provide an intent recognition method based on multi-source evidence fusion reasoning to solve the above problems.

[0005] The solution adopted in this invention is: an intent recognition method based on multi-source evidence fusion reasoning, comprising the following steps:

[0006] Step S1: Obtain multi-source heterogeneous evidence, including at least text, behavioral sequences, and external knowledge data, and transform it into standardized evidence units containing evidence content, source reliability scores, and intent relevance. Construct an initial evidence set based on the evidence reasoning graph model.

[0007] Step S2: Input the initial evidence set into the neural network-based inference engine and perform multi-hop inference through multiple rounds of active query: the first hop analyzes the consistency of the evidence and generates preliminary intent hypotheses and follow-up questions, and subsequent hops dynamically generate new questions based on new evidence to obtain supplementary information in a targeted manner until the termination condition is met;

[0008] Step S3: Dynamic Evidence Fusion Calculation: In each iteration of Step S2, after a new evidence unit is injected, a fusion process based on Dempster-Shafer evidence theory is initiated; a basic probability allocation function is constructed for each new evidence unit, and the confidence quality is allocated to a corresponding subset and uncertainty domain in an identification framework composed of mutually exclusive intent hypotheses based on its reliability and relevance; an improved synthesis rule is adopted to synthesize the BPA of all new evidence in this round with the global BPA generated in the previous iteration. This improved rule adopts a weighted redistribution strategy based on Jousselme distance when there is evidence conflict; after synthesis, the updated hypothesis confidence function and likelihood function are output as the current confidence state.

[0009] Step S4: Output the highest confidence intention conclusion based on the global fusion result, and generate an interpretable evidence chain report containing a complete multi-hop reasoning path and key evidence support.

[0010] A preferred technical solution is that, in step S1, the intent of the target subject to be identified is a behavioral intent in a specific domain, and the multi-source heterogeneous evidence specifically includes:

[0011] Static attribute evidence, whose standardized evidence unit includes the basic attribute description of the target subject, historical state labels and their corresponding timeliness weights;

[0012] Dynamic behavioral evidence is obtained by transforming the original behavioral logs into a time-series behavioral graph network, extracting feature vectors of periodic patterns, dense occurrence patterns, and inactive period occurrence patterns as evidence content, and assigning correlation scores to the feature vectors based on the pattern deviation.

[0013] Semantic expression evidence is obtained by analyzing relevant communication records and documents using natural language processing technology, extracting specific semantic patterns, entity references, and sentiment tendencies contained therein as semantic evidence, and assigning sentiment tendency weights based on the sentiment intensity analysis results.

[0014] External verification evidence is obtained by querying or importing data from authoritative list databases and public information databases through interfaces, and matching it with the current target subject. If the match is successful, evidence units containing relevant information categories and levels are generated.

[0015] The preferred technical solution is that the multi-hop inference performed in step S2 through a multi-round iterative active query method specifically includes the following sub-steps:

[0016] Step S2.1: The neural network-based inference engine receives the initial evidence set, performs context encoding and integration of the evidence units in the set through its encoder, and generates at least two alternative intention hypotheses of initial behavior formalized as probability distributions from its decoder or a specific output head based on this context vector.

[0017] Step S2.2: For each initial hypothesis generated in step S2.1, query a pre-built hypothesis-evidence layer mapping strategy library; this strategy library stores the mapping relationship and invocation weight between different categories of intent hypotheses and multi-level evidence types in a vector or graph structure; based on this mapping relationship, parse and determine a priority-ordered sequence of evidence layers to be invoked and the key evidence pattern features under each layer for each hypothesis.

[0018] Step S2.3: The inference engine constructs structured query predicates or generates precise natural language query instructions based on the target evidence level and key features parsed in step S2.2; these instructions drive the corresponding peripheral data pipelines through the application programming interface to perform targeted collection, retrieval, or real-time calculation of the target data.

[0019] Step S2.4: Extract and vectorize the raw data collected in step S2.3 according to the standardization mode of the corresponding evidence level, encapsulate it into new standardized evidence units, and inject or update the corresponding nodes and edges in the evidence reasoning graph model; then, the reasoning engine performs reasoning again on the enhanced graph model, and analyzes the support, conflict or supplementary relationship between the new evidence units and existing evidence and various competing hypotheses through attention mechanism or graph neural network algorithm, and updates the confidence probability distribution of each hypothesis accordingly;

[0020] Step S2.5: Repeat steps S2.2 to S2.4 to form a closed-loop reasoning loop guided by dynamic hypotheses, planned by a strategy library, and executed by proactive data acquisition. This loop continues until a preset termination condition is met. The condition includes the loop count reaching the upper limit, all preset levels of evidence being traversed, or the confidence level of a certain dominant hypothesis exceeding a set certainty threshold.

[0021] A preferred technical solution is that the method further includes the step of constructing and updating the hypothesis-evidence layer mapping strategy library, specifically:

[0022] The strategy base is stored in the form of a knowledge graph, whose nodes include the intent hypothesis of the behavior, the evidence level category, the evidence type, and the edges connecting them. The weights on the edges represent the importance weights of the evidence in verifying a specific hypothesis. The strategy base is initialized based on domain expert experience rules and historical typical cases. During system operation, the entire link log of each multi-hop inference is continuously recorded, including the initial hypothesis, the evidence invoked, changes in the hypothesis confidence, and the final conclusion. Through an offline analysis module, the contribution of different pieces of evidence to the final confirmation or rejection of a hypothesis is periodically evaluated, and the weights of the corresponding edges in the strategy base are dynamically adjusted using a reinforcement learning algorithm, thereby achieving the self-optimization and evolution of the hypothesis-evidence layer mapping strategy base.

[0023] The preferred technical solution is that the dynamic fusion calculation process described in step S3 specifically includes: defining a basic confidence allocation function for each standardized evidence unit, which allocates confidence values ​​to a subset of a set of mutually exclusive and complete intent hypothesis sets, while reserving a portion of confidence for the uncertainty domain; the confidence allocation ratio of the uncertainty domain is dynamically adjusted by the reliability score of the evidence source; in each round of multi-hop inference, the basic confidence allocation function of the newly generated evidence unit is combined with the current global basic confidence allocation function; when different evidence sources are highly conflicting, a conflict reassignment synthesis rule based on evidence distance or reliability weighting is adopted to reasonably handle conflicting evidence; after each synthesis operation, the trust function, likelihood function, and confidence interval of all single intent hypotheses are recalculated, and the hypothesis with the highest confidence and its confidence interval are used as intermediate results to guide the generation of the next-hop inference problem.

[0024] A preferred technical solution is that, after step S2.1 and before step S2.2, the method further includes a step of multiple hypothesis competition and rapid verification. Specifically, after generating multiple competing initial behavioral intent hypotheses, a rapid verification phase is initiated. In this phase, one to two rounds of lightweight evidence queries are initiated in parallel for each hypothesis. These queries use a pre-built, efficient index interface to selectively retrieve key discriminative evidence that strongly supports or refutes the hypothesis. Subsequently, based on the rapidly obtained limited amount of key evidence, a lightweight scoring model is used to initially evaluate and rank each hypothesis. The top-K hypotheses with the highest scores, where K is an integer greater than or equal to 1, are selected as winning hypotheses to enter the subsequent deep multi-hop reasoning main process. At the same time, hypotheses with excessively low scores are downgraded or archived, thereby improving the computational efficiency of the overall reasoning process while ensuring coverage probability.

[0025] The preferred technical solution is that the step of generating an interpretable chain of evidence report in step S4 specifically includes: the system backtracks and parses the complete multi-hop reasoning path, extracts the core questions raised in each round of reasoning, the specific evidence called to answer the questions, the impact of the evidence on the confidence of each hypothesis, and the intermediate conclusion after that round of reasoning; using a combination of template filling and natural language generation, the information extracted above is organized into a coherent narrative report; the report clearly elaborates on the reasoning logic chain from the initial clues to the final conclusion, and uses visual charts to help display the correlation network and changing trends of key evidence; the final report includes at least the final intent determination result, the overall confidence of the determination result, the list of core evidence supporting the result and its contribution analysis, and the main competing hypotheses that were considered but ultimately excluded and their key reasons for exclusion.

[0026] The preferred technical solution is that the method further includes a closed-loop learning process for model training and domain adaptation. The specific steps are as follows: collecting historically completed intent recognition cases to form a labeled dataset containing complete input evidence, multi-hop inference intermediate processes, final judgment results, and comments; using this dataset to perform supervised fine-tuning of the inference engine, optimizing its ability to generate reasonable hypotheses, raise valid questions, and accurately interpret evidence; simultaneously, using the consistency between the final case result and the system's intermediate evaluation as a reward signal, employing reinforcement learning strategies to further optimize the inference engine's decision-making process; for new vertical domains, by injecting a small number of labeled samples and domain knowledge documents from that domain, combined with prompting engineering and efficient parameter fine-tuning techniques, the core inference model can quickly adapt to the terminology and patterns of the new domain, achieving cross-domain intent recognition capability transfer.

[0027] The preferred technical solution is that the method is applied to intent recognition in the field of high-risk behavioral intent recognition. Its specific steps include: using one or a series of clues that trigger warning rules as input to generate an initial evidence set; the inference engine first generates any one of the following competing hypotheses: suspected abnormal behavior, suspected normal complex operation, or suspected other compliant behavior; based on the hypothesis-evidence layer mapping strategy library, for the suspected abnormal behavior hypothesis, priority is given to calling and analyzing the full-scale relational network graph evidence of the involved subjects to discover complex structures, calling related party attribute evidence to identify high-risk attribute associations, and calling subject behavior pattern mutation evidence to identify deviations from historical baselines; through multi-hop inference, various scattered information is gradually associated and integrated, dynamically fusing new findings at each step, ultimately determining whether a specific abnormal intent exists, and generating a detailed visualized relational path diagram and structured report.

[0028] The preferred technical solution is that the method operates as a workflow of an automated service system: the process begins by receiving intent recognition requests and raw data from the upstream business system, and through the processing of steps S1 to S4, finally returns a structured intent recognition conclusion and report to the upstream business system; the system has both real-time stream processing and batch processing modes, and can flexibly adjust the depth and breadth of multi-hop inference according to the timeliness requirements of the business scenario; the system provides a standard API interface to receive evidence data and return structured recognition results and evidence chains; the core inference and fusion services of the system are deployed in the form of microservices, and its knowledge base, strategy base and model support dynamic updates to ensure continuous iteration and optimization of the recognition model without interrupting service.

[0029] Compared with existing technologies, the intent recognition method based on multi-source evidence fusion reasoning of the present invention has the following technical effects:

[0030] 1. Significantly improves recognition accuracy and coverage of complex intents. By introducing a dynamic hypothesis generation and proactive evidence retrieval mechanism, the system can actively explore and verify clues not directly reflected in the initial evidence. In particular, by employing multi-hop reasoning and hypothesis-evidence layer mapping strategies, it can systematically invoke supplementary evidence at different levels, thereby effectively addressing information fragmentation and concealment, achieving more accurate recognition of complex, novel, and ambiguous intent scenarios, and reducing the false negative rate.

[0031] 2. To enhance decision-making robustness and rationally handle uncertainty and conflict, the Dempster-Shafer evidence theory and its improved rules are used for dynamic fusion, which can quantify the reliability, uncertainty, and direct conflict of evidence from different sources. Through the basic confidence allocation function and the conflict redistribution strategy based on Jousselme distance, the system can more rationally weigh contradictory evidence, avoid overall misjudgment due to local information conflicts, and make the final conclusion more robust and reliable in uncertain environments.

[0032] 3. This method achieves a highly interpretable reasoning process, meeting compliance and auditing requirements. It outputs not only conclusions but also a complete and interpretable chain of evidence report. This report traces the entire process from initial clues and hypothesis formulation to the specific evidence actively invoked in each round and its impact on confidence levels, clearly presenting the logical chain of why this conclusion was reached. This significantly enhances the transparency and credibility of decision-making, fully meeting the mandatory compliance requirements for operational traceability and interpretable results in high-risk areas such as financial risk control and security auditing.

[0033] 4. Enhance system intelligence and adaptive efficiency. By leveraging a pre-built and online-optimizable hypothesis-evidence layer mapping strategy library, the system combines domain knowledge with data-driven learning. Combined with a closed-loop learning process, the system continuously learns optimal evidence retrieval strategies from historical cases and optimizes inference paths through reinforcement learning. This enables the system to self-evolve, quickly adapt to new business scenarios and novel risk patterns, reduce reliance on manual rule updates, and significantly improve the intelligence and efficiency of long-term operation and maintenance.

[0034] 5. Optimize computing resource allocation and improve real-time processing performance. By introducing multiple hypothesis competition and rapid verification steps, initial hypotheses are quickly screened before deep inference, and only the few winning hypotheses with the highest probability are subjected to resource-intensive multi-hop inference and multi-layer evidence fusion. This mechanism effectively avoids unnecessary consumption of computing resources, enabling the system to respond more quickly to business scenarios with high real-time requirements while ensuring coverage probability, thus achieving a balance between efficiency and effectiveness. Attached Figure Description

[0035] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments of the invention and, together with the description, serve to explain the principles of the invention. In these drawings, similar reference numerals are used to denote similar elements. The drawings described below are some embodiments of the invention, but not all embodiments. Other drawings will be readily available to those skilled in the art based on these drawings without any inventive effort.

[0036] Figure 1This is a schematic diagram of the intent recognition method based on multi-source evidence fusion reasoning provided in a specific embodiment of the present invention;

[0037] Figure 2 This is a detailed flowchart of step S2 provided in a specific embodiment of the present invention. Detailed Implementation

[0038] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of this application can be arbitrarily combined with each other.

[0039] The following detailed description of the intent recognition method based on multi-source evidence fusion reasoning, in conjunction with the accompanying drawings and embodiments, is provided in detail.

[0040] For ease of understanding, the following explains some key terms in this embodiment:

[0041] Multi-source heterogeneous evidence refers to a collection of information originating from different channels and possessing different data structures and semantic features, such as text data, behavioral sequence data, and external knowledge data. This evidence is used collectively to support the identification of the target entity's intent.

[0042] A standardized evidence unit refers to a data representation of original, multi-source, heterogeneous evidence that has been preprocessed and given a unified format and structure. This unit typically includes the specific content of the evidence, a reliability score of its source, and a correlation score with the intended target, to facilitate subsequent unified processing and fusion.

[0043] Evidence reasoning graph models are graph structures used to represent and organize evidence and their relationships. In this model, nodes can represent evidence units, intentional assumptions, or entities, while edges represent the support, conflict, or association relationships between them, providing a structured foundation for the reasoning process.

[0044] The inference engine of a neural network refers to a computational module built on a deep learning model. It is designed to receive evidence input, generate intent hypotheses, and drive multiple rounds of proactive queries for reasoning. This engine can learn complex patterns and associations and dynamically adjust the direction of inference based on new evidence.

[0045] Multi-hop reasoning refers to the process of gradually deriving a final conclusion from initial evidence through a series of continuous, purposeful queries and information retrieval steps. Each "hop" generates a new question based on the reasoning result of the previous hop, in order to obtain supplementary information in a targeted manner, thereby gradually approaching the true intention.

[0046] Dempster-Shafer evidence theory is a mathematical framework for dealing with uncertainty and imprecision, allowing confidence to be assigned to a subset of a proposition rather than a single proposition. The theory quantifies the degree of support evidence provides for different hypotheses through a basic probability assignment function and offers synthesis rules to fuse confidence from multiple sources of evidence.

[0047] The Basic Probability Assignment Function (BPA) is a core concept in Dempster-Shafer evidence theory. It assigns the overall confidence quality to all possible subsets of hypotheses within the identification framework. This function quantifies the strength of evidence’s support for a particular hypothesis or combination of hypotheses, as well as its support for uncertainty.

[0048] The trust function and the likelihood function are two important measures in Dempster-Shafer evidence theory used to describe the degree of support for a hypothesis. The trust function represents the degree to which evidence directly supports a hypothesis, while the likelihood function represents the degree to which evidence does not contradict that hypothesis. Together, they define a confidence interval, reflecting the range of confidence in the hypothesis.

[0049] An interpretable chain of evidence report is a detailed document automatically generated by the system after completing intent identification. This report not only provides the final intent conclusion but also clearly elucidates the complete reasoning logic path from initial clues to the final conclusion, including key evidence at each step, its impact on hypothesis confidence, and intermediate conclusions during the reasoning process, thereby enhancing the transparency and auditability of decision-making. Example

[0050] This embodiment provides an intent recognition method, the specific implementation of which is as follows:

[0051] like Figure 1-2As shown, step S1 involves acquiring multi-source heterogeneous evidence, including at least text, behavioral sequences, and external knowledge data. This evidence can come from different data interfaces or storage systems, such as file reading, database queries, or API calls. After acquisition, the raw evidence is transformed into standardized evidence units, including data cleaning, format standardization, and feature extraction. Each evidence unit is then assigned a source reliability score and a relevance score to the potential intent. For example, textual evidence can obtain relevance through keyword matching or a simple classifier, behavioral sequence evidence can obtain relevance through statistical frequency or duration, and external knowledge data is assigned a reliability score based on its authority. Subsequently, an initial evidence set is constructed based on an evidence reasoning graph model. This model is a simple node-edge structure, where nodes represent evidence units and edges represent basic connection relationships.

[0052] Step S2: Input the initial evidence set into a neural network-based inference engine, which can be a pre-trained sequence-to-sequence model or a simple classification network. The engine performs multi-hop inference through multiple rounds of proactive queries. In the first hop inference, the engine performs preliminary analysis of the initial evidence, such as pattern matching or rule judgment, generates one or more preliminary intent hypotheses, and proposes general follow-up questions based on these hypotheses, such as "Are there other related behaviors?" or "Does this subject have any historical abnormal records?" In subsequent inference hops, the engine dynamically generates new follow-up questions based on new evidence, and selectively obtains supplementary information. This process continues until preset termination conditions are met, such as reaching the maximum number of query rounds, all preset evidence types have been checked, or the confidence level of a certain hypothesis reaches a fixed threshold.

[0053] Step S3: Perform dynamic evidence fusion calculation. In each iteration of Step S2, after a new standardized evidence unit is injected, a fusion process based on Dempster-Shafer evidence theory is initiated. A basic probability allocation function (BPA) is constructed for each new piece of evidence, assigning confidence quality to the corresponding subset in the identification framework composed of mutually exclusive intention hypotheses, and allocating a portion of the confidence to the uncertainty domain. For example, one piece of evidence may support "normal transactions," while some evidence may not explicitly support any single hypothesis, thus assigning confidence to the uncertainty domain. Subsequently, an improved synthesis rule is used to synthesize the BPA of all new evidence in this round with the global BPA of the previous round. When evidence conflicts, a weighted redistribution strategy based on Jousselme distance is adopted. The degree of conflict is measured by calculating the distance between evidence, and the conflict confidence is redistributed according to the distance, avoiding simply discarding conflicting information. After synthesis, the updated hypothesis confidence function and likelihood function are output, representing the current confidence state for each intention hypothesis.

[0054] The improved synthesis rule is implemented as follows: First, the degree of conflict between pieces of evidence is measured using Jousselme distance; a larger distance indicates a more severe conflict, while a smaller distance indicates closer evidence. Second, a credibility weight is calculated for each piece of evidence based on the distance; the smaller the distance between the evidence and other evidence, the higher the weight, and vice versa. Third, the conflict credibility is weighted and redistributed according to its credibility weight and the degree to which the evidence supports each hypothesis. Evidence with high credibility and supporting a particular hypothesis receives greater weight in this redistribution. This improvement makes the fusion result more robust under conflicting evidence, avoids the counterintuitive results of classic rules, and preserves the uncertainty information of the evidence.

[0055] Step S4: Output the highest confidence intention conclusion based on the global fusion results, i.e., select the hypothesis with the highest trust function or likelihood function value as the final identification conclusion. Simultaneously, generate an interpretable evidence chain report containing the complete multi-hop reasoning path and supporting key evidence. This report, such as a simple text file, lists the input evidence, generated hypotheses, and final conclusion at each step of the reasoning process, and identifies the key evidence supporting the conclusion. The report can record the questions queried in each round, the evidence obtained, and their impact on the conclusion in chronological or logical order.

[0056] This embodiment overcomes the reliance on single evidence by integrating and standardizing multi-source heterogeneous evidence. Leveraging the proactive querying and multi-hop reasoning of a neural network inference engine, supplementary information can be dynamically acquired, significantly improving the accuracy of identifying complex and concealed intentions in scenarios with ambiguous or fragmented information. Dynamic fusion computation based on Dempster-Shafer evidence theory, particularly the improved conflict resolution strategy, enhances the robustness of the reasoning. The resulting interpretable chain of evidence report provides clear logical support for decision-making, meeting the high demands for transparency and auditability in fields such as financial risk control and cybersecurity.

[0057] In the intent recognition method based on multi-source evidence fusion reasoning, although step S1 proposes a general framework for acquiring and standardizing multi-source heterogeneous evidence, in practical applications, especially when identifying behavioral intent in a specific domain, how to efficiently and accurately select and process evidence from different sources to ensure its effectiveness and relevance in identifying behavioral intent is a problem that needs to be solved. If the evidence type is selected inappropriately or the processing method is generalized, the initial evidence set may contain a large amount of noise or lack key information, affecting the accuracy and efficiency of subsequent reasoning. To address this, this application further proposes that the intent of the target subject to be identified is a behavioral intent in a specific domain, and the multi-source heterogeneous evidence specifically includes static attribute evidence, dynamic behavioral evidence, semantic expression evidence, and external verification evidence.

[0058] The standardized evidentiary unit of static attribute evidence includes a basic attribute description of the target subject, historical state tags, and their corresponding timeliness weights. The basic attribute description may include relatively stable features such as registration information, identity information, and device information. Historical state tags refer to the subject's key state at a specific point in the past, such as account activity status, credit rating, and risk assessment. Timeliness weights measure the importance of these historical state tags in the current reasoning; for example, more recent historical states are given higher weights, while older ones are given lower weights, reflecting the degree of attenuation of their influence on current behavioral intentions. In this way, static attribute evidence provides stable and time-sensitive background information about the subject.

[0059] Dynamic behavioral evidence is obtained by transforming raw behavioral logs into a temporal behavioral graph network. Raw behavioral logs include user operation records, transaction records, network access records, and other events arranged chronologically. Transforming them into a temporal behavioral graph network better captures the correlation and temporal order between behaviors, thereby extracting feature vectors of periodic patterns, densely occurring patterns, and inactive period occurrence patterns as evidence. Periodic patterns refer to recurring sequences of behaviors within a specific time period, such as daily logins or weekly transactions. Densely occurring patterns refer to a large number of behaviors occurring in a short period, which may indicate anomalies or high-frequency activity. Inactive period occurrence patterns refer to behaviors occurring during typically inactive times, such as late at night, which may have special significance. These patterns are extracted into feature vectors through feature engineering, and a correlation score is assigned to the feature vectors based on the pattern deviation. Pattern deviation measures the degree of difference between the current behavioral pattern and the historical baseline or normal pattern; a larger deviation usually indicates a higher correlation with a specific intent.

[0060] Semantic evidence is derived by analyzing relevant communication records and documents using natural language processing (NLP) techniques. This extracts specific semantic patterns, entity references, and sentiment tendencies as semantic evidence. Relevant communication records and documents may include text data such as chat logs, emails, forum posts, and reports. NLP techniques are used to parse this text data, identifying keywords, phrases, or sentence structures related to specific behavioral intentions, as well as key people, places, organizations, or events mentioned in the text. Sentiment analysis determines whether the emotion expressed in the text is positive, negative, or neutral. Furthermore, sentiment tendencies are weighted based on the results of sentiment intensity analysis; for example, strong negative emotions may have a higher weight than mild negative emotions, thus affecting their contribution to intent recognition.

[0061] External verification evidence is obtained by querying or batch importing data from authoritative lists and publicly available information databases via API, and then matching it with the current target entity. Authoritative lists may include blacklists, whitelists, and high-risk personnel databases. Public information databases may contain business registration information, legal records, news reports, etc. This data is obtained through API queries or batch imports and matched with the identity or attributes of the target entity to be identified. Successful matches generate evidence units containing relevant information categories and levels. This external evidence is typically highly reliable and can provide strong corroboration or counter-evidence for intent identification.

[0062] By introducing four specific types of multi-source heterogeneous evidence—static attribute evidence, dynamic behavioral evidence, semantic expression evidence, and external verification evidence—and designing refined content extraction and standardization methods for each type, this application effectively addresses the problems of excessive noise, missing key information, or generalized evidence processing in the initial evidence set when identifying behavioral intent in a specific domain. Specifically, static attribute evidence provides stable background information and historical states of the subject, ensuring its relevance through timeliness weights; dynamic behavioral evidence captures deep patterns and deviations in behavior through a time-series behavioral graph network, making abnormal behavior identification more sensitive; semantic expression evidence mines deep semantic and emotional cues from text, enriching the understanding of the subject's intent; and external verification evidence introduces authoritative third-party information, providing strong external corroboration for intent judgment. These refined evidence units not only improve the quality and relevance of the initial evidence set in step S1 but also provide more accurate and discriminative input for subsequent multi-hop inference and dynamic evidence fusion calculations based on neural network-based inference engines, thereby significantly improving the accuracy and reliability of behavioral intent identification in a specific domain and providing more solid data support for the final generated interpretable evidence chain report.

[0063] While methods for intent recognition based on multi-source heterogeneous evidence can acquire and standardize various evidence types and utilize neural network inference engines for initial reasoning, they often struggle to accurately capture deeper intents in complex and ever-changing real-world scenarios. Especially when intent hypotheses are unclear or evidence is ambiguous, efficiently and purposefully acquiring more key information to verify or revise hypotheses is a critical challenge for improving the accuracy and efficiency of inference.

[0064] To address this, this application further proposes a multi-hop inference method using a multi-round iterative active query approach, specifically including the following sub-steps: First, the neural network-based inference engine receives the initial evidence set, performs context encoding and integration of the evidence units in the set through its encoder, and generates at least two alternative intention hypotheses of initial behavior, formalized as probability distributions, from its decoder or a specific output head based on this context vector. The neural network inference engine can adopt various architectures, such as a Transformer-based encoder-decoder model, graph neural network, or hybrid model. The encoder is responsible for mapping standardized evidence units (such as text embeddings, behavioral sequence feature vectors, attribute labels, etc.) from different sources and formats to a unified low-dimensional vector space and capturing the contextual relationships and semantic associations between them. For example, the encoder can be a multilayer perceptron, recurrent neural network, or Transformer encoder, integrating evidence information through a self-attention mechanism. The integrated context vector represents a comprehensive representation of all currently known evidence. The decoder or a specific output head then outputs multiple intention hypotheses of initial behavior with competing relationships based on this context vector, through a multi-classification layer or sequence generation model. These assumptions are presented in the form of probability distributions, representing the inference engine's initial assessment of the likelihood of different intentions. For example, the output is a vector containing assumptions such as "normal transaction intention" and their corresponding probabilities.

[0065] Secondly, for each initial hypothesis generated above, a pre-built "hypothesis-evidence layer mapping strategy library" is queried. This strategy library stores the mapping relationships and invocation weights between different categories of intent hypotheses and multi-level evidence types in a vector or graph structure. Based on this mapping relationship, a priority-ordered sequence of evidence layers to be invoked and the key evidence pattern features under each layer are determined for each hypothesis. The "hypothesis-evidence layer mapping strategy library" is a core knowledge component that predefines the correlation between different intent hypotheses and the evidence to be verified. This strategy library can take the form of a knowledge graph, where nodes represent intent hypotheses, evidence layers (such as static attribute layers, dynamic behavior layers, and external verification layers), and specific evidence types, and edges represent the mapping relationships between them, with weights indicating the importance or priority of the evidence type in verifying a specific hypothesis. The parsing process involves traversing or querying this graph, dynamically generating an ordered evidence query plan based on the type and confidence level of the current hypothesis. This plan not only specifies the types of evidence to be queried but may also include specific patterns or features that need attention in these pieces of evidence. For example, transaction frequency within a specific time period, login behavior of a specific IP address, etc.

[0066] Furthermore, based on the target evidence level and key features parsed above, the inference engine constructs structured query predicates or generates precise natural language query instructions. These instructions drive the corresponding external data pipelines through application programming interfaces (APIs) to perform targeted collection, retrieval, or real-time computation of the target data. Constructing query predicates means transforming abstract evidence requirements into executable data query statements. For example, if the target evidence level is "dynamic behavioral evidence" and the key feature is "high-frequency, low-value transaction pattern," the inference engine might generate an SQL query statement, such as `SELECT FROM transaction_logs WHERE user_id=[target_id] AND amount<1000 AND frequency>10_per_hour`. For natural language query instructions, the inference engine can utilize natural language generation technology to transform the structured query intent into human-readable text and send it to external data services with natural language understanding capabilities. These instructions call backend data services or data pipelines through standardized application programming interfaces, such as RESTful APIs or RPC interfaces. These pipelines are responsible for efficiently collecting, retrieving, or performing real-time computation of the required target data from various data sources.

[0067] Subsequently, the collected raw data is feature-extracted and vectorized according to the standardized pattern of the corresponding evidence level, encapsulated into new standardized evidence units, and injected into or updated the corresponding nodes and edges in the evidence reasoning graph model. The reasoning engine then performs further reasoning on the enhanced graph model, analyzing the support, conflict, or supplementary relationships between the new evidence units and existing evidence and competing hypotheses through attention mechanisms or graph neural network algorithms, thereby updating the confidence probability distribution of each hypothesis. The collected raw transaction records, login logs, communication content, etc., need to undergo preprocessing and feature engineering to conform to the system's standardized evidence unit format. This includes data cleaning, structure transformation, feature extraction, and vectorization. These new standardized evidence units are then injected into the evidence reasoning graph model as new nodes or to update the attributes of existing nodes, and edges are established with other relevant evidence or entities. For example, a new transaction record can be used as an edge connected to a user node. After the graph model is updated, the reasoning engine runs its reasoning logic again. Through attention mechanisms or graph neural networks, the model can identify the correlation between new evidence and existing evidence, as well as their support, conflict, or supplementary effects on different intent hypotheses. Based on this analysis, the system dynamically adjusts and updates the confidence probability distribution of all competing intent hypotheses, reflecting the information gain brought by new evidence.

[0068] Finally, the above query strategy, data collection, and inference update process are repeated to form a closed-loop inference cycle guided by dynamic hypotheses, planned by a strategy library, and executed through proactive data acquisition. This cycle continues until preset termination conditions are met, including but not limited to reaching the maximum number of iterations, completing the traversal of all preset levels of evidence, or exceeding a set certainty threshold for the confidence level of a dominant hypothesis. This repeated execution process constructs an iterative and adaptive inference cycle. In each iteration, the system uses the strategy library to guide the next step of evidence acquisition based on the most credible hypothesis, actively queries to acquire new data, integrates the new data into the inference model, and updates the hypothesis confidence level. This closed loop ensures the dynamism and goal-orientation of the inference process. Termination conditions are key to controlling the depth and efficiency of inference. For example, a maximum number of query rounds can be set to avoid infinite loops; or the process can stop when all levels of evidence related to the current hypothesis have been explored; or, when the confidence level of a certain intentional hypothesis reaches a preset high threshold, such as 95%, the hypothesis is considered sufficiently certain, and the inference can be terminated early, outputting a conclusion.

[0069] By employing a multi-round iterative proactive query approach to perform multi-hop inference, this application effectively addresses the problem of accurately determining deep intents based solely on initial evidence in complex intent recognition scenarios. The scheme first generates multiple competing initial intent hypotheses and then intelligently plans a priority-ranked evidence query sequence using a pre-built "hypothesis-evidence layer mapping strategy library," achieving precise and goal-oriented evidence acquisition. The inference engine constructs query instructions based on the plan, driving the peripheral data pipeline to collect key data in a targeted manner, ensuring the timeliness and accuracy of the required information. Newly collected data is dynamically injected into the evidence inference graph model, and its support, conflict, or supplementary relationships with existing hypotheses are analyzed in real time through attention mechanisms or graph neural network algorithms, thereby dynamically updating the confidence level of each hypothesis. This closed-loop inference cycle, guided by dynamic hypotheses, planned by the strategy library, and executed through proactive data acquisition, enables the system to adaptively adjust its evidence acquisition strategy according to the progress of inference until a preset termination condition is met, significantly improving the accuracy, efficiency, and robustness of intent recognition. Especially when information is incomplete or ambiguous, it can gradually converge to a more reliable intent conclusion through progressively deeper evidence exploration.

[0070] In some of the embodiments described above in this application, a "hypothesis-evidence layer mapping strategy library" is proposed to guide the acquisition and analysis of evidence in multi-hop reasoning. However, if this strategy library is statically preset, it may not be able to adapt to constantly changing domain knowledge and data patterns, resulting in limited reasoning efficiency and accuracy. Especially when facing complex and ever-changing intent recognition scenarios, its guiding role may not be optimal.

[0071] To address this, this application further proposes a method for constructing and updating the aforementioned "hypothesis-evidence layer mapping strategy library." Specifically, the strategy library is stored in the form of a knowledge graph, where nodes include the intent hypothesis of a behavior, evidence level categories, specific evidence types, and edges connecting them. The weights on the edges represent the importance weight of that type of evidence in verifying a specific hypothesis. This "hypothesis-evidence layer mapping strategy library," as a key component guiding multi-hop reasoning, adopts a knowledge graph storage structure, aiming to efficiently represent and manage complex relationships. Nodes in the knowledge graph can be specifically represented as: the intent hypothesis of a behavior, evidence level categories (e.g., static attribute layer, dynamic behavior layer, semantic expression layer, etc.), and specific evidence types (e.g., account registration time, transaction frequency, keyword mentions, etc.). The edges connecting these nodes represent the relationships between different entities. For example, the intent hypothesis of a certain behavior may be associated with multiple evidence level categories, and each evidence level category contains multiple specific evidence types. The weights attached to the edges numerically represent the importance or support of that specific evidence type or evidence level category in verifying or refuting the intent hypothesis of a certain behavior. This knowledge graph-based structure enables the policy base to store domain knowledge in a structured and scalable manner, facilitating subsequent queries, updates, and reasoning. For example, weights can be floating-point numbers between 0 and 1, representing the degree of relevance or importance.

[0072] The initialization of the strategy library is based on domain expert experience rules and historical typical cases. The initial construction of the strategy library is fundamental to ensuring the system can effectively start and possess preliminary reasoning capabilities. This is typically achieved through two main methods: first, combining domain expert experience rules to structurally input the experts' understanding of the relationship between specific intentions and relevant evidence into the knowledge graph; second, analyzing historically labeled typical cases, extracting the association patterns between intentions and evidence from these cases, and transforming them into nodes and edges in the knowledge graph. This initialization method ensures that the strategy library possesses a certain degree of professionalism and practicality from the outset of system deployment.

[0073] During system operation, a full-link log is continuously recorded for each multi-hop inference, including the initial hypothesis, invoked evidence, changes in hypothesis confidence, and the final conclusion. To achieve dynamic optimization of the policy library, the system needs comprehensive recording of its own operational process. The full-link logging mechanism aims to capture the complete lifecycle information of each multi-hop inference. This includes: the initial intent hypothesis generated by the system at the start of inference; all evidence units actively queried and invoked during inference to verify or refute these hypotheses and their sources; how the confidence levels (e.g., trust function, likelihood function) of each competing hypothesis dynamically change over time as new evidence is injected and integrated; and the final intent recognition conclusion output by the system. This log data forms the basis for subsequent offline analysis and policy library optimization, providing a "footprint" of system behavior that allows for tracing and evaluating every decision point and evidence contribution during the inference process.

[0074] The offline analysis module periodically evaluates the contribution of different pieces of evidence to the final confirmation or rejection of a hypothesis, and uses reinforcement learning algorithms to dynamically adjust the weights of corresponding edges in the policy library, thereby achieving self-optimization and evolution of the "hypothesis-evidence layer mapping policy library." The offline analysis module is the core of the policy library's self-optimization. It periodically processes end-to-end log data recorded during system operation. This module uses statistical analysis and causal inference to evaluate which types or levels of evidence played a crucial role in correctly identifying or excluding a particular intent hypothesis in historical cases—that is, their "contribution." Based on this, reinforcement learning algorithms are used to dynamically adjust the weights of the edges connecting hypotheses and evidence in the knowledge graph. The reinforcement learning model views the reasoning process as a series of decisions, using the accuracy of the final identification result as a reward signal. Through continuous trial and error and learning, the model can discover better evidence query strategies and adjust the weights of edges in the policy library accordingly, enabling the system to prioritize querying evidence that is more effective and discriminative in identifying the target intent in future reasoning. This mechanism allows the policy library to continuously learn and evolve based on actual operational results, thereby continuously improving the efficiency and accuracy of multi-hop reasoning.

[0075] By constructing a "hypothesis-evidence layer mapping strategy library" in the form of a knowledge graph and initializing it with domain expert experience and historical cases, this application provides structured and preliminarily intelligent evidence query guidance for multi-hop inference. Furthermore, by continuously recording the entire inference process log and periodically evaluating the contribution of different pieces of evidence to the final confirmation or rejection of a hypothesis using an offline analysis module, and dynamically adjusting the weights of corresponding edges in the strategy library using a reinforcement learning algorithm, this application achieves self-optimization and evolution of the strategy library. This enables the inference engine to adaptively adjust the evidence query priority and strategy based on actual operating results and constantly changing domain knowledge, thereby ensuring that the most discriminative key evidence is always efficiently and accurately acquired and utilized during multi-hop inference, significantly improving the efficiency and accuracy of intent recognition, and effectively avoiding the performance degradation caused by a static strategy library.

[0076] While some embodiments of this application propose dynamic evidence fusion computation based on Dempster-Shafer evidence theory, practical applications still face challenges. These include how to accurately define the basic confidence allocation function for each standardized evidence unit, dynamically adjust the confidence allocation ratio of the uncertainty domain according to the quality of the evidence, and employ more robust synthesis rules to effectively handle highly conflicting evidence to ensure the accuracy and stability of the fusion results. Improper handling of these issues may lead to inaccurate confidence assessments of intent hypotheses, thereby affecting the effective guidance of subsequent multi-hop inference.

[0077] In response, this application further proposes that the aforementioned dynamic fusion calculation process specifically includes: defining a basic confidence allocation function for each standardized evidence unit, which allocates confidence values ​​to a subset of a set of mutually exclusive and complete intent hypothesis sets, while reserving a portion of confidence for the uncertainty domain; the confidence allocation ratio of the uncertainty domain is dynamically adjusted by the reliability score of the evidence source; in each round of multi-hop inference, the basic confidence allocation function of the newly generated evidence unit is combined with the current global basic confidence allocation function; when different evidence sources are highly conflicting, a conflict reassignment synthesis rule based on evidence distance or reliability weighting is adopted to reasonably handle conflicting evidence; after each synthesis operation, the trust function, likelihood function, and confidence interval of all single intent hypotheses are recalculated, and the hypothesis with the highest confidence and its confidence interval are used as intermediate results to guide the generation of the next-hop inference problem.

[0078] Specifically, a basic confidence assignment function is defined for each standardized unit of evidence. This function, the core of Dempster-Shafer evidence theory, quantifies the confidence of an evidence in supporting subsets of intent hypotheses within the identification framework. For example, confidence values ​​can be assigned to subsets of one or more mutually exclusive intent hypotheses based on factors such as evidence content and intent relevance, using pre-defined rules, machine learning models, or expert knowledge. If evidence strongly supports a particular intent hypothesis, most of the confidence is assigned to that hypothesis; if evidence supports multiple hypotheses but cannot distinguish between them, confidence is assigned to the union of these hypotheses. Simultaneously, the function reserves a portion of confidence for the uncertainty domain to represent the part of the evidence that cannot explicitly support any specific intent hypothesis.

[0079] The reliability allocation ratio for the uncertainty region is dynamically adjusted by the evidence source reliability score. A higher reliability score indicates more credible and definitive information, thus the reliability ratio allocated to the uncertainty region should be reduced accordingly. Conversely, a lower reliability score indicates higher ambiguity or uncertainty in the evidence, requiring more reliability to be allocated to the uncertainty region to reflect the uncertainty of the information. This dynamic adjustment can be achieved through a mapping function, ensuring that the fusion process fully considers the quality of the evidence itself and avoids excessive influence of low-quality evidence on the determination of certainty.

[0080] In each round of multi-hop reasoning, the basic confidence assignment function (BPA) of newly generated evidence units is synthesized with the current global basic confidence assignment function. Here, the "current global basic confidence assignment function" represents the cumulative confidence state of all processed evidence prior to the current reasoning round. The synthesis of the BPA of new evidence with the global BPA allows the system to progressively accumulate and update the confidence levels for each intent hypothesis, reflecting the dynamic and incremental nature of the reasoning process.

[0081] When there is a high degree of conflict between different sources of evidence, a conflict redistribution synthesis rule based on evidence distance or reliability weighting is employed to reasonably handle conflicting evidence. For example, the Jousselme distance or other distance metrics between conflicting evidence can be calculated to quantify the degree of conflict. Evidence with greater conflict may be assigned a lower weight in its reliability during synthesis, or the conflicting portion may be reassigned to the uncertainty region, or the conflicting reliability may be proportionally distributed to non-conflicting hypotheses based on distance. Alternatively, by combining the reliability scores of evidence sources, when evidence conflicts occur, evidence with higher reliability scores is given priority, and its reliability weight is increased, while the weight of evidence with lower reliability is decreased. The reliability portion of the conflict can be reassigned to non-conflicting hypotheses proportionally, or allocated to the uncertainty region. This improved synthesis rule aims to address the limitations of traditional DS theory in handling highly conflicting evidence, avoiding unreasonable reliability allocation due to conflict, thereby improving the robustness and accuracy of the fusion results.

[0082] After each synthesis operation, the trust function, likelihood function, and confidence interval for all individual intent hypotheses are recalculated. The hypothesis with the highest current confidence and its confidence interval are used as intermediate results to guide the generation of the next-hop inference question. The trust function represents the minimum confidence support for a given intent hypothesis, the likelihood function represents the maximum confidence support for a given intent hypothesis, and the confidence interval [Bel(A),Pl(A)] represents the confidence range for hypothesis A. Using these calculation results as intermediate results is crucial. It not only provides the current confidence state for each intent hypothesis, but more importantly, this "hypothesis with the highest current confidence and its confidence interval" can be directly fed back to the inference engine to guide the generation of the next round of proactive query questions. For example, if the confidence of a hypothesis increases significantly, the inference engine may generate a query question aimed at further confirming the hypothesis or excluding its competitors. If the confidence of multiple hypotheses is close, a query aimed at distinguishing these competing hypotheses may be generated.

[0083] Through the aforementioned dynamic evidence fusion calculation process, this method can more precisely quantify the degree of support each evidence unit provides for the intent hypothesis and dynamically adjust uncertainty based on the reliability of the evidence, thereby improving the accuracy of confidence allocation. Especially when facing highly conflicting evidence, the use of conflict reassignment synthesis rules based on evidence distance or reliability weighting effectively avoids the limitations of traditional fusion methods in conflict handling, making the fusion results more robust and credible. Furthermore, the real-time updating and output of the hypothesis with the highest confidence and its confidence interval after each fusion provides clear and dynamic guidance for the generation of questions in subsequent multi-hop inference, enabling the inference process to focus more on key information, avoid blind queries, and significantly improve the efficiency and accuracy of the entire intent recognition process.

[0084] In some of the above embodiments, the intent recognition method, through a neural network-based inference engine, generates multiple candidate initial behavior intent hypotheses after receiving an initial evidence set, and performs multi-hop inference based on these hypotheses through multi-round iterative active querying. However, when the number of initially generated hypotheses is large, performing comprehensive and deep multi-hop inference on each hypothesis may lead to excessive consumption of computational resources and reduced inference efficiency, especially in scenarios requiring rapid response, where such overhead is unacceptable.

[0085] To address this, this application further proposes a rapid verification phase after generating multiple competing initial behavioral intent hypotheses. In this phase, one to two rounds of lightweight evidence queries are initiated in parallel for each hypothesis. These queries, through a pre-built, efficient indexing interface, target and retrieve "key discriminatory evidence" that strongly supports or refutes the hypothesis. Subsequently, based on the rapidly acquired limited key evidence, a lightweight scoring model is used to initially evaluate and rank each hypothesis. The top-K hypotheses with the highest scores, where K is an integer greater than or equal to 1, are selected as winning hypotheses and enter the subsequent deep, multi-hop inference main process. Simultaneously, hypotheses with excessively low scores are downgraded or archived, thereby improving the overall computational efficiency of the inference process while ensuring coverage probability.

[0086] Specifically, after the inference engine generates multiple competing initial behavioral intent hypotheses based on initial evidence, the system immediately initiates a rapid verification phase. This phase aims to quickly and efficiently filter these initial hypotheses to avoid wasting computational resources on hypotheses with low probability, thereby optimizing the efficiency of subsequent deep inference. During the rapid verification phase, the system initiates one to two rounds of lightweight evidence queries in parallel for each generated initial behavioral intent hypothesis. "Lightweight" here means that the depth and breadth of the queries are strictly limited, typically involving only a small amount of readily available evidence with strong discriminative power for the hypothesis. The parallel query mechanism ensures that the verification process for multiple hypotheses can be performed simultaneously, significantly shortening the overall verification time. To achieve the efficiency of lightweight evidence queries, the system utilizes pre-built, high-efficiency indexing interfaces. These interfaces can be query APIs optimized for specific data sources (such as user profile databases, behavior log indexes, and external knowledge bases), enabling rapid location and extraction of the required information without complex full data scanning or deep analysis. The goal of lightweight evidence queries is to selectively retrieve "key discriminative evidence" that strongly supports or refutes a specific hypothesis. This evidence is typically information that is highly relevant to the core features of the hypothesis and has a decisive influence. Through pre-defined rules or models, the system can identify and prioritize this type of evidence. After quickly acquiring a small amount of key evidence, the system uses a lightweight scoring model to initially evaluate and rank the initial hypotheses. This scoring model is usually simple in structure and has low computational cost, enabling it to quickly calculate the support or confidence score of each hypothesis based on the acquired key evidence. For example, it could be a rule-based expert system, a simple linear regression model, or a pre-trained small classifier. After evaluation by the lightweight scoring model, the system selects the Top-K hypotheses with the highest scores, where K is an integer greater than or equal to 1. These selected Top-K hypotheses are considered the most likely "winning hypotheses" and will qualify for the subsequent deep multi-hop inference main process, undergoing more comprehensive and detailed evidence analysis. Hypotheses that scored too low in the initial evaluation will be downweighted or directly archived by the system. Deweighting means that these hypotheses will have a lower priority in subsequent inferences, or that fewer computational resources will be allocated to them; archiving means that these hypotheses are temporarily shelved and no longer participate in the current round of deep inference, but their information may be retained for subsequent analysis or as negative samples for learning.

[0087] By introducing the aforementioned steps of multiple hypothesis competition and rapid verification, this application effectively solves the problem of wasted computational resources and low inference efficiency caused by an excessive number of initial hypotheses in multi-hop inference. After the inference engine generates multiple competing initial behavioral intent hypotheses, the system no longer blindly performs deep inference on all hypotheses, but instead initiates a rapid verification phase. In this phase, by initiating one to two rounds of lightweight evidence queries in parallel and utilizing a pre-built, efficient index interface to retrieve key discriminative evidence, information that has a decisive impact on each hypothesis can be quickly obtained at extremely low computational cost. Subsequently, based on this small amount of key evidence, a lightweight scoring model is used to initially evaluate and rank the hypotheses, thereby efficiently selecting the top-K winning hypotheses with the highest scores and allowing them to enter the subsequent deep multi-hop inference main process. At the same time, hypotheses with excessively low scores are downgraded or archived, avoiding unnecessary resource investment in low-probability hypotheses. This mechanism significantly improves the computational efficiency and response speed of the overall reasoning process while ensuring coverage of all possible intent assumptions. This allows the system to focus more quickly on the most core and most likely intent assumptions, thus providing more practical and efficient intent recognition services in scenarios with limited resources or high timeliness requirements.

[0088] In some of the embodiments described above in this application, a method is proposed to output the highest confidence intention conclusion based on the global fusion result and generate an interpretable evidence chain report containing a complete multi-hop reasoning path and key evidence support. However, in practical applications, if the report content lacks sufficient detail and structured presentation, it may be difficult to clearly demonstrate the complex reasoning process, making it difficult for users to understand the deep logic behind the system's specific judgments, thereby affecting the trust and auditability of the system's decisions.

[0089] In response, this application further proposes the following steps for generating an interpretable chain of evidence report: The system backtracks and parses the complete multi-hop reasoning path, extracting the core questions raised in each round of reasoning, the specific evidence invoked to answer those questions, the impact of that evidence on the confidence of each hypothesis, and the intermediate conclusions after that round of reasoning; Using a combination of template filling and natural language generation, the above structured information is organized into a coherent narrative report; The report clearly elucidates the logical chain of reasoning from the initial clues to the final conclusion, and uses visual charts to assist in displaying the correlation network and changing trends of key evidence; The final report includes at least the final intent determination result, the overall confidence level of the determination result, a list of core evidence supporting the result and its contribution analysis, and the main competing hypotheses that were considered but ultimately excluded, along with the key reasons for their exclusion.

[0090] Specifically, the system's ability to backtrack and parse the complete multi-hop reasoning path means that after the multi-hop reasoning process is completed, the system can access and analyze the stored reasoning logs or reasoning graph structures to trace back each reasoning decision from the initial evidence to the final conclusion. This includes identifying the starting point, intermediate nodes, and final decision point of the reasoning, and understanding the logical connections between each step.

[0091] Extracting the core questions raised in each round of inference refers to the process by which the inference engine dynamically generates questions for targeted acquisition of supplementary information based on the current confidence state and policy library in each iteration of multi-hop inference. These questions are recorded by the system and precisely extracted during report generation to demonstrate the driving force behind the inference.

[0092] Extracting the specific evidence invoked to answer the question refers to the system recording and identifying standardized units of evidence actually acquired, processed, and utilized from multi-source heterogeneous evidence in each round of reasoning to respond to the core question. The identifiers, key content summaries, and source information of this evidence are extracted.

[0093] Extracting the impact of this evidence on the reliability of each hypothesis refers to how each newly injected evidence unit changes the trust function, likelihood function, or reliability interval of each intended hypothesis during the dynamic evidence fusion calculation process. The system quantifies and records this impact, for example through reliability increments and reliability weight adjustments, to reveal the key role of the evidence in decision-making.

[0094] Extracting intermediate conclusions after each round of reasoning refers to the process by which the system identifies the intent hypothesis with the highest confidence level and its corresponding confidence interval based on the updated hypothesis confidence function and likelihood function after each round of evidence fusion calculation. These intermediate results reflect the stage-by-stage judgments in the reasoning process.

[0095] The method of combining template filling and natural language generation to organize the above-mentioned structured information into a coherent narrative report involves pre-designing a series of report templates that define the overall structure and key information points of the report. Then, using natural language generation (NLG) technology, the extracted structured data (such as reasoning paths, questions, evidence, impact values, intermediate conclusions, etc.) is dynamically filled into the templates to generate fluent and easy-to-understand natural language descriptions, ensuring the coherence and readability of the report.

[0096] The report clearly outlines the logical chain of reasoning from initial clues to final conclusions. This means that the report not only lists the reasoning steps, but also demonstrates, through clear narration and logical connectors, how each step of reasoning connects to the previous step, how it leads to the next step, and how evidence is gradually accumulated or the hypothesis is revised, ultimately forming a complete decision-making chain.

[0097] Visualizing the relationship network and trends of key evidence through charts and graphs involves using various visualization techniques such as graph structures, time series graphs, and bar charts to intuitively present the relationships between different evidence units, the changing trends in the impact of evidence on the confidence of hypotheses, and the evolution of the confidence of each intent hypothesis during the reasoning process. This helps users quickly understand complex information. The final report must include at least the final intent determination result, meaning that one of the core contents of the report is the conclusion of the target subject's intent determined by the system after multi-hop reasoning and evidence fusion. The final report must include at least the overall confidence level of the determination result, meaning that the report will clearly provide a quantitative level of confidence in the final intent determination result, such as a percentage value or a confidence interval, to reflect the system's confidence level in this conclusion. The final report must include at least a list of core evidence supporting the result and its contribution analysis, meaning that the report will list the key evidence units that play a decisive role in the final determination result and analyze the magnitude and manner in which these pieces of evidence contribute to the final conclusion during the reasoning process. The final report should include at least the main competing hypotheses that were considered but ultimately excluded, along with the key reasons for their exclusion. This means that the report not only presents the final conclusions but also lists other intentional hypotheses that were carefully considered during the reasoning process but were ultimately systematically excluded, and details the specific reasons for the exclusion of these hypotheses, such as being rejected by conflicting evidence or a continuous decline in reliability.

[0098] Through the aforementioned technical solutions, the system can generate highly interpretable evidence chain reports. These reports not only provide the final intent identification results but also, by tracing and analyzing the complete reasoning path, detail the logical evolution from initial clues to the final conclusion. By extracting and presenting the core issues in each round of reasoning, the specific evidence invoked and its impact on hypothesis confidence, as well as intermediate conclusions, the report clearly reveals the internal mechanism of the system's decision-making. The combination of template filling and natural language generation allows complex structured reasoning information to be presented in a coherent and easily understandable narrative, greatly improving the report's readability. Furthermore, visual charts are used to illustrate the relationship network and changing trends of key evidence, further enhancing the user's intuitive understanding of the reasoning process. The final report includes the final judgment, overall confidence level, core evidence and its contribution analysis, as well as excluded competing hypotheses and their reasons, providing users with comprehensive, transparent, and auditable decision-making basis. This not only enhances user trust in the intent identification results but also provides valuable support for system debugging, optimization, and compliance review, effectively solving the problem that traditional reports struggle to clearly present complex reasoning logic.

[0099] In some of the above implementations, an intent recognition method based on multi-source evidence fusion reasoning is proposed. Through multi-hop reasoning and dynamic evidence fusion, it can identify target intent from complex heterogeneous data. However, in practical applications, intent recognition tasks often face challenges such as constantly evolving data patterns, the emergence of new types of evidence, and the need to quickly adapt to different business domains. If the inference engine's capabilities are static and lack continuous learning and domain adaptation mechanisms, its recognition accuracy and robustness may decline over time. Furthermore, when facing new vertical domains, it requires significant manpower for redevelopment or manual optimization, which limits its generalization ability and long-term effectiveness.

[0100] To address this, this application further proposes a closed-loop learning process for model training and domain adaptation. This closed-loop learning process aims to ensure that the intent recognition method can continuously learn from new data and feedback, and effectively adapt to different application scenarios or vertical domains. Through automated and semi-automated mechanisms, it continuously improves the performance and generalization ability of the inference engine, thereby overcoming the limitations of static models.

[0101] To achieve continuous optimization of the inference engine, a high-quality labeled dataset is needed. This dataset is formed by collecting historical completed intent recognition cases, which include complete input evidence, multi-hop inference intermediate processes, final judgment results, and comments. Input evidence refers to the original, multi-source, heterogeneous data used in the recognition process; the multi-hop inference intermediate processes record each step the inference engine takes when processing this evidence, such as the generated hypotheses, the questions posed, the order of evidence invocation, and the changes in the confidence of each hypothesis; the final judgment result is the intent recognition conclusion confirmed by human experts or the business system; and the comments indicate where the model performs well or has shortcomings. This detailed case data provides rich training signals for subsequent supervised fine-tuning and reinforcement learning.

[0102] Building upon this foundation, the inference engine is subjected to supervised fine-tuning using this dataset to optimize its ability to generate reasonable hypotheses, formulate valid questions, and accurately interpret evidence. Supervised fine-tuning is a crucial step in refining the core capabilities of the inference engine using the aforementioned labeled dataset. Specifically, the inference engine can be viewed as a complex sequence-to-sequence (Seq2Seq) model or a graph neural network model, with its input being an initial set of evidence and its output being intent hypotheses, follow-up questions, and evidence interpretations. By using the "input evidence" from the labeled data as the model's input, and the "reasonable hypotheses," "valid questions," and "accurate evidence interpretations" as the model's expected outputs, optimization algorithms such as backpropagation are employed to adjust the model parameters, making its output closer to the results annotated by experts. For example, the model can be trained to generate initial intent hypotheses consistent with expert judgments given a set of evidence, or, under specific hypotheses, to formulate follow-up questions similar to expert query strategies, and to make more accurate judgments on the support or conflict relationships of newly acquired evidence.

[0103] Simultaneously, the consistency between the final outcome of the case and the intermediate evaluations of the system is used as a reward signal to further optimize the decision-making process of the inference engine through reinforcement learning. Reinforcement learning is used here to optimize the decision-making strategy of the inference engine, particularly in multi-hop inference processes, such as how to select the optimal query path, when to terminate inference, and how to handle conflicting evidence. The system treats each multi-hop inference process of the inference engine as a sequential decision-making process, where each step affects the final intent recognition result. When the final judgment of the inference engine is consistent with the true result in the labeled data, or when its intermediate evaluation highly matches expert comments, the system provides a positive reward signal; otherwise, it provides a negative reward. Through reinforcement learning algorithms such as policy gradient or Q-learning, the inference engine can learn strategies to make better decisions in different situations, thereby improving its overall inference efficiency and accuracy.

[0104] Furthermore, for new vertical domains, by injecting a small number of labeled samples and domain knowledge documents from that domain, combined with prompting engineering and efficient parameter fine-tuning techniques, the core inference model can quickly adapt to the terminology and patterns of the new domain, achieving cross-domain intent recognition capability transfer. To enable the rapid deployment of intent recognition methods to new vertical domains, this is typically achieved as follows: First, collect a small number of representative labeled samples from the new domain, reflecting its unique terminology, behavioral patterns, and intent types. Second, inject knowledge documents from the new domain, such as industry dictionaries, regulatory documents, or expert guides, to enrich the model's domain semantic understanding. Based on this, prompting engineering techniques can be employed to guide the inference engine to better understand the tasks and data of the new domain by designing specific input templates or instructions. Simultaneously, combined with efficient parameter fine-tuning techniques, such as LoRA or Adapter, only a small number of newly added or modified parameters in the model are fine-tuned, rather than the entire model. This allows for efficient adaptation to the specific patterns of the new domain while maintaining the original general capabilities, achieving rapid knowledge transfer and lightweight model deployment.

[0105] By introducing a closed-loop learning process of model training and domain adaptation, this method effectively addresses the adaptability issue of intent recognition models when facing dynamically changing data patterns and new business domains. Specifically, by continuously collecting historical cases and constructing a labeled dataset containing complete reasoning processes, supervised fine-tuning of the inference engine is possible, significantly improving its ability to generate hypotheses, formulate questions, and interpret evidence, thereby enhancing the accuracy and efficiency of inference. Simultaneously, reinforcement learning strategies are used to guide the inference engine to optimize its decision-making process through reward signals, making it more intelligent and robust in multi-hop inference path selection and conflict handling. More importantly, for new vertical domains, the combination of a small number of labeled samples, domain knowledge documents, prompting engineering, and efficient parameter fine-tuning techniques enables the core inference model to quickly and efficiently adapt to the terminology and patterns of the new domain, achieving cross-domain intent recognition capability transfer. This not only significantly reduces the cost and time of model deployment in new domains but also ensures that the intent recognition system can maintain high performance and high accuracy over the long term, thereby improving the system's practicality and generalization ability.

[0106] In some embodiments described above, this application proposes an intent recognition method based on multi-source evidence fusion reasoning. This method can process heterogeneous evidence from multiple sources and perform multi-hop reasoning to identify intent. However, when applied to highly sensitive and complex scenarios such as high-risk behavioral intent recognition, the sheer number and complexity of potential abnormal behaviors, coupled with extremely high requirements for accuracy and interpretability, present challenges that may arise from relying solely on general methods. For example, failure to effectively focus on key risk indicators, or difficulty in quickly distinguishing genuine malicious intent from complex but compliant operations, may lead to inefficient evidence collection and even affect the validity and reliability of the final identification results.

[0107] To address this, this application further proposes an intent recognition method applicable to the field of high-risk behavioral intent recognition. This method uses one or a series of clues that trigger warning rules as input to generate an initial evidence set. In high-risk scenarios, intent recognition is often not an indiscriminate, comprehensive scan, but rather triggered by specific "clues." These clues can be triggered by pre-set risk warning rules of the system, such as large abnormal transactions, logins from different locations, or access to sensitive data, or they can be based on human reports or external intelligence. These clues, as initial, directional information, provide a starting point and direction for subsequent evidence collection and reasoning, helping to quickly focus on potential risk points and avoid blind searches.

[0108] The inference engine first generates competing hypotheses such as "suspected anomalous behavior," "suspected normal complex operation," and "suspected other compliant behavior." For high-risk behaviors, the inference engine generates a set of competing intent hypotheses covering multiple possibilities in the initial stage. For example, "suspected anomalous behavior" directly points to potential risks; "suspected normal complex operation" considers that some high-risk behaviors may simply be complex manifestations of legitimate business, avoiding misjudgment; and "suspected other compliant behavior" serves as a fallback, covering other non-abnormal situations that require further confirmation. This mechanism of multiple competing hypotheses helps to comprehensively assess risks, reduces bias caused by a single hypothesis, and provides clear verification targets for subsequent evidence collection.

[0109] Based on the "Hypothesis-Evidence Layer Mapping Strategy Library," for the "suspected anomalous behavior" hypothesis, priority is given to targeting and analyzing the full network graph evidence of the involved entities to discover complex structures. Related party attribute evidence is invoked to identify high-risk attribute associations, and entity behavior pattern mutation evidence is invoked to identify deviations from historical baselines. This is a specific evidence invocation strategy for the "suspected anomalous behavior" hypothesis. Full network graph evidence, by constructing and analyzing a complete network graph of the target entity and its related entities, can reveal hidden, multi-layered complex relationships, such as multi-layered nested company structures, abnormal fund flow networks, shared devices or IP addresses, etc. In high-risk scenarios, anomalous behavior is often masked by complex structures; therefore, analyzing the full network graph is crucial for discovering these complex structures. Related party attribute evidence refers to background information, risk labels, blacklist records, etc., of entities associated with the target entity (such as individuals, companies, devices, etc.). By invoking this type of evidence, it is possible to quickly identify whether the target entity is associated with known high-risk entities, thereby assessing the risk level of its behavior. Entity behavior pattern mutation evidence focuses on changes in the target entity's own behavior patterns. By comparing current behavior with the subject's historical behavioral baseline, significant deviations from the norm can be identified, such as abnormal fluctuations in transaction frequency, amount, time, and location. In high-risk scenarios, abrupt changes in behavioral patterns are often important signals of unusual intent, such as sudden account activity, a surge in transaction volume, or abnormal access permissions. The "hypothesis-evidence layer mapping strategy library" plays a crucial role here, guiding the inference engine to intelligently and prioritize the selection of key evidence types that best support or refute the hypothesis when faced with "suspected abnormal behavior," thereby improving the efficiency and relevance of evidence collection.

[0110] By employing multi-hop reasoning, the system progressively connects and integrates scattered information, dynamically fusing new findings at each step to ultimately determine the existence of a specific anomalous intent and generate a detailed, visualized path diagram and a structured report. In high-risk scenarios, evidence of anomalous intent is often scattered across different data sources, with hidden connections between them. The multi-hop reasoning mechanism iteratively expands the chain of evidence from initial clues, linking seemingly unrelated information. Each newly discovered piece of evidence is dynamically fused, updating the confidence levels for each hypothesis. Ultimately, the system not only provides a determination of whether a specific anomalous intent exists but also generates a detailed, visualized path diagram, clearly demonstrating the reasoning process and key evidence nodes from initial clues to the final conclusion, along with a structured report, providing strong support and interpretability for decision-makers.

[0111] Through the above technical solutions, this application can significantly improve the efficiency and accuracy of identifying high-risk behaviors. Specifically, by using clues that trigger warning rules as input, the system can quickly respond to potential risks, avoiding indiscriminate processing of massive amounts of data, thereby improving the targeting of the initial evidence set construction. The inference engine generates competing hypotheses such as "suspected abnormal behavior," "suspected normal complex operation," and "suspected other compliant behavior" in the initial stage. This multi-faceted hypothesis generation mechanism helps to comprehensively cover all possibilities of high-risk behaviors, effectively capturing anomalies while avoiding misjudgments of complex but compliant behaviors, thus improving the robustness of identification. Furthermore, for the "suspected abnormal behavior" hypothesis, the system can prioritize and analyze the full set of relational network graph evidence based on the "hypothesis-evidence layer mapping strategy library" to discover complex structures, related party attribute evidence to identify high-risk attribute associations, and evidence of sudden changes in subject behavior patterns to identify deviations from historical baselines. This highly customized and prioritized evidence retrieval strategy enables the reasoning process to quickly focus on the most discriminative types of evidence in high-risk scenarios, effectively penetrating the disguise of anomalous behavior and revealing its underlying intentions, thus greatly improving the ability and accuracy of detecting anomalous intentions. Ultimately, by progressively linking and integrating scattered information through multi-hop reasoning and dynamically fusing new findings at each step, this application can form a complete and credible chain of evidence, ultimately determining whether a specific anomalous intention exists. Simultaneously, the generation of detailed, visualized path diagrams and structured reports not only provides decision-makers with clear judgment results, but more importantly, it presents the reasoning process and key evidence support in an intuitive and interpretable manner. This has irreplaceable value for auditing, compliance review, and risk decision-making in high-risk scenarios, significantly enhancing the credibility and operability of the identification results.

[0112] In some of the embodiments described above in this application, an intent recognition method based on multi-source evidence fusion reasoning is proposed. This method can identify complex intents through multi-hop reasoning and dynamic evidence fusion. However, in practical applications, the challenge lies in how to efficiently and stably integrate this complex reasoning method into existing business processes, enabling it to adapt to different business scenario requirements, while ensuring continuous system optimization and uninterrupted operation.

[0113] In this regard, this application further proposes that the above method operates as a workflow of an automated service system. The process begins by receiving intent recognition requests and raw data from the upstream business system. Through steps S1 to S4, it ultimately returns structured intent recognition conclusions and reports to the upstream business system. This system has both real-time stream processing and batch processing modes, and can flexibly adjust the depth and breadth of multi-hop inference according to the timeliness requirements of the business scenario. Furthermore, the system provides a standard API interface to receive evidence data and return structured recognition results and evidence chains. The core inference and fusion services of this system are deployed in the form of microservices, and its knowledge base, policy base, and model support dynamic updates, ensuring continuous iteration and optimization of the recognition model without service interruption.

[0114] Specifically, operating the aforementioned intent recognition method as a workflow of an automated service system means that the entire intent recognition process is encapsulated as an end-to-end automated service, from receiving external requests to outputting results, without human intervention. This workflow can be driven by a service orchestration engine or message queue system. For example, an upstream business system triggers an intent recognition request by publishing a message or calling an interface, and the service system automatically consumes these requests and initiates the corresponding processing flow. The process begins by receiving the intent recognition request and raw data from the upstream business system. Through processing steps S1 to S4, it finally returns a structured intent recognition conclusion and report to the upstream business system. This means that the service system defines clear input and output interfaces. The intent recognition request can contain the target entity identifier to be analyzed, contextual information, and initial raw data. This raw data is transformed into standardized evidence units in step S1. After multi-hop reasoning in step S2 and dynamic evidence fusion calculation in step S3, the intent conclusion and interpretability evidence chain report are finally generated in step S4. These structured conclusions and reports are then returned to the upstream business system that initiated the request through a preset communication mechanism, realizing a business loop.

[0115] The system's core inference and fusion services are deployed as microservices. Its knowledge base, policy base, and model support dynamic updates, ensuring continuous iteration and optimization of the recognition model without service interruption. The microservice architecture breaks down the complex system into multiple independent, loosely coupled service units. For example, the inference engine can be a microservice, the evidence fusion module another, and knowledge base management and policy base management can also be independent microservices. This deployment method improves the system's scalability, fault tolerance, and development efficiency. Simultaneously, the dynamic updates of the knowledge base, policy base, and inference model mean that the system can be updated without service interruption when new knowledge, optimized policies, or improved model versions become available. This can be achieved through technologies such as hot reloading, blue-green deployment, or canary deployment, ensuring the system can continuously learn and adapt to new data patterns and business needs without causing any interruption to ongoing business operations. By running the aforementioned intent recognition method as a workflow of an automated service system and providing standard API interfaces, seamless integration with upstream business systems is achieved, greatly enhancing the method's practicality and operability. The system features both real-time stream processing and batch processing modes, enabling it to flexibly adjust the depth and breadth of multi-hop inference based on varying timeliness requirements of different business scenarios. This ensures both response speed and the comprehensiveness and accuracy of inference. Furthermore, the core inference and fusion services are deployed as microservices, coupled with a dynamic update mechanism for the knowledge base, strategy base, and models. This ensures continuous iterative optimization and adaptive adjustments without service interruption, effectively addressing constantly changing business environments and data patterns, and significantly improving the accuracy, efficiency, and stability of intent recognition. Example

[0116] The following example will provide a more detailed explanation of the above technical solution:

[0117] In a financial risk control scenario, the system needs to identify whether user A has high-risk behavioral intentions. Traditional intent recognition methods, whether based on static rules or end-to-end machine learning models, face problems such as fragmented evidence, ambiguous information, lack of proactive exploration capabilities, and opaque decision-making processes, leading to inaccurate identification and difficulty in interpreting complex and hidden intentions. This method effectively solves these technical problems through multi-source evidence fusion reasoning.

[0118] First, in step S1, when the system receives a warning clue about user A, for example, if user A recently made a large transfer with an unusual counterparty, it will initiate the evidence acquisition and standardization process. The system will acquire information from multiple heterogeneous data sources, including: Static attribute evidence: collecting user A's basic identity information, historical risk tags, account opening time, etc., and assigning weights based on their timeliness. Dynamic behavioral evidence: extracting user A's transaction logs, login records, fund flows, and other raw behavioral data, and transforming them into a time-series behavioral graph network. From this network, periodic patterns, dense occurrence patterns, and inactive period occurrence patterns are identified. Specifically, these include fixed monthly repayments, multiple small transfers within a short period, and unusual logins and large transactions late at night. The feature vectors of these patterns serve as evidence content, and a correlation score is assigned based on their deviation from user A's historical behavioral baseline. For example, a high deviation from large transactions late at night results in a high correlation score. Semantic expression evidence: performing natural language processing on user A's communication records with counterparties to extract specific semantic patterns, entity references, and sentiment tendencies. Based on the sentiment intensity analysis results, sentiment tendency weights are assigned. External verification evidence is obtained by querying authoritative list databases and public information databases, such as enterprise registration information, through an interface. The query results are then matched with user A and their associates. If a match is successful, evidence units containing relevant information categories and levels are generated. For example, "User A's associate B is on a blacklist, with a level of high risk." All this raw evidence is processed and transformed into standardized evidence units, each containing evidence content, a source reliability score, and relevance to potential intent. These standardized evidence units are used to construct an initial evidence set, represented in the form of an evidence reasoning graph model, where nodes represent entities or evidence units, and edges represent relationships between them. Next, in step S2, the initial evidence set is input into a neural network-based inference engine, performing multi-hop inference through multiple rounds of proactive queries. This overcomes the limitations of traditional methods that passively rely on the initial evidence set and cannot actively explore information.

[0119] Step S2.1: The inference engine first performs contextual encoding and integration on the initial evidence set, and based on this, generates multiple alternative initial behavioral intent hypotheses, such as "suspected abnormal behavior," "suspected normal complex operation," and "suspected other compliant behavior." Multiple Hypothesis Competition and Rapid Verification: Before entering deep inference, the system initiates a rapid verification phase. For each of the above initial hypotheses, one to two rounds of lightweight evidence queries are initiated in parallel. For example, for the "suspected abnormal behavior" hypothesis, the system will quickly retrieve whether user A is on the internal high-risk list or whether there are any recently marked abnormal transaction records through an efficient indexing interface. A lightweight scoring model is used to initially evaluate and rank the hypotheses, selecting the top-K hypotheses with the highest scores to enter the subsequent deep multi-hop inference main process, while hypotheses with too low a score are downgraded or archived, thereby improving overall inference efficiency.

[0120] Step S2.2: For the selected "suspected anomalous behavior" hypotheses, the inference engine queries the pre-built "hypothesis-evidence layer mapping strategy library". This strategy library is stored in the form of a knowledge graph, where nodes include the intent hypothesis of the behavior, the evidence level category, the specific evidence type, and the edges connecting them. The weight on the edge indicates the importance of that type of evidence for verifying a specific hypothesis. For example, for the "suspected anomalous behavior" hypothesis, the strategy library may instruct the priority to call "full-scale association network graph evidence" to discover complex structures, followed by "related party attribute evidence" to identify high-risk attribute associations, and then "subject behavior pattern mutation evidence".

[0121] Step S2.3: Based on the target evidence hierarchy and key features parsed from the strategy library, the inference engine constructs structured query predicates or generates precise natural language query instructions. For example, it generates the instruction "Query the transaction network diagram of user A and all its related parties, and identify whether there are indirect fund transfer paths exceeding three layers." This instruction drives the corresponding external data pipeline through the application programming interface to perform targeted data collection or real-time calculation of the target data.

[0122] Step S2.4: The collected raw data is feature-extracted and vectorized according to the standardized pattern of the corresponding evidence level, encapsulated into new standardized evidence units, and injected into or updated the corresponding nodes and edges in the evidence reasoning graph model. Subsequently, the reasoning engine performs further reasoning on the enhanced graph model, analyzing the support, conflict, or supplementary relationships between the new evidence units and existing evidence and various competing hypotheses through attention mechanisms or graph neural network algorithms, and updating the confidence probability distribution of each hypothesis accordingly. For example, if it is found that user A's transaction network does indeed have a multi-layered complex structure, the confidence of the "suspected abnormal behavior" hypothesis will be significantly increased.

[0123] Step S2.5: The system repeats steps S2.2 to S2.4, forming a closed-loop reasoning loop guided by dynamic hypotheses, planned by a strategy library, and executed by proactive data acquisition. This loop continues until a preset termination condition is met, such as the confidence level of the "suspected abnormal behavior" hypothesis exceeding a set certainty threshold (e.g., 90%), or all preset levels of evidence have been traversed.

[0124] In step S3, the system performs dynamic evidence fusion calculations. This addresses the problem that traditional methods cannot effectively quantify evidence uncertainty and handle evidence conflicts. In each iteration of step S2, when a new standardized evidence unit (e.g., the evidence "User A's transaction network has a complex structure" obtained from S2.4) is injected, the system initiates a fusion process based on Dempster-Shafer evidence theory. A Basic Probability Assignment Function (BPA) is constructed for each new evidence unit. Based on its source reliability score and its relevance to intent, the reliability quality is assigned to a corresponding subset and uncertainty region within an identification framework composed of mutually exclusive intent assumptions. For example, for the evidence "User A's transaction network has a complex structure," its BPA might assign 0.7 reliability to {"suspected abnormal behavior"}, 0.1 reliability to {"suspected abnormal behavior", "suspected normal complex operation"}, and the remaining 0.2 reliability to the uncertainty region. The reliability allocation ratio of the uncertainty region is dynamically adjusted by the evidence source reliability score. The system employs an improved synthesis rule to synthesize the BPA (Browser Aptitude Test) of all new evidence in the current round with the global BPA generated in the previous iteration. When different evidence sources are highly conflicting—for example, one piece of evidence strongly supports "abnormal behavior," while another new piece of evidence strongly supports "normal operation"—the improved rule uses a weighted redistribution strategy based on Jousselme distance to reasonably handle conflicting evidence and avoid the non-intuitive results that traditional DS (Digital Strategy) theory may produce when there is conflict. After each synthesis operation, the system recalculates the trust function, likelihood function, and confidence interval for all single intent hypotheses, and uses the hypothesis with the highest confidence and its confidence interval as intermediate results to guide the generation of the next-hop inference problem.

[0125] Finally, in step S4, based on the global fusion results, the system outputs the highest confidence intention conclusion and generates an interpretable evidence chain report containing a complete multi-hop reasoning path and key evidence support. This solves the problem of lack of interpretability in traditional methods. The system backtracks and parses the complete multi-hop reasoning path, extracting the core questions raised in each round of reasoning (e.g., "Does a complex transaction network exist?"), the specific evidence invoked to answer the questions (e.g., "User A's transaction network diagram"), the impact of this evidence on the confidence of each hypothesis (e.g., the confidence of "suspected abnormal behavior" increases from 0.6 to 0.8), and the intermediate conclusions after that round of reasoning. The system uses a combination of template filling and natural language generation to organize the above structured information into a coherent narrative report. The report clearly elucidates the reasoning logic chain from the initial clues to the final conclusion and uses visual charts to help display the correlation network and changing trends of key evidence. The final report should include at least the final intent determination result (e.g., “User A has an anomalous intent”), the overall confidence level of the determination result (e.g., 92%), a list of core evidence supporting the result and their contribution analysis (e.g., “complex transaction network” contributes 30%, “high-risk attributes of related parties” contribute 25%), and the main competing hypotheses that were considered but ultimately excluded (e.g., “suspected normal complex operation”) and the key reasons for their exclusion (e.g., “lack of evidence to support that it is a normal operation and conflict with evidence of anomalous behavior”).

[0126] The entire method operates as an automated service system workflow. It begins by receiving intent recognition requests and raw data from the upstream business system. Through the processing steps S1 to S4 described above, it finally returns structured intent recognition conclusions and reports to the upstream business system. The system supports both real-time stream processing and batch processing modes, and can flexibly adjust the depth and breadth of multi-hop inference according to the timeliness requirements of business scenarios. Furthermore, the system includes a closed-loop learning process for model training and domain adaptation. It collects historical cases for supervised fine-tuning and reinforcement learning, and utilizes a small number of labeled samples and domain knowledge to quickly adapt to new domains, ensuring continuous iteration and optimization of the recognition model.

[0127] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that an article or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such an article or apparatus. Without further limitations, an element defined by the phrase "comprising..." does not exclude the presence of other identical elements in the article or apparatus that includes that element.

[0128] The above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. The present invention has been described in detail with reference to preferred embodiments. Those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications and substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An intent recognition method based on multi-source evidence fusion reasoning, characterized in that, Includes the following steps: Step S1: Obtain multi-source heterogeneous evidence, including at least text, behavioral sequences, and external knowledge data, and transform it into standardized evidence units containing evidence content, source reliability scores, and intent relevance. Construct an initial evidence set based on the evidence reasoning graph model. Step S2: Input the initial evidence set into the neural network-based inference engine and perform multi-hop inference through multiple rounds of active query: the first hop analyzes the consistency of the evidence and generates preliminary intent hypotheses and follow-up questions, and subsequent hops dynamically generate new questions based on new evidence to obtain supplementary information in a targeted manner until the termination condition is met; Specifically, it includes the following sub-steps: Step S2.1: The neural network-based inference engine receives the initial evidence set, performs context encoding and integration of the evidence units in the set through its encoder, and generates at least two alternative intention hypotheses of initial behavior formalized as probability distributions from its decoder or a specific output head based on this context vector. Step S2.2: For each intent hypothesis generated in step S2.1, query a pre-built hypothesis-evidence layer mapping strategy library; this strategy library stores the mapping relationship and invocation weight between different categories of intent hypotheses and multi-level evidence types in a vector or graph structure; based on this mapping relationship, parse and determine a priority-ordered sequence of evidence layers to be invoked and the key evidence pattern features under each layer for each intent hypothesis. Step S2.3: The inference engine constructs structured query predicates or generates precise natural language query instructions based on the target evidence level and key features parsed in step S2.2; these instructions drive the corresponding peripheral data pipelines through the application programming interface to perform targeted collection, retrieval, or real-time calculation of the target data. Step S2.4: Extract and vectorize the raw data collected in step S2.3 according to the standardization mode of the corresponding evidence level, encapsulate it into new standardized evidence units, and inject or update the corresponding nodes and edges in the evidence reasoning graph model; then, the reasoning engine performs reasoning on the evidence reasoning graph model again, and analyzes the support, conflict or supplementary relationship between the new evidence units and existing evidence and various competing hypotheses through attention mechanism or graph neural network algorithm, and updates the confidence probability distribution of each hypothesis accordingly; Step S2.5: Repeat steps S2.2 to S2.4 to form a closed-loop reasoning loop guided by dynamic assumptions, planned by a strategy library, and executed by proactive data acquisition; This cycle continues until a preset termination condition is met, including the number of cycles reaching the upper limit, all preset levels of evidence being traversed, or the confidence level of a dominant hypothesis exceeding a set certainty threshold. Step S3: Dynamic Evidence Fusion Calculation: In each iteration of Step S2, after a new evidence unit is injected, a fusion process based on Dempster-Shafer evidence theory is initiated; a basic probability allocation function is constructed for each new evidence unit, and the confidence quality is allocated to a corresponding subset and uncertainty domain in an identification framework composed of mutually exclusive intent assumptions based on its reliability and relevance; an improved synthesis rule is adopted to synthesize the BPA of all new evidence in this round with the global BPA generated in the previous iteration. This synthesis rule adopts a weighted redistribution strategy based on Jousselme distance when there is evidence conflict. After synthesis, the updated hypothesis confidence functions and likelihood functions are output as the current confidence state. Step S4: Output the highest confidence intention conclusion based on the global fusion result, and generate an interpretable evidence chain report containing a complete multi-hop reasoning path and key evidence support.

2. The method according to claim 1, characterized in that, In step S1, the multi-source heterogeneous evidence specifically includes: Static attribute evidence, whose standardized evidence unit includes the basic attribute description of the target subject, historical state labels and their corresponding timeliness weights; Dynamic behavioral evidence is obtained by transforming the original behavioral logs into a time-series behavioral graph network, extracting feature vectors of periodic patterns, dense occurrence patterns, and inactive period occurrence patterns as evidence content, and assigning correlation scores to the feature vectors based on the pattern deviation. Semantic expression evidence is obtained by analyzing relevant communication records and documents using natural language processing technology, extracting specific semantic patterns, entity references, and sentiment tendencies contained therein as semantic evidence, and assigning sentiment tendency weights based on the sentiment intensity analysis results. External verification evidence is obtained by querying or importing data from authoritative list databases and public information databases through interfaces, and matching it with the current target subject. If the match is successful, evidence units containing relevant information categories and levels are generated.

3. The method according to claim 1, characterized in that, The method further includes the step of constructing and updating the hypothesis-evidence layer mapping strategy library, specifically: The strategy base is stored in the form of a knowledge graph, whose nodes include the intent hypothesis of the behavior, the evidence level category, the evidence type, and the edges connecting them. The weights on the edges represent the importance weights of the evidence in verifying a specific hypothesis. The strategy base is initialized based on domain expert experience rules and historical typical cases. During system operation, the entire link log of each multi-hop inference is continuously recorded, including the initial hypothesis, the evidence invoked, changes in the hypothesis confidence, and the final conclusion. Through an offline analysis module, the contribution of different pieces of evidence to the final confirmation or rejection of a hypothesis is periodically evaluated, and the weights of the corresponding edges in the strategy base are dynamically adjusted using a reinforcement learning algorithm, thereby achieving the self-optimization and evolution of the hypothesis-evidence layer mapping strategy base.

4. The method according to claim 1, characterized in that, The dynamic evidence fusion calculation process described in step S3 specifically includes: A basic confidence assignment function is defined for each standardized evidence unit. This function assigns confidence values ​​to a subset of a set of mutually exclusive and complete intention hypotheses, while reserving a portion of confidence for the uncertainty domain. The confidence assignment ratio for the uncertainty domain is dynamically adjusted by the reliability score of the evidence source. In each round of multi-hop inference, the basic confidence assignment function of the newly generated evidence unit is synthesized with the current global basic confidence assignment function. When there is a high degree of conflict between different evidence sources, a conflict redistribution synthesis rule based on the Jousselme distance or reliability weighting between evidence is adopted to reasonably handle conflicting evidence. After each synthesis operation, the trust function, likelihood function, and confidence interval of all individual intention hypotheses are recalculated, and the hypothesis with the highest confidence and its confidence interval are used as intermediate results to guide the generation of the next-hop inference problem.

5. The method according to claim 1, characterized in that, The method, after step S2.1 and before step S2.2, also includes a step of multiple hypothesis competition and fast verification, specifically as follows: After generating multiple competing initial behavior intention hypotheses, a rapid verification phase is initiated; in this phase, one to two rounds of lightweight evidence queries are initiated in parallel for each hypothesis. These queries use a pre-built, efficient indexing interface to target and retrieve key discriminative evidence that strongly supports or refutes the hypothesis. Subsequently, based on the small amount of key evidence obtained quickly, a lightweight scoring model is used to initially evaluate and rank the hypotheses. The top-K hypotheses with the highest scores, where K is an integer greater than or equal to 1, are selected as the winning hypotheses and enter the subsequent deep multi-hop inference main process. At the same time, hypotheses with low scores are downgraded or archived, thereby improving the computational efficiency of the overall inference process while ensuring the possibility of coverage.

6. The method according to claim 1, characterized in that, The step S4, which generates an interpretable chain of evidence report containing a complete multi-hop reasoning path and key evidence support, specifically includes: The system backtracks and analyzes the complete multi-hop reasoning path, extracting the core questions raised in each round of reasoning, the specific evidence invoked to answer the questions, the impact of the evidence on the confidence of each hypothesis, and the intermediate conclusions after that round of reasoning. Using a combination of template filling and natural language generation, the extracted information is organized into a coherent narrative report. The report clearly elucidates the logical chain of reasoning from the initial clues to the final conclusion, and uses visual charts to help display the correlation network and changing trends of key evidence. The final report includes at least the final intent determination result, the overall confidence level of the determination result, a list of core evidence supporting the result and their contribution analysis, and the main competing hypotheses that were considered but ultimately excluded and the key reasons for their exclusion.

7. The method according to claim 1, characterized in that, The method also includes a closed-loop learning process for model training and domain adaptation, with the following specific steps: Collect completed historical intent recognition cases to form a labeled dataset containing complete input evidence, multi-hop reasoning intermediate processes, final judgment results, and comments; The inference engine is then subjected to supervised fine-tuning using this dataset to optimize its ability to generate reasonable hypotheses, ask valid questions, and accurately interpret evidence. Simultaneously, the consistency between the final case results and intermediate system evaluations is used as a reward signal, and a reinforcement learning strategy is employed to further optimize the inference engine's decision-making process. For new vertical domains, a small number of labeled samples and domain knowledge documents are injected, combined with prompting engineering and efficient parameter fine-tuning techniques, enabling the core inference model to quickly adapt to the terminology and patterns of the new domain, achieving cross-domain intent recognition capability transfer.

8. The method according to claim 2, characterized in that, The method is applied to intent recognition in the field of high-risk behavioral intent recognition, and its specific steps include: Using one or more clues that trigger warning rules as input, an initial set of evidence is generated. The reasoning engine first generates one of the following competing hypotheses: suspected abnormal behavior, suspected normal complex operation, or suspected other compliant behavior. Based on the hypothesis-evidence layer mapping strategy library, for the suspected abnormal behavior hypothesis, it prioritizes the targeted invocation and analysis of the full-scale relational network graph evidence of the involved subjects to discover complex structures, invokes the attribute evidence of related parties to identify high-risk attribute associations, and invokes the subject's behavioral pattern mutation evidence to identify deviations from historical baselines. Through multi-hop reasoning, it gradually associates and integrates various types of scattered information, dynamically integrates new discoveries at each step, and finally determines whether there is a specific abnormal intent, generating a detailed visualized relational path diagram and a structured report.

9. The method according to any one of claims 1 to 8, characterized in that, The method operates as a workflow of an automated service system: the process begins by receiving intent recognition requests and raw data from the upstream business system, and through the processing steps S1 to S4, finally returns a structured intent recognition conclusion and report to the upstream business system; the system has both real-time stream processing and batch processing modes, and can flexibly adjust the depth and breadth of multi-hop inference according to the timeliness requirements of the business scenario; the system provides a standard API interface to receive evidence data and return structured recognition results and evidence chains. The core reasoning and fusion services of the system are deployed in the form of microservices. Its knowledge base, strategy base and model support dynamic updates, ensuring continuous iteration and optimization of the recognition model without interrupting the service.

Images