Method for realizing accurate intention recognition in AI question and answer application

Abstract

The invention discloses a method for realizing accurate intention recognition in an AI question and answer application. The method comprises the following steps: pre-defining a semantic analysis dimension; constructing a semantic analysis dimension tree before system deployment based on semantic analysis dimensions, and defining a hierarchical relationship between the dimensions and a semantic path from a root node to a leaf node; making a dimension value rule, deeply associating the dimension value rule with nodes and semantic paths in the semantic analysis dimension tree, and constructing a dynamic rule constraint engine; and analyzing user input based on a structured framework of a semantic analysis dimension tree in combination with a dimension value rule, generating a structured semantic intention, and performing semantic task distribution and execution. According to the method, a hierarchical semantic analysis dimension tree is constructed, intention analysis is converted from point mapping to structured path searching, and the problem that structured analysis of the composite intention is missing is effectively solved; through deep collaboration of a dynamic rule constraint engine and a semantic analysis dimension tree, logic consistency and stable output of an analysis process are ensured.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of natural language processing and knowledge retrieval, in particular to a method for realizing accurate intent recognition in AI question and answer application. BACKGROUND

[0002] In recent years, with the strong capabilities of Large Language Models (LLM) in natural language understanding and generation tasks, LLM has become a core technology component in many industry applications. However, general-purpose large language models have inherent limitations when dealing with specific domain, real-time, or private knowledge issues, such as knowledge update lag, possible "hallucinations", and inability to access non-training data. Retrieval-Augmented Generation (RAG) technology has emerged as a key bridge connecting large language models and external knowledge bases. AI question and answer applications implemented through RAG can retrieve relevant information from large-scale, updatable knowledge sources and provide it as context to large language models, significantly improving the accuracy, timeliness, and credibility of generated content.

[0003] Currently, there are two main implementation paths for intent recognition technology in AI question and answer applications:

[0004] One of them is "general-purpose model encoding + single semantic analysis dimension matching", which uses vector similarity or traditional retrieval algorithms to perform nearest neighbor search in the knowledge base, and the closest result is output as the intent;

[0005] The other is "pre-training intent classification model", which first trains an intent classification model to divide user queries into predefined categories before retrieval, and then performs directional retrieval according to the categories.

[0006] The existing technology has the following core defects:

[0007] (1) Low accuracy of complex or ambiguous intent understanding: For complex intents or ambiguous expressions with multiple superpositions, the understanding accuracy is usually less than 90%;

[0008] (2) High implementation cost: If the accuracy is improved by increasing model parameters or expanding training data, a large amount of computing power and time cost is required, making it difficult to balance accuracy and cost.

[0009] From a technical point of view, the root cause of the above defects is that the existing technology outputs "point" vectors or labels, and this "point mapping" method loses the hierarchy and combination relationship between the internal semantic elements of complex intents, i.e., there is a "lack of structured analysis". SUMMARY

[0010] In order to solve the above technical problems, the purpose of the present application is to provide a method for realizing accurate intent recognition in AI question and answer application, so as to overcome the technical problems such as "structured analysis missing" caused by the lack of effective structured intermediate representation layer when the user input contains multiple semantic constraints in the existing question and answer application; by introducing the core structured analysis framework of "hierarchical semantic analysis dimension tree", and making it deeply cooperate with "dynamic rule constraint engine" and "structured compliance oriented scenario model", the high-precision and high-consistency intent structured analysis is realized.

[0011] The present application solves the above problems by the following technical solutions:

[0012] A method for realizing accurate intent recognition in AI question and answer application, comprising:

[0013] Step S101: for a target semantic application scenario, a set of explicit and independent semantic analysis dimensions is defined in advance, the semantic analysis dimensions at least include two of domain dimension, requirement dimension, index dimension, time dimension and object dimension;

[0014] Step S102: based on the semantic analysis dimensions, a static semantic analysis dimension tree with strict hierarchical relationship is constructed before system deployment, the semantic analysis dimension tree includes a root node representing the most abstract user intent, intermediate nodes representing each semantic analysis dimension, and leaf nodes representing the specific values of each dimension, the semantic analysis dimension tree defines the hierarchical relationship between the dimensions and the semantic path from the root node to the leaf node;

[0015] Step S103: the dimension value rule is formulated and associated with the nodes and semantic paths in the semantic analysis dimension tree, a dynamic rule constraint engine is constructed, the dimension value rule includes node activation rule and path constraint rule;

[0016] Step S104: receiving user input, based on the structured framework of the semantic analysis dimension tree, and combining the dimension value rule to analyze the user input, generating a structured semantic intent;

[0017] S105: according to the generated structured semantic intent, the semantic task distribution and execution are carried out.

[0018] As a further improvement of the present application, in the step S101, the specific method for defining semantic analysis dimensions comprises:

[0019] S101.1: business research is carried out on the target application scenario, and historical user query records are analyzed to mine the semantic elements implied in the user query;

[0020] S101.2: constructing an analysis system comprising at least two independent semantic analysis dimensions based on the semantic elements, each semantic analysis dimension representing a specific analysis perspective of the user intent;

[0021] S101.3: configuring attribute information for each defined semantic analysis dimension, the attribute information comprising: dimension name, value set, value data type, default value strategy and priority weight.

[0022] As a further improvement of the application, in the step S102, the method of constructing the semantic analysis dimension tree comprises:

[0023] S102.1: dimension weight ordering and logical layering;

[0024] S102.2: tree structure construction of the semantic analysis dimension tree;

[0025] S102.3: path semantic definition and registration.

[0026] As a further improvement of the application, the dimension weight ordering and logical layering specifically comprises: logical layering and weight ordering of the semantic analysis dimensions according to business domain knowledge to determine the dependency and dependence relationship between dimensions;

[0027] And / or the tree structure construction of the semantic analysis dimension tree specifically comprises: based on the logical layering result, explicitly constructing a tree structure composed of root nodes, intermediate nodes and leaf nodes;

[0028] And / or the path semantic definition and registration specifically comprises: explicitly defining the complete business semantics for each legal path from the root node to the leaf node in the semantic analysis dimension tree, and registering it as an identifiable intent pattern.

[0029] As a further improvement of the application, in the step S103, the specific method of formulating dimension value rules comprises:

[0030] S103.1: node activation rule formulation and binding;

[0031] S103.2: path constraint rule formulation and association;

[0032] S103.3: rule-tree depth association storage and engine construction.

[0033] As a further improvement of the application, the node activation rule formulation and binding specifically comprises: formulating fine-grained activation rules for key nodes in the semantic analysis dimension tree, and directly binding to the corresponding nodes to form node activation rules; the node activation rules comprise: keyword matching rules, sentence pattern matching rules and semantic similarity trigger rules;

[0034] The path constraint rule is formulated and associated, and the specific method is that: based on the topological relationship of the semantic analysis dimension tree, the path constraint rules between different paths in the semantic analysis dimension tree are formulated, and the path constraint rules are associated to the relevant path or node combination; meanwhile, the following rules need to be followed when specifying the path constraint rules: mutual exclusion path rule, dependent path rule and conflict resolution and priority rule;

[0035] The rule-tree depth association storage and engine construction, and the specific method is that: the node activation rule and the path constraint rule are stored in depth association with the node ID and path hash of the semantic analysis dimension tree, and a "rule-node-path" three-dimensional mapping network is constructed; the three-dimensional mapping network constitutes the knowledge base of the dynamic rule constraint engine.

[0036] As a further improvement of the application, the keyword matching rule is: defining the keywords, phrases or synonyms that must appear in the user input to activate the corresponding node;

[0037] And / or the sentence pattern matching rule is: defining a set of text patterns for matching specific intent expression methods;

[0038] And / or the semantic similarity trigger rule is: for abstract concept nodes, defining a trigger threshold based on vector similarity.

[0039] As a further improvement of the application, the mutual exclusion path rule is: defining that the paths in conflict in semantics cannot be activated at the same time;

[0040] And / or the dependent path rule is: defining that the activation of a node / path must be based on the activation of another node / path;

[0041] And / or the conflict resolution and priority rule is: when the user input can activate multiple conflict paths, providing a decision rule based on business weight to resolve through the pre-defined domain priority rule.

[0042] As a further improvement of the application, in the step S104, the specific method of analyzing the user input includes:

[0043] S104.1: semantic pre-processing and feature extraction of user input;

[0044] S104.2: after the semantic features of the pre-processing, the semantic analysis dimension tree is analyzed in coordination, the user input, the complete semantic analysis dimension tree structure information, and the current session context are jointly organized into prompt information; and reasoning in the structure space of the semantic analysis dimension tree is guided, and the preliminary path hypothesis and its confidence are output.

[0045] In the path reasoning process or after the path reasoning process, the dynamic rule constraint engine is called in real time to perform conflict detection, dependency verification and confidence calibration on the preliminary path hypothesis;

[0046] S104.3: output one or more high-confidence and complete semantic paths after rule verification and iterative optimization, and specific values of each leaf node in the semantic path, to form a structured semantic intent.

[0047] As a further improvement of the present application, in the step S105, the specific method of semantic task distribution and execution is:

[0048] mapping the generated structured semantic intent explicit semantic path and node value into a precise query condition of a database or a calling parameter of a specific task executor;

[0049] performing retrieval or task corresponding to the query condition or calling parameter, and fusing the matching result and the similarity retrieval result based on the semantic vector;

[0050] sorting the fused result according to a preset weight, and returning the most relevant knowledge fragment or task execution result.

[0051] Compared with the prior art, the present application has the following advantages and beneficial effects:

[0052] (1) The present application proposes and constructs a "hierarchical semantic analysis dimension tree", which is used as the only core technical framework for solving the above-mentioned structured analysis problem of complex intent. The technical essence of the framework is: a plurality of predefined semantic analysis dimensions such as domain, requirement, index, etc. are organized through a tree-shaped data structure, the parent-child, brother and other hierarchical relationships between the dimensions are clearly defined, and a complete semantic path from the root node representing the most abstract user intent to the leaf node representing the specific value of each dimension is defined. The analysis process of the user query is redefined as a "path finding" task under this tree-shaped framework. Therefore, the technical essence of this method is to construct and apply the structured analysis framework of "hierarchical semantic analysis dimension tree" to accurately analyze the user input intent and the value of each dimension, and to realize low-cost and high-precision recall.

[0053] (2) The present application takes explicit multi-dimensional semantic analysis dimensions such as field, requirement, index, and time as the core, rather than a single dimension; a hierarchical semantic analysis dimension tree is constructed, the tree structure of the root node, the intermediate node, and the leaf node defines the dependency, dependence, and path relationship between the dimensions, and the flat dimension list is converted into a three-dimensional structure with semantic depth and logical hierarchy, which is the basic framework for realizing the structured analysis of complex intent. And formulate the dimension value rule including keyword matching, dimension consistency check, and constrain the analysis direction of the large model; the technical core support is the dynamic rule constraint engine technology, which is bound with the depth of the semantic analysis dimension tree node, not only including the surface node activation rule, but also including the path constraint rule such as mutual exclusion, dependence, and conflict resolution in a deeper level, which ensures the overall semantic logic consistency in the multi-path reasoning process, which is the key mechanism to avoid contradictory analysis results and improve output stability.

[0054] (3) The present application effectively solves the problem of lack of structured analysis of complex intent by constructing a "hierarchical semantic analysis dimension tree" and changing the intent analysis from "point mapping" to "structured path finding"; through the depth cooperation of the dynamic rule constraint engine and the semantic analysis dimension tree, the logic consistency of the analysis process is ensured, and the output is stable; in the complex intent scenario, the recognition accuracy is improved to more than 95%, without additional large-scale training cost; it is suitable for complex knowledge base, ChatBI index hit, MCP tool retrieval and other scenes, and has wide applicability and practicality. BRIEF DESCRIPTION OF DRAWINGS

[0055] Figure 1 A method flowchart for realizing accurate intent recognition in an AI question and answer application. DETAILED DESCRIPTION

[0056] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0057] Embodiment 1:

[0058] In combination with the accompanying Figure 1 The present application provides a method for realizing accurate intent recognition in an AI question and answer application. The core innovation of the method is to construct and apply a "hierarchical semantic analysis dimension tree" as the structured framework and unique core technology route of the entire intent analysis process, and to innovate the analysis paradigm from the "point mapping" of the prior art such as vector matching and single label classification to "structured path finding and matching". The overall process of the method mainly includes the following steps:

[0059] S101: defining explicit multi-semantics analysis dimensions;

[0060] For target semantic application scenarios such as complex knowledge base, ChatBI, MCP tool, a set of explicit, independent semantic analysis dimensions such as two or more of the fields, requirements, indicators, time, objects, etc. are defined in advance. This step prepares the basic semantic elements for building a static, field-specific core framework, emphasizing its prior definition, which belongs to the routine technical preparation and is briefly described here. It specifically includes the following sub-steps: analyzing scene requirements, building dimension system, configuring dimension attributes, etc.

[0061] The specific method for defining semantic analysis dimensions includes:

[0062] S101.1: Conducting business research on the target application scenario and analyzing historical user query records to mine semantic elements implied in user queries;

[0063] S101.2: Based on the semantic elements, building an analysis system containing at least two independent semantic analysis dimensions, each representing a specific analysis perspective of user intent;

[0064] S101.3: Configuring attribute information for each defined semantic analysis dimension, including dimension name, value set, value data type, default value strategy, and priority weight.

[0065] S102: Building a "hierarchical semantic analysis dimension tree" - the entity construction of the core innovative framework;

[0066] This step is the first technical focus and innovation point of the present application, aiming to systematically transform the discrete, flat semantic analysis dimensions defined in S101 into a static semantic analysis dimension tree with strict hierarchy, path dependence, and business logic relationships before the system goes online and any user query occurs, thereby providing a basic framework for solving the "structured analysis missing" problem. The nodes of the semantic analysis dimension tree are not dynamically generated "sub-questions", but pre-defined business semantic analysis dimensions and their legal values. The semantic analysis dimension tree includes a root node representing the most abstract user intent, intermediate nodes representing each semantic analysis dimension, and leaf nodes representing specific values of each dimension. The semantic analysis dimension tree defines the hierarchical relationship between dimensions and the semantic path from the root node to the leaf node. The construction of the semantic analysis dimension tree includes the following three sub-steps:

[0067] S102.1: Dimension weight ordering and logical layering;

[0068] Instead of simply listing dimensions, the semantic analysis dimensions are logically layered and weighted according to business domain knowledge. For example, the "domain dimension" and "demand dimension" are defined as core first-level dimensions as the main trunk of the tree; the "index dimension" and "product dimension" are defined as second-level key parameter dimensions as the main branches; and the "time dimension" and "space dimension" are defined as third-level auxiliary constraint dimensions. This hierarchical definition defines the dependency relationship between dimensions, which is the premise of building a tree structure.

[0069] S102.2: Tree structure construction of semantic analysis dimension tree;

[0070] Based on the logical layering results, a tree structure composed of root nodes, intermediate nodes and leaf nodes is explicitly constructed, and the meaning of each node is as follows:

[0071] Root node: represents the most abstract "user intent";

[0072] Intermediate node: represents each semantic analysis dimension itself, such as "domain dimension" node and "demand dimension" node. The parent-child relationship between nodes reflects the hierarchy of dimensions, such as the "financial domain" node, which is a child node of the "domain dimension" node and a parent node of the "demand dimension" node;

[0073] Leaf node: represents the specific value of each dimension, such as "sales", "2025", and "A product". The leaf node is the final landing point of intent analysis;

[0074] Overall structure: A complete tree path may be: the root node is the intent, the domain dimension is finance, the demand dimension is data query, the index dimension is sales, the product dimension is smart TV, and the time dimension is 2025. This path itself defines a complete and structured intent template.

[0075] S102.3: Path semantic definition and registration;

[0076] For each legal path in the semantic analysis dimension tree from the root node to the leaf node, the complete business semantics is explicitly defined and registered as a recognizable intent pattern by the system. For example, the semantics of the path "finance → trend analysis → profit rate → quarter" is defined as "financial profit rate quarter trend analysis intent". At this point, the analysis of user intent is formally redefined as: in the "hierarchical semantic analysis dimension tree", find one or more most matching semantic paths from the root node to the leaf node for user input. The tree structure provides a structured search space and target blueprint for the entire scheme.

[0077] In summary, the technical logic of constructing the "hierarchical semantic analysis dimension tree" is to organize independent semantic analysis dimensions into a structured semantic space through a hierarchical tree structure, so that the intent parsing process has a clear path dependence and semantic inheritance relationship. This dimension tree itself defines all possible and legal intent structures. The parsing task is redefined as finding the most matching semantic path in this tree for user input. This provides the rule engine with a precise target, i.e., a specific node or path, and provides a structured output target for the scenario model.

[0078] The innovative technical feature of this step is to propose an "intent matching based on pre-defined structured templates" method. It converts the intent recognition problem into a structured pattern matching problem by introducing a static, hierarchical semantic analysis dimension tree as an intermediate representation layer, thereby providing an irreplaceable basic framework for subsequent precise rule constraints and high-precision recall. This is a major breakthrough over the existing "dynamic problem decomposition" or "single label classification" paradigm.

[0079] S103: Formulate dimension value rules and associate semantic analysis dimension tree - configure dynamic behavior criteria for the core framework;

[0080] This step is another technical focus and innovation point of the present application, aiming to build a dynamic rule constraint engine that is deeply bound and works cooperatively with the "hierarchical semantic analysis dimension tree". The rules of this engine are not isolated, but take the nodes and paths of the semantic analysis dimension tree as the action objects, ensuring that the parsing process is logically correct and the output is stable within the framework. The dimension value rules include node activation rules and path constraint rules. Formulating dimension value rules includes the following three sub-steps:

[0081] S103.1: Formulation and binding of node activation rules;

[0082] Formulate fine-grained activation rules for key nodes in the semantic analysis dimension tree, especially leaf nodes, and directly bind them to the corresponding nodes to form node activation rules. Node activation rules include keyword matching rules, sentence pattern matching rules, and semantic similarity triggering rules; wherein:

[0083] Keyword matching rules: define the keywords, phrases, or synonyms that must appear in user input to activate the corresponding node; for example, the leaf node "sales" is bound to the keyword rule: {"must contain": ["sales", "revenue", "income"], "synonyms": ["flow", "money sold"]}.

[0084] Sentence pattern matching rule: define a set of text patterns for matching specific intent expressions; for example, bind the sentence pattern rule to the node "contrast report": { "pattern": ["… contrast …", "… compared to …", "… and … difference"]} ;

[0085] Semantic similarity trigger rule: for abstract concept nodes, define a trigger threshold based on vector similarity;

[0086] Technical effect: these rules serve as the "trigger" for the node, ensuring that the semantic fragments in the user input can be accurately mapped to the specific nodes in the tree, which is the starting point for path finding.

[0087] S103.2: Path constraint rule formulation and association;

[0088] Based on the topological relationship of the tree structure of the semantic analysis dimension tree, formulate the path constraint rule that controls the logic between paths, and associate it to the relevant path or node combination. At the same time, the following rules need to be followed when formulating the constraints.

[0089] Mutually exclusive path rule: define paths that are semantically conflicting and cannot be activated simultaneously. For example, bind the rule to the "data query" and "scheme recommendation" nodes, stating that the two are mutually exclusive;

[0090] Dependent path rule: define the activation of certain nodes / paths as a prerequisite for the activation of other nodes / paths. For example, create a rule for the "contrast benchmark dimension" node: IF "demand dimension = contrast report" THEN MUST "contrast benchmark dimension!= empty". This rule directly depends on the value of the "demand dimension" node;

[0091] Conflict resolution and priority rule: when user input may activate multiple conflicting paths, provide decision rules based on business weight. For example, "query sales" may simultaneously activate the "financial field" and "market field" nodes, and resolve them through predefined field priority rules.

[0092] S103.3: Rule-tree depth association storage and engine construction;

[0093] Store all the above rules in depth association with the node ID and path hash of the semantic analysis dimension tree, and construct a "rule-node-path" three-dimensional mapping network. This network constitutes the core knowledge base of the dynamic rule constraint engine. The uniqueness of this engine lies in that its execution logic is completely driven by the hierarchy and path relationship of the semantic analysis dimension tree. During the parsing process, the engine can quickly retrieve and apply all node-level and path-level rules associated with the currently activated node, achieving real-time verification and guidance.

[0094] In summary, the technical logic of this step of formulating dimension value rules and associating semantic analysis dimension tree lies in: through the rule system, the process of path exploration in the semantic analysis dimension tree is accurately controlled, ensuring that the output results meet both the language rules and the business logic. The coordination between the rule engine and the semantic analysis dimension tree is the key to realizing the logical consistency and output stability of the scheme.

[0095] The innovative technical feature of this step is the "dynamic rule constraint engine based on predefined structure". This engine is deeply integrated with the static hierarchical semantic analysis dimension tree, and all rules are targeted at the static nodes and paths of the tree. This "static structure + dynamic rule" coordination mechanism makes the entire parsing process have both the explicitness of the structured framework and the precision of the rule constraints, thereby realizing high-precision and high-stability intent parsing without the need for large-scale training. This is fundamentally different from methods that rely on the model's own ability for dynamic decomposition and reasoning.

[0096] S104: User input parsing and structured intent generation based on the semantic analysis dimension tree framework;

[0097] This step is a specific embodiment of the cooperation between the tree built by the core framework S102 and the rule engine built by S103. The parsing process is no longer a black box reasoning, but a path finding within a structured framework constrained by rules. The parsing process involves the synergistic effect of large language models and the dimension value rules.

[0098] Receiving user input, based on the structured framework of the semantic analysis dimension tree, and combining the dimension value rules to parse the user input and generate a structured semantic intent.

[0099] The specific method of step S104 includes:

[0100] S104.1: Semantic preprocessing and feature extraction of user input;

[0101] This step is to complete the regular cleaning, word segmentation, entity recognition of user input. It includes semantic cleaning and standardization, semantic entity recognition and linking, coreference resolution and semantic completion, semantic relation extraction;

[0102] S104.2: Collaborative parsing under the guidance of the framework;

[0103] Based on the pre-processed semantic features, collaborative analysis is performed in the semantic analysis dimension tree, the user input, complete semantic analysis dimension tree structure information, and the current session context are collectively organized into prompt information; and reasoning is guided within the structure space of the semantic analysis dimension tree to output preliminary path hypotheses and their confidence; during or after the path reasoning process, the dynamic rule constraint engine is called in real time to detect conflicts, verify dependencies, and calibrate confidence for the preliminary path hypotheses;

[0104] Specifically comprising the following steps:

[0105] 1. Context construction: the user input, complete "hierarchical semantic analysis dimension tree" structure information, and the current session context are collectively organized into prompt information.

[0106] 2. In-tree path reasoning: the analysis model is guided to reason within the given tree structure space. The task of the model is to output the possible activated node sequence (i.e. path hypothesis) and its confidence, rather than a simple label or vector;

[0107] 3. Real-time intervention of rule engine: the dynamic rule constraint engine is called immediately during the reasoning process or after obtaining the preliminary path hypothesis. The engine applies the rules bound in S103 to perform real-time verification according to the currently activated nodes, such as checking path conflicts, verifying dependency integrity, and feeding back the verification results of pass, conflict, and missing dependency to the analysis process.

[0108] 4. Iterative calibration: according to the feedback of the rule engine, the analysis process may reject some paths that violate the rules, supplement the necessary nodes that are missing, or recalibrate the confidence of multiple paths, forming an iterative optimization process of rules and models.

[0109] S104.3: Structured intent output: output one or more high-confidence semantic paths after rule verification and iterative optimization, as well as the specific values of each leaf node in the path, forming the structured intent object.

[0110] The final output is one or more high-confidence, complete semantic paths, and the specific values of each leaf node in the path.

[0111] The output is directly a structured intent object that can be seamlessly executed by a machine, for example: { "path": "intent / finance / comparison report / sales / smart TV / 2025", "value": { "field": "finance", "demand": "comparison report", "indicator": "sales", …}}.

[0112] S105: Precise task execution and knowledge recall based on structured intent;

[0113] According to the structured semantic intent-specific semantic path and node value generated in S104, the precise query condition of the database or the calling parameter of the specific task executor can be directly and accurately mapped. For example, the path and value are converted into the WHERE clause of the SQL query, or a specific analysis tool is called, thereby achieving precise and efficient recall of knowledge.

[0114] Perform the retrieval or task corresponding to the query condition or calling parameter, and fuse the matching result with the similarity retrieval result based on the semantic vector;

[0115] Sort the fused results according to the preset weight, and return the most relevant knowledge fragment or task execution result.

[0116] Specifically includes the following steps:

[0117] S105.1: Establish a mapping management of semantic nodes to specific tasks, including task semantic registration, semantic pattern matching, task priority scheduling, and task dependency analysis;

[0118] S105.2: Convert semantic intent into executable task instances, including parameter semantic mapping, parameter semantic verification, task instance generation, and execution context construction;

[0119] S105.3: Execute semantic tasks and monitor the execution process, including task semantic routing, execution process semantic tracking, semantic anomaly detection, and execution result semantic labeling;

[0120] S105.4: Integrate the execution results of multiple tasks and generate the final output, including result semantic fusion, result semantic enhancement, semantic consistency verification, and semantic result return.

[0121] This step is the embodiment of the final technical effect of the innovative framework, which can be briefly explained.

[0122] In summary, the above five steps and their detailed sub-steps constitute a complete and rigorous method for realizing precise intent recognition, and the core and key lies in steps S102 and S103. The "hierarchical semantic analysis dimension tree" constructed in step S102 provides a preset structured intermediate representation layer and target space for complex intent analysis, converting the problem into path finding. The dynamic rule constraint engine integrated with the tree depth constructed in step S103 provides real-time logical guarantee for ensuring correct path finding within the structure. The two constitute an inseparable collaborative whole: the tree provides the structure and target for the rules, and the rules guarantee the correct traversal and analysis of the tree. Step S104 is the specific embodiment of the collaborative body in the workflow, step S101 is the foundation, and step S105 is the application. This overall scheme systematically solves the fundamental defects in the background technology and realizes high-precision and high-stability intent recognition.

[0123] Example 2:

[0124] In a specific embodiment, the method of the present application is described in detail taking the "product sales query scenario of ChatBI (Conversational Business Intelligence) in the financial field" as an example. This embodiment aims to demonstrate the parameter configuration and specific application process of the method of the present application in a real business environment.

[0125] A. System architecture building;

[0126] Adopting micro-service architecture, the following core components are included:

[0127] Semantic analysis engine: carrying the core method of the present patent, responsible for semantic analysis dimension tree management and intent analysis;

[0128] Rule engine: responsible for the management, execution and verification of dimension value rules;

[0129] Task scheduling center: maintaining the "intent-task" mapping relationship, responsible for task distribution;

[0130] Business execution engine: including multiple specialized task executors;

[0131] Collaborative control module: responsible for coordinating the interaction and data flow between the semantic analysis engine, rule engine and scenario-based model, ensuring their collaborative work.

[0132] B. Detailed implementation steps;

[0133] Step 1: Define explicit multiple semantic analysis dimensions;

[0134] For the "financial ChatBI" scenario, by analyzing 15,832 historical user query logs, a complete system including 45 semantic analysis dimensions is constructed. The following are the detailed definitions of 8 core dimensions. For brevity, not all dimensions are listed:

[0135] 1. Domain dimension:

[0136] Function: distinguish the business domain attribution of the query.

[0137] Value set: {finance, human resources, marketing, technology R&D, supply chain, comprehensive}.

[0138] Priority weight: 0.9, high priority.

[0139] 2. Demand dimension:

[0140] Function: identify the core demand type of the user.

[0141] Value set: {Data query, trend analysis, comparison report, anomaly diagnosis, root cause analysis, solution recommendation, prediction request}.

[0142] Priority weight: 0.95, highest priority.

[0143] 3. Indicator Dimensions:

[0144] Function: Specify the specific business metrics that users care about.

[0145] The set of values ​​includes 28 financial indicators such as sales revenue, profit margin, ROE, gross profit margin, net profit margin, operating income, debt-to-equity ratio, and cash flow.

[0146] Priority weight: 0.93.

[0147] 4. Time dimension:

[0148] Function: Limits the time range for analysis.

[0149] Value set: {2022, 2023, 2024, 2025, Q1, Q2, Q3, Q4, this month, last week, today}.

[0150] Priority weight: 0.85.

[0151] 5. Product Dimension:

[0152] Function: Specifies the specific product to be searched.

[0153] Value set: {Smart TV series, air conditioner product line, refrigerator business unit, mobile phone products, IoT modules, chip business} and other 15 product lines.

[0154] Priority weight: 0.88.

[0155] 6. Regional Dimension:

[0156] Function: Defines the administrative or sales area where business takes place.

[0157] Value set: {North China, East China, South China, Southwest China, Northwest China, Northeast China, Overseas, Global}.

[0158] Priority weight: 0.75.

[0159] 7. Granularity dimension:

[0160] Function: Specifies the level of detail in the data summary.

[0161] Value set: {Total, Monthly, Weekly, Daily, Store-level, Channel-level}.

[0162] Priority weight: 0.70.

[0163] 8、Comparison reference dimension:

[0164] Function: In comparative analysis, specify the reference object of comparison.

[0165] Value set: {Last year's same period, Last month, Competitor A, Competitor B, Industry average, Budget target}.

[0166] Priority weight: 0.65.

[0167] Step 2: Build hierarchical semantic analysis dimension tree;

[0168] Based on the defined semantic analysis dimensions, a three-stage process is used to build the semantic analysis dimension tree. Below, we will construct the relevant semantic analysis dimension tree according to the dimensions of ChatBI built above.

[0169] 1、Determine the main stem of the tree structure;

[0170] First, build a basic tree stem that only contains the basic root node and each first-level dimension node. We use the yaml file format to build it, and the constructed yaml is as follows:

[0171] ```yaml;

[0172] Root node:

[0173] Node ID: "root";

[0174] Node name: "Intention root node";

[0175] Child nodes: ["domain_node"];

[0176] # First-level dimension;

[0177] - Node ID: "domain_node";

[0178] Node name: "Domain dimension";

[0179] Child nodes: ["finance_domain", "hr_domain", "market_domain"];

[0180] # Second-level dimension;

[0181] - Node ID: "finance_domain";

[0182] Node name: "Finance domain";

[0183] Child nodes: ["requirement_node"];

[0184] - Node ID: "requirement_node"

[0185] Node Name: "Requirement Dimension"

[0186] Sub-nodes: ["data_query", "trend_analysis", "comparison_report"]

[0187] 2. Establish core dimension relationships

[0188] Determine the main path of domain dimension → requirement dimension → indicator dimension, product dimension, time dimension, and space dimension as auxiliary branches, and comparison dimension as a special scenario branch. Complete the complete expansion of the financial domain. The following describes several typical dimensions in the financial domain.

[0189] The yaml for the financial domain expansion is as follows:

[0190] ```yaml

[0191] # Detailed definition of requirement dimension

[0192] - Node ID: "data_query"

[0193] Node Name: "Data Query"

[0194] Sub-nodes: ["metric_node", "product_node", "time_node", "spatial_node"]

[0195] Task Mapping: "financial_data_query_task"

[0196] - Node ID: "trend_analysis"

[0197] Node Name: "Trend Analysis"

[0198] Sub-nodes: ["metric_node", "time_node", "product_node"]

[0199] Task Mapping: "trend_analysis_task"

[0200] # Complete definition of indicator dimension

[0201] - Node ID: "metric_node"

[0202] Node Name: "Indicator Dimension"

[0203] Child nodes: ["sales_metric", "profit_metric", "roe_metric", "gross_margin_metric"];

[0204] - Node ID: "sales_metric";

[0205] Node name: "Sales Revenue";

[0206] Node type: "leaf node";

[0207] Task mapping: "sales_data_query_task";

[0208] Parameter mapping:

[0209] - Parameter name: "metric_type";

[0210] Parameter value: "sales_amount";

[0211] - Parameter name: "data_source";

[0212] Parameter value: "sales_system";

[0213] #Complete definition of product dimensions:

[0214] - Node ID: "product_node";

[0215] Node name: "Product Dimension";

[0216] Child nodes: ["tv_product", "ac_product", "refrigerator_product", "phone_product"];

[0217] - Node ID: "tv_product";

[0218] Node name: "Smart TV";

[0219] Node type: "leaf node";

[0220] Parameter mapping:

[0221] - Parameter name: "product_line";

[0222] Parameter value: "smart_tv";

[0223] - Parameter name: "business_unit";

[0224] Parameter name: "time_range"

[0225] # Time-space dimension complete definition:

[0226] - Node ID: "time_node"

[0227] Node name: "Time Dimension"

[0228] Child nodes: ["current_year", "last_year", "current_quarter", "current_month"]

[0229] - Node ID: "current_year"

[0230] Node name: "2025"

[0231] Node type: "Leaf Node"

[0232] Parameter mapping:

[0233] - Parameter name: "time_range"

[0234] Parameter value: "2025"

[0235] - Parameter name: "granularity"

[0236] Parameter value: "yearly"

[0237] …

[0238] ```

[0239] 3. Structure verification test

[0240] To verify the robustness of the previously constructed dimension structure, multiple structure use cases are constructed to test and adjust the constructed structure. The results of the structure use cases are as follows:

[0241] ```yaml

[0242] Verification use case:

[0243] - Test path: "root→finance_domain→data_query→sales_metric→tv_product→current_year"

[0244] Expected result: "Smart TV 2025 sales query"

[0245] Actual result: "Matched successfully"

[0246] Response Time: "23ms";

[0247] Test Path: "root→finance_domain→trend_analysis→profit_metric→ac_product";

[0248] Expected Result: "Air conditioner product profit rate trend analysis";

[0249] Actual Result: "Match success";

[0250] Response Time: "19ms";

[0251] The YAML structure constructed above is a specific example of the "hierarchical semantic analysis dimension tree" in the financial ChatBI scenario. Through the compact implementation of the two stages, the construction of the semantic analysis dimension tree is completed, providing a complete structured semantic framework for intent parsing.

[0252] Step three: Develop dimension value rules and associate semantic analysis dimension tree;

[0253] A complete rule management system is developed to accurately associate rules with semantic analysis dimension tree nodes. Thirty-five specific rules are developed, and some rule examples and associated semantic analysis dimension tree examples are explained below.

[0254] 1. Semantic matching rules;

[0255] 1.1. Semantic matching rule example:

[0256] Domain dimension - finance: {"must contain keyword semantics": ["income", "cost", "profit", "revenue", "financial report", "ROE", "sales", "gross profit rate"], "optional keywords": ["account", "tax", "fee", "rate"], "matching threshold": 1}. That is, at least one word in the "must contain keyword semantics" list must appear in the user input.

[0257] Time dimension - 2025: {"exact semantic match": ["2025", "2025 year"], "fuzzy semantic match": ["this year", "this year", "current fiscal year"], "contextual semantic inference": {"current year": 2025}}. The system will automatically map "this year" to "2025 year" based on the current server time.

[0258] Product dimension - A product: {"product alias mapping": {"A product": ["A product", "A series", "A type", "product A"], "B product": ["B product", "B series", "B type", "product B"]}}.

[0259] 1.2. Rules are directly associated to specific nodes of the semantic analysis dimension tree (in yaml format):

[0260] ```yaml;

[0261] # Leaf node semantic matching rules;

[0262] - Rule ID: "NODE_RULE_015";

[0263] Applicable node: "sales_metric" # Sales leaf node;

[0264] Rule type: "semantic trigger";

[0265] Rule content:

[0266] Must semantics: ["sales", "sales volume", "revenue", "income amount", "sales amount"];

[0267] Optional semantics: ["how much", "data", "value", "total"];

[0268] Minimum match number: 1;

[0269] Weight: 0.95;

[0270] - Rule ID: "NODE_RULE_023";

[0271] Applicable node: "current_year" # 2025 leaf node;

[0272] Rule type: "time semantic parsing";

[0273] Rule content:

[0274] Exact semantics: ["2025", "2025 year"];

[0275] Fuzzy semantics: ["this year", "this year", "current fiscal year"];

[0276] Dynamic inference:

[0277] Baseline time: "2025-01-01";

[0278] Semantic mapping: {"this year": "2025 year", "this year": "2025 year"};

[0279] - Rule ID: "NODE_RULE_045";

[0280] Applicable node: "tv_product" # Smart TV leaf node;

[0281] Rule Type: "Product Semantic Mapping";

[0282] Rule Content:

[0283] Product Alias:

[0284] Smart TV: ["TV", "Television", "Smart TV", "Smart TV", "Smart TV"];

[0285] Brand Association: ["Changhong TV", "CHiQ TV"];

[0286] 2. Semantic Consistency Verification Rules;

[0287] 2.1. Semantic Consistency Verification Rules Example:

[0288] Rule R001: IF indicator dimension IN ["Sales", "Profit Rate", "Revenue"] THEN demand dimension MUST IN ["Data Query", "Trend Analysis", "Comparison Report", "Abnormal Diagnosis"]. This rule prevents logical contradictions such as "Query Sales Recommendation Plan".

[0289] Rule R015: IF demand dimension = "Comparison Report" THEN comparison benchmark dimension MUST NOT EMPTY. Ensure that when performing comparison analysis, there must be a clear comparison object.

[0290] Rule R038: IF time dimension = "Today" THEN granularity dimension MUST IN ["Daily", "Store Level"]. When querying today's data, the granularity cannot be "Monthly" or "Weekly".

[0291] 2.2. Rules directly associated with specific nodes of semantic analysis dimension tree (YAML format):

[0292] ```yaml;

[0293] # Semantic Consistency Verification Rules;

[0294] - Rule ID: "PATH_RULE_001";

[0295] Rule Type: "Path Semantic Dependency";

[0296] Rule Content:

[0297] IF node path contains ["sales_metric"] # Sales node;

[0298] THEN must contain path ["data_query"] OR ["trend_analysis"] OR ["comparison_report"];

[0299] Error handling: "Sales figure metric must be used with data query, trend analysis or comparison report requirement";

[0300] - Rule ID: "PATH_RULE_015";

[0301] Rule type: "Path semantic integrity";

[0302] Rule content:

[0303] IF node path contains ["comparison_report"] # comparison report node;

[0304] THEN must contain any child node under path ["comparison_dimension"];

[0305] Error handling: "Comparison report must specify comparison baseline";

[0306] - Rule ID: "PATH_RULE_038";

[0307] Rule type: "Path semantic rationality";

[0308] Rule content:

[0309] IF node path contains ["current_year"] # 2025 year node;

[0310] AND node path contains ["data_query"] # data query node;

[0311] THEN granularity node must be ["yearly_granularity"] # yearly granularity;

[0312] Error handling: "Yearly data query should use yearly granularity";

[0313] In implementation, the rule engine is designed as an active component, not only checking after parsing, but also being called by the scenario model in real time through API during the parsing process, to judge the rule compliance of the intermediate result, realizing deep cooperation.

[0314] These rules all take the nodes and paths of the hierarchical semantic analysis dimension tree as the action objects, which are the "traffic rules" to ensure that the hierarchical semantic analysis dimension tree framework is correctly applied.

[0315] Step four: user input parsing and structured intent generation based on the semantic analysis dimension tree framework;

[0316] Take the user input "Help me check the sales of A product this year, and preferably compare it with last year" as an example to illustrate the analysis process:

[0317] 1. Input preprocessing: The system receives the input, performs text cleaning and word segmentation. Identify the key entities "A product" (mapped to the product dimension), "this year" (time dimension), "sales" (indicator dimension), and "last year" (comparison benchmark dimension).

[0318] 2. Collaborative parsing process:

[0319] a. The model receives the preprocessed text, dimension tree structure, and related rules as input.

[0320] b. The model performs multi-path reasoning and initially outputs the possible activated paths and their weights. This reasoning is conducted within the hierarchical semantic analysis dimension tree topology.

[0321] c. The rule engine is called in real-time to detect conflicts and calibrate weights for the initial output paths. For example, check if "data query" and "comparison report" conflict due to rule R001.

[0322] d. The model adjusts the output based on the feedback from the rule engine.

[0323] e. This process may iterate one or more times until the output result meets the rule constraints and has the highest confidence.

[0324] 3. Structured output parsing: The model returns the following JSON result:

[0325] {

[0326] "intent": "A product 2025 vs. 2024 sales comparison analysis",

[0327] "dimension values": {

[0328] "field": "finance",

[0329] "requirement": "comparison report", Note: Here the model correctly understands the "comparison" requirement under the hierarchical semantic analysis dimension tree framework and rule constraints, and finds the "comparison report" node and its associated "comparison benchmark" sub-node, rather than simply parsing it as "data query";

[0330] "indicator": "sales",

[0331] "product": "A product",

[0332] "Time":"2025",

[0333] "Comparison Baseline":"Last Year",

[0334] "Granularity":"Annual",

[0335] "Region":"Global"

[0336] },

[0337] "Weight":{

[0338] "Overall":0.96,

[0339] "Field":0.99,

[0340] "Requirement":0.94,

[0341] "Indicator":0.98,

[0342] "Product":0.97,

[0343] "Time":0.95,

[0344] "Comparison Baseline":0.93

[0345] };

[0346] };

[0347] 4. Post-processing of rules: The system calls the rule engine for verification. Both rules R001 and R015 are satisfied, there is no logical conflict, and the result is confirmed as valid.

[0348] This example clearly demonstrates the core role of the "Hierarchical Semantic Analysis Dimension Tree" framework: it accurately parses vague expressions like "Compare..." into a complete structured path through the tree structure ("Comparison Report" node and its child node "Comparison Baseline") and associated rules. This process embodies the technical advantages of parsing within a structured framework.

[0349] Step Five: Precise Task Execution and Knowledge Retrieval Based on Structured Intent;

[0350] 1. Knowledge Base Indexing: The knowledge base of this implementation has been pre-processed, with each record containing structured fields such as product line, time, indicator type, region, and a text content field. We have constructed a vector index for the text content and an Elasticsearch inverted index for the structured fields.

[0351] 2. Multi-strategy collaborative retrieval:

[0352] Structured Exact Matching: First, perform an exact query in Elasticsearch:

[0353] json

[0354] {

[0355] "query":{

[0356] "bool":{

[0357] "must":[

[0358] {"term":{"product line":"product A"}},

[0359] {"term":{"Indicator Type":"Sales Revenue"}},

[0360] {"term":{"year":2025}},

[0361] {"term":{"data granularity":"annual"}} ]

[0363] }

[0364] }

[0365] };

[0366] This step returns the annual summary data for Product A's sales in 2025.

[0367] Comparative data retrieval: Based on "Comparison benchmark: Same period last year", execute the query again, change the year to 2024, and return the sales data of product A in 2024.

[0368] Semantic Vector Retrieval: The intent description "Comparative analysis of sales revenue of Product A in 2025 and 2024" is converted into a vector, and a similarity search is performed in the Faiss index. As a supplement, some documents about sales revenue analysis methods are returned.

[0369] 3. Result fusion and sorting:

[0370] The sales data for Product A in 2025 and 2024 that are exactly matched are given the highest weight (1.0).

[0371] Other annual and quarterly data that are exactly matched have the next lowest weight (0.6).

[0372] The document with the lowest weight (0.3) returned by semantic search was the analysis method document.

[0373] The system sorts data according to weighted scores to ensure that the most relevant data is listed first.

[0374] Precise knowledge retrieval: The final retrieved knowledge fragments include:

[0375] Core data: {A product, 2025, sales, 120 million yuan, annual} and {A product, 2024, sales, 100 million yuan, annual};

[0376] Auxiliary information: sales growth calculation method, A product historical sales trend chart (2022-2025);

[0377] These structured data are passed to a large language model to generate the final user answer: "A product sales in 2025 is 120 million yuan, compared with 100 million yuan in 2024, an increase of 20%."

[0378] Finally, based on the high-precision and structured intent representation generated by the "hierarchical semantic analysis dimension tree" framework, the system realizes the accurate recall of knowledge. Test results show that in the financial ChatBI scenario of 1000+ intents, the implementation achieves an intent recognition accuracy of 96.2%, significantly higher than each of the comparative documents, which proves the effectiveness and advancement of the "hierarchical semantic analysis dimension tree" core framework proposed in this patent in solving the problem of complex intent structured analysis.

[0379] Although the present application has been described herein with reference to the explanatory embodiments thereof, the above-described embodiments are merely preferred embodiments of the present application, and the embodiments of the present application are not limited to the above-described embodiments. It should be understood by those skilled in the art that many other modifications and embodiments can be designed, which will fall within the scope and spirit of the principles disclosed in this application.

Claims

1. A method for achieving accurate intent recognition in AI question-answering applications, characterized in that, include: Step S101: For the target semantic application scenario, predefine a set of explicit and independent semantic analysis dimensions, wherein the semantic analysis dimensions include at least two of the following: domain dimension, requirement dimension, indicator dimension, time dimension, and object dimension; Step S102: Based on the semantic analysis dimensions, before system deployment, construct a static semantic analysis dimension tree with strict hierarchical relationships. The semantic analysis dimension tree includes a root node representing the most abstract user intent, intermediate nodes representing each semantic analysis dimension, and leaf nodes representing the specific values ​​of each dimension. This semantic analysis dimension tree defines the hierarchical relationships between dimensions and the semantic path from the root node to the leaf node. Step S103: Formulate dimension value rules and deeply associate the dimension value rules with the nodes and semantic paths in the semantic analysis dimension tree to build a dynamic rule constraint engine. The dimension value rules include node activation rules and path constraint rules. Step S104: Receive user input, parse the user input based on the structured framework of the semantic analysis dimension tree and in combination with the dimension value rules, and generate structured semantic intent; Step S105: Distribute and execute semantic tasks based on the generated structured semantic intent.

2. The method for achieving accurate intent recognition in AI question-answering applications according to claim 1, characterized in that, In step S101, the specific method for defining the semantic analysis dimension includes: S101.1: Conduct business research on the target application scenario and analyze historical user query records to uncover the semantic elements hidden in user queries; S101.2: Based on the semantic elements, construct an analysis system containing at least two independent semantic analysis dimensions, where each semantic analysis dimension represents a specific analytical perspective of the user's intent; S101.3: Configure attribute information for each defined semantic analysis dimension. The attribute information includes: dimension name, value set, value data type, default value strategy, and priority weight.

3. The method for achieving accurate intent recognition in AI question-answering applications according to claim 1, characterized in that, In step S102, the method for constructing the semantic analysis dimension tree includes: S102.1: Dimensional weighting and logical layering; S102.2: Construction of the tree structure of the semantic analysis dimension tree; S102.3: Path semantic definition and registration.

4. The method for achieving accurate intent recognition in AI question-answering applications according to claim 3, characterized in that, The specific method for the dimension weight ranking and logical layering is as follows: based on business domain knowledge, the semantic analysis dimensions are logically layered and weighted to determine the subordinate and dependent relationships between dimensions; And / or the tree structure construction of the semantic analysis dimension tree, specifically, is as follows: based on the result of the logical layering, an explicit tree structure consisting of root nodes, intermediate nodes, and leaf nodes is constructed; And / or the path semantic definition and registration, specifically, is as follows: for each legal path from the root node to the leaf node in the semantic analysis dimension tree, explicitly define its complete business semantics and register it as an identifiable intent pattern.

5. The method for achieving accurate intent recognition in AI question-answering applications according to claim 1, characterized in that, In step S103, the specific method for formulating the dimension value rules includes: S103.1: Node activation rule formulation and binding; S103.2: Path constraint rule formulation and association; S103.3: Rule-Tree Deeply Associative Storage and Engine Construction.

6. The method for achieving accurate intent recognition in AI question-answering applications according to claim 5, characterized in that, The node activation rule formulation and binding method is as follows: fine-grained activation rules are formulated for key nodes in the semantic analysis dimension tree and directly bound to the corresponding nodes to form node activation rules; the node activation rules include: keyword matching rules, sentence pattern matching rules and semantic similarity triggering rules; The path constraint rules are formulated and associated in the following way: based on the topological relationship of the semantic analysis dimension tree, path constraint rules are formulated between different paths in the semantic analysis dimension tree, and the path constraint rules are associated with the relevant paths or node combinations; at the same time, the following rules must be followed when specifying the path constraint rules: mutual exclusion path rules, dependent path rules, and conflict resolution and priority rules. The specific method for the rule-tree deep association storage and engine construction is as follows: the node activation rules and the path constraint rules are deeply associated with the node IDs and path hashes of the semantic analysis dimension tree to construct a three-dimensional mapping network of "rule-node-path"; this three-dimensional mapping network constitutes the knowledge base of the dynamic rule constraint engine.

7. The method for achieving accurate intent recognition in AI question-answering applications according to claim 6, characterized in that, The keyword matching rule is defined as follows: the keywords, phrases, or synonyms that must appear in the user input to activate the corresponding node are defined. And / or the sentence pattern matching rules are: a set of text patterns defined for matching specific intention expressions; And / or the semantic similarity triggering rule is: for abstract concept nodes, a triggering threshold based on vector similarity is defined.

8. The method for achieving accurate intent recognition in AI question-answering applications according to claim 6, characterized in that, The mutually exclusive path rule is defined as follows: semantically conflicting paths cannot be activated simultaneously. And / or the dependency path rule is defined as follows: the activation of one node / path must be contingent upon the activation of another node / path; And / or the conflict resolution and priority rules are as follows: when user input can activate multiple conflicting paths, a decision rule based on business weight is provided to resolve the conflict through a predefined domain priority rule.

9. The method for achieving accurate intent recognition in AI question-answering applications according to claim 1, characterized in that, In step S104, the specific method for parsing user input includes: S104.1: Perform semantic preprocessing and feature extraction on user input; S104.2: The preprocessed semantic features are collaboratively parsed in the semantic analysis dimension tree, and the user input, the complete semantic analysis dimension tree structure information, and the current session context are jointly organized into prompt information; and the prompt information is guided to perform reasoning within the structural space of the semantic analysis dimension tree, and the preliminary path hypothesis and its confidence level are output. During or after the path reasoning process, the dynamic rule constraint engine is invoked in real time to perform conflict detection, dependency verification, and confidence calibration on the initial path assumptions. S104.3: Output one or more high-confidence and complete semantic paths after rule verification and iterative optimization, as well as the specific values ​​of each leaf node in the semantic path, forming a structured semantic intent.

10. A method for achieving accurate intent recognition in AI question-answering applications according to any one of claims 1-9, characterized in that, In step S105, the specific method for semantic task distribution and execution is as follows: The generated structured semantic paths and node values ​​with clear semantic intent are mapped to precise query conditions in the database or call parameters of a specific task executor. Execute the retrieval or task corresponding to the query conditions or call parameters, and fuse the matching results with the similarity retrieval results based on semantic vectors; The merged results are sorted according to preset weights, and the most relevant knowledge fragments or task execution results are returned.

Images