Enterprise regulation intelligent question-answering system and method based on semantic understanding

Abstract

The invention relates to the technical field of intelligent question answering, and discloses an enterprise regulation intelligent question answering system and method based on semantic understanding, and the system comprises a data collection layer which is used for obtaining a multi-level complex regulation query statement input by a user; the semantic understanding layer is used for performing word segmentation, dependency syntax analysis and semantic role labeling on a query statement; and the knowledge reasoning layer is used for carrying out multi-step evidence fusion and answer verification through a logical reasoning engine. The method comprises the following steps: S1, acquiring a multi-level complex regulation query statement input by a user; s2, performing word segmentation, dependency syntax analysis and semantic role labeling on the query statement; and S3, analyzing the nested structure statement by using a hierarchical recurrent neural network. By introducing a multi-level semantic understanding mechanism and combining dependency syntactic analysis, semantic role labeling and a level recurrent neural network, deep structure analysis of the composite query statement is achieved, and the real intention which is not shown by a user can be accurately recognized.

Description

Technical Field

[0001] This invention relates to the field of intelligent question-answering technology, specifically to an intelligent question-answering system and method for enterprise regulations based on semantic understanding. Background Technology

[0002] In recent years, intelligent question-answering platforms based on natural language processing technology have been gradually applied to enterprise internal management, solving the problem of limited and non-semantic understanding-based retrieval methods for a large number of institutional documents. These platforms typically integrate internal enterprise institutional documents comprehensively, enabling efficient and intelligent application of these documents, and have broad application prospects in knowledge base construction, intelligent customer service, legal consultation, and other fields.

[0003] Existing systems suffer from a significant drop in semantic parsing accuracy when processing complex long sentences or multi-level regulatory queries. This is especially true when users pose complex questions containing multiple conditions or comparisons. Existing models based on keyword matching or shallow semantic analysis struggle to provide effective parsing. These methods typically only capture surface-level lexical features and fail to deeply deconstruct the logical structure of sentences or the user's unspoken true intent. Consequently, when faced with non-preset template queries in an enterprise environment, the system's accuracy in intent recognition and semantic parsing drops significantly, frequently returning irrelevant, fragmented, or even incorrect answers. This severely restricts the practicality and user experience of intelligent question-answering systems in real-world business scenarios. To address these issues, this invention designs an intelligent question-answering system and method for enterprise regulations based on semantic understanding. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent question-and-answer system and method for enterprise regulations based on semantic understanding, which solves the problem of inaccurate intent recognition leading to a decrease in the practicality of the question-and-answer system in the background art.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: A semantic understanding-based intelligent question-and-answer system for enterprise regulations includes: The data acquisition layer is used to acquire multi-level complex regulatory query statements input by users. These query statements contain multiple conditions, comparison relationships, and implicit intentions.

[0006] The semantic understanding layer is used to perform word segmentation, dependency parsing and semantic role labeling on the query statement, construct a semantic structure graph containing roles, conditions, comparison relations and implicit intentions, and use the semantic structure graph as inference input.

[0007] The knowledge reasoning layer generates multi-hop reasoning paths based on the semantic structure graph, the preset enterprise rules and regulations knowledge graph, and the multi-level semantic parsing model, and performs multi-step evidence fusion and answer verification through the logical reasoning engine.

[0008] The feedback optimization layer is used to dynamically optimize semantic parsing and inference strategies based on user feedback signals.

[0009] The output layer is used to output structured answers and interpretable reasoning paths based on a multi-dimensional confidence fusion mechanism that integrates semantic similarity, logical consistency, temporal rationality, and authority.

[0010] The system also includes a dynamic regulatory timeliness awareness module, which monitors regulatory revisions and automatically updates the knowledge graph and reasoning chain.

[0011] It also includes a semantic interpretability visualization interface, which is used to show users the reasoning path and semantic matching process.

[0012] Preferably, the semantic understanding layer, when constructing the semantic structure graph, further includes: Subject-verb-object triples are formed using dependency parsing, and word vector weights are calculated using a self-attention mechanism.

[0013] Based on the following formula: ; Adjusting word weights, where w i The term frequency is calculated based on the TF-IDF values ​​of the corporate regulations corpus, with a value ranging from 0 to 1. The smaller the value, the lower the term frequency. L represents the sentence length in characters, including spaces and punctuation, and is generally between 10 and 50 characters. α and β are adjustable parameters used to increase the importance of low-frequency keywords while balancing the importance of frequency and length to semantics. This formula increases the prominence of key information by increasing the weight of low-frequency words and penalizing overly long sentences.

[0014] A rule engine is introduced to identify comparison relationships, temporal and conditional logic, and a hierarchical recurrent neural network is used to perform multi-level parsing of nested structures.

[0015] Preferably, the knowledge reasoning layer, when generating multi-hop reasoning paths, further includes: A graph traversal algorithm is used to generate candidate inference chains and calculate path relevance scores, as shown in the following formula: ; Among them W i V represents the node weight, indicating the importance of a node in the inference path. Its value ranges from 0 to 1, with larger values ​​indicating greater criticality. i The node verification score represents the authenticity and accuracy of the node information, with a value ranging from 0 to 1. A higher value indicates greater credibility. m is the total number of nodes in the path.

[0016] A three-layer Tree-LSTM network is used to parse nested structures. The input is a dependency syntax tree node vector, and the hidden layer dimension is 128. Breadth-first search is used to traverse the knowledge graph to generate inference chains.

[0017] Limit the maximum search depth to avoid circular reasoning and use a state transition probability model to optimize path selection.

[0018] Preferably, the knowledge reasoning layer further includes the following during answer verification: The evidence fusion matrix is ​​used, as shown below: ; Measuring the correlation between different reasoning paths, where e i This represents the weight of the i-th piece of evidence in the reasoning path, with a value ranging from 0 to 1.

[0019] Based on conditional confidence c i Relevance score r of reasoning steps i Calculate the overall verification score: ; Where c i This represents the conditional confidence level at step i, ranging from 0 to 1. A higher value indicates a more credible premise; r i The value represents the relevance score of the reasoning step i, ranging from 0 to 1. The higher the value, the greater the impact of the step on the final conclusion; n is the total number of reasoning steps.

[0020] A threshold θ is set to control the answer acceptance standard, with a value ranging from 0.7 to 1, and is combined with the logical consistency formula: ; Perform deductive verification, where R represents the validity of the current reasoning step, and A and B are the correlation judgment conditions between the previous and subsequent steps.

[0021] Preferably, the feedback optimization layer further includes: Constructing a user intent evolution graph: ; Where V is the set of nodes, representing different intentions and semantic states; E is the set of edges, representing the transition relationships between states; and f is the state transition function, which determines how the system transitions from one intention to another.

[0022] Using the learning rate adjustment function: ; The self-learning speed of the dynamic control model is determined by the following: t is the current iteration number, τ is the decay period (usually set to several thousand iterations), and η0 is the initial learning rate (range 0.01-0.1).

[0023] Preferably, the output layer, when generating an answer, further includes: A sequence-to-sequence text generation model is used to generate natural language responses; Based on a multi-dimensional confidence fusion mechanism, the four dimensions of semantic similarity, logical consistency, temporal rationality, and regulatory authority are combined for weighted calculation to form the final confidence score; the top Z best answers are output, where Z is an integer, usually set to 3-5, along with an index of source regulatory clauses and a summary of reasoning path.

[0024] Preferably, the knowledge reasoning layer further includes the following during semantic encoding: Use the LSTM model to extract node embedding vectors; Construct a semantic similarity function: ; Measuring the consistency between nodes and regulatory provisions, where V node V represents the node embedding vector. rule The embedding vector of the rule matching item is represented; multi-node features are fused through the context attention module to generate a global semantic representation.

[0025] Preferably, when processing long texts, the system further includes: employing block processing technology to divide long paragraphs of regulatory text into multiple semantic blocks according to themes; constructing independent graph structures within each semantic block and merging them into a complete relational network through cross-block connections; and using dynamic windowing technology to adaptively adjust the sentence length, with the window length being: ; Where K is the base window length, ranging from 5 to 10; len(text) represents the length of the current sentence, in units of characters.

[0026] Preferably, the dynamic regulatory timeliness perception module further includes: real-time monitoring of update events in the regulatory database; automatic adjustment of node weights and reasoning path priorities in the knowledge graph; and marking and excluding expired clauses to ensure the timeliness of the answers.

[0027] The semantic interpretability visualization interface further includes: providing a graphical interface to display semantic relationship graphs and multi-hop reasoning paths; highlighting key semantic matching nodes and logical dependencies; and supporting interactive user queries and feedback input.

[0028] According to a second aspect of the present invention, a semantic understanding-based intelligent question-answering method for enterprise regulations as described above is also proposed, comprising the following steps: Step S1: Obtain the multi-level complex rule query statement input by the user.

[0029] Step S2 involves performing word segmentation, dependency parsing, and semantic role labeling on the query statement to construct a semantic structure graph that includes roles, conditions, comparison relationships, and implicit intentions.

[0030] Step S3: Use a rule engine to identify comparison relationships, tenses and conditional logic in the query, and use a hierarchical recurrent neural network to parse nested structure statements.

[0031] Step S4: Based on the enterprise rules and regulations knowledge graph and multi-level semantic parsing model, a multi-hop reasoning path is generated through a graph traversal algorithm, and the path relevance score is calculated.

[0032] Step S5 involves multi-step evidence fusion and answer verification through a logical reasoning engine, using the evidence fusion matrix and threshold to control answer adoption.

[0033] Step S6: Based on deep semantic parsing, perform fine-grained semantic encoding and intent recognition on the inference nodes.

[0034] Step S7: Based on the multi-dimensional confidence fusion mechanism of semantic similarity, logical consistency, temporal rationality and authority, output structured answers and references.

[0035] Step S8: Construct a user intent evolution graph and dynamically optimize semantic parsing and reasoning strategies based on user feedback.

[0036] Step S9: Automatically update the knowledge graph by combining the dynamic legal timeliness perception module.

[0037] Compared with the prior art, the beneficial effects achieved by the present invention are: 1. This invention addresses the problem that existing systems struggle to effectively parse complex long sentences, multiple conditions, and implicit logic. By introducing a multi-level semantic understanding mechanism, combined with dependency parsing, semantic role labeling, and hierarchical recurrent neural networks, this invention achieves deep structural parsing of compound query statements. It can accurately identify the user's unspoken true intent, significantly improving the accuracy and completeness of semantic understanding.

[0038] 2. In response to the common multi-condition and multi-level reasoning scenarios in corporate regulations, this invention constructs multi-hop reasoning paths based on knowledge graphs and introduces a logical reasoning engine and evidence fusion mechanism to achieve logical deduction and answer verification of multiple conditional relationships. This effectively avoids answer errors or fragmentation caused by broken logical chains or incomplete reasoning. By constructing a user intent evolution graph and a feedback optimization mechanism, the system can continuously optimize semantic parsing and reasoning strategies based on actual user feedback, gradually adapting to complex queries without preset templates and enhancing the system's practicality in dynamic corporate environments.

[0039] 3. In this invention, the system not only outputs structured answers, but also provides a visual display of the reasoning path and a highlighted explanation of the semantic matching process, enabling users to understand the logic behind the answer generation and enhancing the credibility of the system's output and user experience. By integrating a dynamic regulatory timeliness awareness module, the system can monitor regulatory changes in real time and automatically update the knowledge graph and reasoning path, ensuring that the answers are always based on the latest and most valid regulatory clauses. Attached Figure Description

[0040] Figure 1 This is a diagram of the overall system architecture of the present invention; Figure 2 This is a flowchart of the semantic understanding process of the present invention; Figure 3 This is a flowchart of the knowledge reasoning process of the present invention; Figure 4 This is the overall flowchart of the question-and-answer method of the present invention. Detailed Implementation

[0041] 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, and 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.

[0042] Example 1; Please see Figures 1-4 A semantic understanding-based intelligent question-and-answer system for enterprise regulations, comprising: The system comprises the following layers: a data acquisition layer, which acquires multi-level complex regulatory query statements input by the user, including multiple conditions, comparison relationships, and implicit intentions; a semantic understanding layer, which performs word segmentation, dependency parsing, and semantic role labeling on the query statements to construct a semantic structure graph containing roles, conditions, comparison relationships, and implicit intentions, and uses this semantic structure graph as inference input; a knowledge reasoning layer, which generates multi-hop reasoning paths based on the semantic structure graph, a pre-set enterprise regulatory knowledge graph, and a multi-level semantic parsing model, and performs multi-step evidence fusion and answer verification through a logical reasoning engine; a feedback optimization layer, which dynamically optimizes semantic parsing and reasoning strategies based on user feedback signals; and an output layer, which outputs structured answers and interpretable reasoning paths based on a multi-dimensional confidence fusion mechanism of semantic similarity, logical consistency, temporal rationality, and authority.

[0043] The system also includes a dynamic regulatory timeliness awareness module, which monitors regulatory revisions and automatically updates the knowledge graph and reasoning chain; and a semantic interpretability visualization interface, which displays the reasoning path and semantic matching process to users.

[0044] The semantic understanding layer, when constructing the semantic structure graph, further includes: Dependency parsing is used to form subject-verb-object triples, and word vector weights are calculated through a self-attention mechanism. Based on the following formula: ; Adjusting word weights, where w i The term frequency is calculated based on the TF-IDF values ​​of the corporate regulations corpus, with a value ranging from 0 to 1. The smaller the value, the lower the term frequency. L represents the sentence length in characters, including spaces and punctuation, and is generally between 10 and 50 characters. α and β are adjustable parameters used to increase the importance of low-frequency keywords while balancing the importance of frequency and length to semantics. This formula increases the prominence of key information by increasing the weight of low-frequency words and penalizing overly long sentences.

[0045] The system first acquires the user's query statement through a data acquisition layer. This statement may contain multiple conditions, comparative relationships, or implicit intentions. Then, the semantic understanding layer converts continuous text into lexical units through word segmentation, constructs a syntax tree using dependency parsing, and extracts subject-verb-object triples. Finally, semantic role labeling identifies the semantic function of each component. During this process, the system uses a self-attention mechanism to calculate word vector weights and adjusts these weights according to the aforementioned formula. For identified comparative relationships, tenses, and conditional logic, the system calls its built-in rule engine for specialized parsing. For complex sentences with nested structures, a hierarchical recurrent neural network is used for multi-level parsing, ultimately constructing a complete semantic structure graph that clearly represents the role relationships, conditional constraints, and implicit intentions in the query statement.

[0046] When generating multi-hop reasoning paths, the knowledge reasoning layer further includes: A graph traversal algorithm is used to generate candidate inference chains and calculate path relevance scores, as shown in the following formula: ; Among them W i V represents the node weight, indicating the importance of a node in the inference path. Its value ranges from 0 to 1, with larger values ​​indicating greater criticality. iA node verification score represents the authenticity and accuracy of the node information, ranging from 0 to 1, with higher values ​​indicating greater credibility. m represents the total number of nodes in the path. The maximum search depth is limited to avoid circular reasoning, and a state transition probability model is used to optimize path selection. The formula multiplies the weight and score of each node and sums the results to obtain the overall path relevance. This formula is designed to quantify the reliability of different paths, thereby selecting the optimal path.

[0047] The knowledge reasoning layer further includes the following during answer verification: The evidence fusion matrix is ​​used, as shown below: ; Measuring the correlation between different reasoning paths, where e i This represents the weight of the i-th piece of evidence in the reasoning path, with a value ranging from 0 to 1. This matrix measures the correlation between different reasoning paths to assess whether multiple reasoning clues are consistent. Specifically, each element e in the matrix... i This indicates the co-occurrence frequency of a certain premise or intermediate reasoning conclusion across multiple paths. By calculating the structural characteristics of this matrix, the system can identify the main logical routes and eliminate non-mainstream misleading reasoning.

[0048] Based on conditional confidence c i Relevance score r of reasoning steps i Calculate the overall verification score: ; Where c i This represents the conditional confidence level at step i, ranging from 0 to 1. A higher value indicates a more credible premise; r i The value represents the relevance score of the i-th reasoning step, ranging from 0 to 1. A higher value indicates a greater impact of that step on the final conclusion; n is the total number of reasoning steps. The overall verification score V is the average of these weighted sums to comprehensively evaluate the overall reliability of the reasoning process. This formula is designed to quantitatively assess the credibility of the results, ensuring that the final answer has sufficient logical support. For example, when verifying whether an employee can apply for a special subsidy, each step of the system's reasoning needs to meet certain confidence and relevance requirements to obtain a high V value.

[0049] A threshold θ is set to control the answer acceptance standard, with a value ranging from 0.7 to 1, and is combined with the logical consistency formula: ; Perform deductive verification, where R represents the validity of the current reasoning step, and A and B are the correlation judgment conditions between the previous and subsequent steps.

[0050] In the above embodiments, the knowledge reasoning layer, based on a pre-defined enterprise regulations knowledge graph and a multi-level semantic parsing model, uses a graph traversal algorithm to generate multi-hop reasoning paths and calculates the path relevance score R. p Where the node weight W i And verification score V i The system jointly evaluates the credibility of each path. It limits the maximum search depth to avoid circular reasoning and optimizes path selection using a state transition probability model. During the answer verification phase, the system uses an evidence fusion matrix to measure the relevance of different reasoning paths, calculates the overall verification score V, and applies a threshold θ and a logical consistency formula L. cons Perform deductive verification to ensure the logical rigor of the answer.

[0051] The feedback optimization layer further includes: Constructing a user intent evolution graph: ; In this diagram, V is the set of nodes, representing different intentions and semantic states; E is the set of edges, representing the transition relationships between states; and f is the state transition function, determining how the system transitions from one intention to another. In a corporate policy context, this diagram can represent the user's journey from inquiring about leave policies to understanding reimbursement standards, providing contextual support for subsequent questions.

[0052] Using the learning rate adjustment function: ; The dynamic control of the model's self-learning speed is defined by the formula: t is the current iteration number, τ is the decay period (typically set to several thousand iterations), and η0 is the initial learning rate, ranging from 0.01 to 0.1. This formula aims to control the model's rapid learning in the early stages, gradually slowing down the changes in later stages to avoid overfitting and improve stability and generalization ability.

[0053] The output layer, when generating the answer, further includes: A sequence-to-sequence text generation model is used to generate natural language responses.

[0054] Based on a multi-dimensional confidence fusion mechanism, the final confidence score is calculated by weighting four dimensions: semantic similarity, logical consistency, temporal rationality, and regulatory authority.

[0055] Output the top Z best answers, where Z is an integer, usually set to 3-5, along with an index of source regulations and a summary of the reasoning path.

[0056] The feedback optimization layer constructs a user intent evolution graph M, dynamically captures changes in user intent through a state transition function, and uses a learning rate adjustment function η. tThe model's self-learning speed is controlled to continuously optimize semantic parsing and inference strategies. The output layer uses a sequence-to-sequence text generation model to generate natural language answers, and outputs the top Z optimal answers and their inference path summaries based on a multi-dimensional confidence fusion mechanism of semantic similarity, logical consistency, temporal rationality, and authority.

[0057] The knowledge reasoning layer, during semantic encoding, further includes: Use an LSTM model to extract node embedding vectors.

[0058] Construct a semantic similarity function: ; Measuring the consistency between nodes and regulatory provisions, where V node V represents the node embedding vector. rule The embedding vector of the rule matching item is represented; multi-node features are fused through the context attention module to generate a global semantic representation.

[0059] When processing long texts, the system further includes: using block processing technology to divide long paragraphs of regulations into multiple semantic blocks according to themes; constructing an independent graph structure within each semantic block, and merging them into a complete relational network through cross-block connections.

[0060] Use dynamic windowing technology to adaptively adjust sentence length; window length: ; Where K is the base window length, ranging from 5 to 10; len(text) represents the length of the current sentence in characters; and W... max This represents the maximum window length. The formula adaptively adjusts the window size to ensure that long sentences are segmented appropriately without affecting semantic coherence.

[0061] The dynamic regulation timeliness perception module further includes: real-time monitoring of update events in the regulation database; automatic adjustment of node weights and reasoning path priorities in the knowledge graph; and marking and excluding expired clauses to ensure the timeliness of answers.

[0062] The semantic interpretability visualization interface further includes: providing a graphical interface to display semantic relationship graphs and multi-hop reasoning paths; highlighting key semantic matching nodes and logical dependencies; and supporting interactive user queries and feedback input.

[0063] The working principle of this invention is as follows: The dynamic regulatory timeliness perception module achieves intelligent maintenance of regulatory timeliness based on continuous monitoring and adaptive update mechanisms. This module monitors the regulatory database update events in real time through a preset regulatory database interface, either by polling at fixed time intervals or through an event-driven approach. These events include changes such as adding regulations, revising clauses, and repealing declarations. When a regulatory update event is detected, the module automatically triggers the knowledge graph update process: First, semantic parsing is performed on the newly added or revised regulatory clauses to extract key entities, attributes, and relationships, generating new knowledge nodes. Then, the weights of the affected nodes are recalculated using a node importance evaluation algorithm, which comprehensively considers the node's centrality in the inference network, its citation frequency, and timeliness. Simultaneously, the module adjusts the priority of inference paths related to the updated clauses and reorders the available inference chains based on the path timeliness coefficient and confidence score.

[0064] For expired clauses, expired nodes are first marked as "historical versions" in the knowledge graph to maintain their accessibility but reduce their weight in the reasoning process; then, the timeliness filter automatically excludes the marked expired content in the answer generation stage to ensure that the output answer strictly complies with the current effective laws and regulations.

[0065] The semantic interpretability visualization interface is based on graphics rendering and interactive response mechanisms to visualize the internal reasoning process of the system. The interface builds a multi-layered visualization framework through a front-end graphics engine: it displays the semantic relationship graph in the base layer, displays the multi-hop reasoning path in the reasoning layer, and clearly presents the reasoning chain from query to conclusion using a temporal flow layout.

[0066] During the visualization process, the interface uses a key node identification algorithm to determine the content that needs to be highlighted. Based on the structural importance, semantic matching degree, and confidence score of nodes in the inference path, it automatically identifies key semantically matching nodes. Through a logical dependency analysis algorithm, it identifies and highlights clusters of nodes with strong logical connections. The highlighting adopts a multi-dimensional visual encoding scheme, including color saturation to represent confidence level, border thickness to represent node importance, and animation effects to represent logical flow.

[0067] Users can trigger a detailed information display panel through clicks, hovers, and other interactive actions, presenting the complete attributes, source terms, and confidence analysis of nodes. An embedded feedback collection component receives user confirmations, corrections, or supplementary information regarding the inference path. These feedback signals are transmitted to the system's feedback optimization layer via a standardized interface, forming a complete optimization loop. Simultaneously, the interface provides an inference path export function, supporting the output of the complete semantic analysis process in a structured format for easy subsequent review and archiving.

[0068] Example 2; Please see Figures 1-4In this embodiment of the invention, an intelligent question-and-answer method for enterprise regulations based on semantic understanding includes the following steps: Step S1: Obtain the user's multi-level complex regulatory query statement; Step S2: Perform word segmentation, dependency parsing, and semantic role labeling on the query statement to construct a semantic structure graph containing roles, conditions, comparison relationships, and implicit intentions; Step S3: Use a rule engine to identify comparison relationships, tenses, and conditional logic in the query, and use a hierarchical recurrent neural network to parse nested structured statements; Step S4: Based on the enterprise regulatory knowledge graph and multi-level semantic parsing model, generate multi-hop reasoning paths through a graph traversal algorithm and calculate path relevance scores; Step S5: Perform multi-step evidence fusion and answer verification through a logical reasoning engine, and use the evidence fusion matrix and threshold to control answer adoption; Step S6: Based on deep semantic parsing, perform fine-grained semantic encoding and intent recognition on reasoning nodes; Step S7: Based on a multi-dimensional confidence fusion mechanism of semantic similarity, logical consistency, time reasonableness, and authority, output structured answers and references; Step S8: Construct a user intent evolution graph and dynamically optimize semantic parsing and reasoning strategies based on user feedback; Step S9: Automatically update the knowledge graph in conjunction with a dynamic regulatory timeliness perception module.

[0069] The working principle of this invention embodiment is as follows: In step S1, the system obtains the multi-level complex rule query statement input by the user through the data acquisition interface. This process includes receiving the input text, encoding conversion and preprocessing, to provide a standardized data foundation for subsequent semantic analysis.

[0070] Step S2 performs a three-level language analysis on the query statement. First, word segmentation is used to divide the continuous text into a sequence of words, establishing basic language units. Then, dependency parsing is performed to construct a grammatical dependency tree and extract core syntactic components such as subject, verb, and object. Finally, semantic role labeling is performed to identify and label the semantic roles of each component, including agent, patient, time, and place. Through these three levels of continuous analysis, the system constructs a complete semantic structure graph, clearly representing the role relationships, conditional constraints, comparative relationships, and implicit intentions in the query statement.

[0071] Step S3 specifically handles the complex logical structure in the query. The system calls the preset rule engine to identify the comparison relationship, temporal information and conditional logic in the query based on regular expressions and logical templates. For complex statements with nested structures, a hierarchical recurrent neural network is used for layered parsing. Semantic dependencies are deconstructed layer by layer in a bottom-up manner to ensure accurate capture and representation of complex semantic relationships.

[0072] Step S4 initiates the knowledge graph-based reasoning process. Based on a pre-defined enterprise regulations knowledge graph, the system employs a graph traversal algorithm combining breadth-first search and depth-first search to generate multiple candidate reasoning paths within the knowledge network. These paths are then scored using the path relevance score formula R. p Each path is quantitatively evaluated, taking into account the importance of nodes (W). i And verify credibility V i Two dimensions are used to filter out the most relevant reasoning paths.

[0073] Step S5 implements deep answer verification. The system performs multi-step evidence fusion through a logical reasoning engine, constructs an evidence fusion matrix to evaluate the consistency and complementarity between different evidence sources, calculates the overall verification score V based on conditional confidence, and controls the answer adoption standard through a preset threshold θ to ensure the reliability and accuracy of the output answer.

[0074] Step S6 performs fine-grained semantic analysis. Based on deep semantic parsing, the embedding vectors of inference nodes are extracted using an LSTM network, and multi-node features are fused through a contextual attention mechanism. On this basis, fine-grained intent recognition is performed to accurately capture the user's true query intent.

[0075] Step S7 generates the final output. Based on a multi-dimensional confidence fusion mechanism, the system comprehensively calculates the weighted scores of four dimensions: semantic similarity, logical consistency, temporal rationality, and regulatory authority, to form the final confidence assessment. The output includes a structured answer and corresponding references, including index information of relevant regulations and a summary of the reasoning process.

[0076] Step S8 achieves continuous system optimization by constructing a user intent evolution graph M=(V,E,f) to model the state transition process of user intent; and dynamically adjusting the semantic parsing and inference strategies based on user feedback signals, using the learning rate adjustment function η. t Control the model update speed to achieve continuous improvement in system performance.

[0077] Step S9 ensures the timeliness and accuracy of the answer by combining the dynamic legal timeliness perception module to monitor the update events of the legal database in real time; automatically identifying newly added, revised or repealed regulations and updating the node content and relationships in the knowledge graph accordingly; and ensuring that the output answer strictly complies with the requirements of currently effective regulations through the timeliness verification mechanism.

[0078] Example 3; Please see Figures 1-4 This paper presents a specific application scenario of an intelligent question-and-answer system and method for enterprise regulations based on semantic understanding. The following example, a typical enterprise employee query, is used to explain in detail the system's workflow, parameter settings, and output results.

[0079] Users enter the query "If an employee takes more than 10 days of leave during the probationary period, can the company terminate the employment contract?" through the system interface. The data acquisition layer receives the query and performs preprocessing, including encoding conversion and text cleaning, to ensure that the input data is standardized.

[0080] The semantic understanding layer first segments the query statement into word sequences, including "employee," "probation period," "leave," "exceeding," "10 days," "company," "termination," and "labor contract." Then, a syntactic dependency tree is constructed through dependency parsing, extracting subject-verb-object triples: "company" as the subject, "termination" as the verb, and "labor contract" as the object. Semantic role labeling further identifies the semantic roles of each component: "employee" as the agent, "probation period" as the time condition, and "leave exceeding 10 days" as the conditional constraint. When constructing the semantic structure graph, the system uses a self-attention mechanism to calculate word vector weights and adjusts these weights according to formula S. The word frequency w... i The values ​​are based on a corporate regulations corpus, with the word frequency of "leave request" being 0.2 and "relevant" being 0.3; the sentence length L is 20 characters; and the adjustable parameters α and β are set to 0.5 and 0.3 respectively to increase the importance of low-frequency keywords.

[0081] The knowledge reasoning layer, based on a pre-defined enterprise regulations knowledge graph, uses a graph traversal algorithm to generate multi-hop reasoning paths. Path 1 is associated with the regulation clause "Probationary Period Management Regulations," Path 2 with "Leave Policy," and Path 3 with "Conditions for Termination of Employment Contract." The system calculates the path relevance score R. p Where the node weight W i The node's importance in the inference path is indicated by values ​​of 0.8, 0.9, and 0.7; the node verification score V... i The values ​​represent the authenticity and accuracy of the node information, taking values ​​of 0.9, 0.8, and 0.9 respectively; the total number of path nodes, m, is 3. The calculated R... p =(0.8×0.9)+(0.9×0.8)+(0.7×0.9)=2.07. The system limits the maximum search depth to 5 to avoid circular reasoning and uses a state transition probability model to optimize path selection, prioritizing paths with higher scores.

[0082] During the answer verification phase, the system uses an evidence fusion matrix E to measure the correlation between different reasoning paths, with evidence weights e. i The values ​​are 0.9, 0.8, and 0.85. Based on the conditional confidence level c... i Relevance score r of reasoning steps i Calculate the overall verification score V, where c takes values ​​of 0.9, 0.8, and 0.85, and r... iThe values ​​are 0.9, 0.8, and 0.9, with n being 3. The calculated value is V = (0.9 × 0.9 + 0.8 × 0.8 + 0.85 × 0.9) / 3 = 2.215 / 3 ≈ 0.738. The system sets the threshold θ to 0.7. Since V ≥ θ, the answer is accepted. Furthermore, considering the logical consistency formula L... cons Deductive verification is performed, where R represents the validity of the current reasoning step, and A and B are the judgment conditions for the preceding and following steps, ensuring logical rigor.

[0083] The feedback optimization layer constructs a user intent evolution graph M, where the node set V represents different intent states, the edge set E represents state transition relationships, and the state transition function f is dynamically adjusted based on user feedback. The system uses a learning rate adjustment function η. t To control the self-learning speed, with the initial learning rate η0 set to 0.05, the current iteration number t set to 100, and the decay period τ set to 1000, the following calculations are obtained: η t =0.05×(1-100 / 1000) 2 =0.0405, enabling continuous optimization of semantic parsing and reasoning strategies.

[0084] The output layer uses a sequence-to-sequence text generation model to generate natural language answers. Based on a multi-dimensional confidence fusion mechanism, it performs a weighted calculation based on four dimensions: semantic similarity, logical consistency, temporal rationality, and regulatory authority, resulting in a final confidence score of 0.85. The system outputs the top Z best answers, where Z is set to 3, including the answer text, the index of the source regulatory clauses, and a summary of the reasoning path.

[0085] The dynamic regulation timeliness awareness module monitors the regulation database for updates in real time. Upon detecting a revision to the "Probationary Period Management Regulations," it automatically adjusts the weights of relevant nodes and the priority of reasoning paths in the knowledge graph, marking and excluding outdated clauses to ensure the timeliness of answers. The semantic interpretability visualization interface displays the semantic relationship graph and multi-hop reasoning paths through a graphical interface, highlighting key semantic matching nodes, emphasizing the logical dependency between "leave exceeding 10 days" and "termination of employment contract," and supporting interactive user queries and feedback input.

[0086] Working Principle: The intelligent question-answering system for enterprise regulations operates on a layered, collaborative intelligent processing architecture. First, the data acquisition layer acquires multi-level, complex user-input queries. Then, the semantic understanding layer performs word segmentation, dependency parsing, and semantic role labeling to construct a semantic structure graph containing roles, conditions, comparison relationships, and implicit intentions. This layer uses a self-attention mechanism and weight adjustment formula to optimize word vector representations and leverages a rule engine and hierarchical recurrent neural network to handle complex logic and nested structures. The knowledge reasoning layer, based on the enterprise regulations knowledge graph, uses a graph traversal algorithm to generate multi-hop reasoning paths and performs multi-step verification through path relevance scoring and evidence fusion matrices. It then uses a logical consistency formula and preset thresholds to filter reliable answers. The feedback optimization layer constructs an intent evolution graph based on user feedback and dynamically optimizes semantic parsing and reasoning strategies through a learning rate adjustment function. Finally, the output layer generates natural language answers based on a multi-dimensional confidence fusion mechanism, along with the reasoning path and clause index. In addition, the system also uses a dynamic regulatory timeliness perception module and a semantic interpretability visualization interface to realize real-time updates of the knowledge graph and visualization of the reasoning process, thereby improving the system's timeliness, accuracy and user understandability.

[0087] The intelligent question answering method operates according to a multi-step semantic understanding and reasoning process. It begins by acquiring a complex regulatory query input by the user (S1), followed by word segmentation, syntactic analysis, and semantic role labeling to construct a semantic structure graph (S2). Next, the system identifies comparative relationships, tenses, and conditional logic within the query and uses a hierarchical recurrent neural network to parse nested structures (S3). In the knowledge reasoning stage, multi-hop reasoning paths are generated based on the enterprise regulatory knowledge graph, and their relevance is evaluated (S4). Then, a logical reasoning engine performs multi-step evidence fusion and answer verification (S5). The system further performs fine-grained semantic encoding and intent recognition on the reasoning nodes (S6), and outputs structured answers and references based on a multi-dimensional confidence fusion mechanism (S7). To continuously optimize system performance, the method also constructs a user intent evolution graph, dynamically adjusting parsing and reasoning strategies based on user feedback (S8). Finally, combined with a dynamic regulatory timeliness awareness module, the knowledge graph is automatically updated to ensure the timeliness and accuracy of the answers (S9). The entire process achieves end-to-end intelligent processing from semantic parsing to credible answer generation.

[0088] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A semantic understanding-based intelligent question-and-answer system for enterprise regulations, characterized in that, include: The data acquisition layer is used to acquire multi-level complex regulatory query statements input by users, which contain multiple conditions, comparison relationships and implicit intentions; The semantic understanding layer is used to perform word segmentation, dependency parsing and semantic role labeling on the query statement, construct a semantic structure graph containing roles, conditions, comparison relations and implicit intentions, and use the semantic structure graph as inference input; The knowledge reasoning layer generates multi-hop reasoning paths based on the semantic structure graph, the preset enterprise rules and regulations knowledge graph, and the multi-level semantic parsing model, and performs multi-step evidence fusion and answer verification through the logical reasoning engine. The feedback optimization layer is used to dynamically optimize semantic parsing and inference strategies based on user feedback signals. The output layer is used to output structured answers and interpretable reasoning paths based on a multi-dimensional confidence fusion mechanism that integrates semantic similarity, logical consistency, temporal rationality, and authority. The system also includes a dynamic regulatory timeliness awareness module, which monitors regulatory revisions and automatically updates the knowledge graph and reasoning chain; It also includes a semantic interpretability visualization interface, which is used to show users the reasoning path and semantic matching process.

2. The intelligent question-and-answer system for enterprise regulations based on semantic understanding according to claim 1, characterized in that, The semantic understanding layer, when constructing the semantic structure graph, further includes: Dependency parsing is used to form subject-verb-object triples, and word vector weights are calculated through a self-attention mechanism. Based on the following formula: ； Adjust the word weight S, where w i The value represents the word frequency, ranging from 0 to 1; L represents the sentence length, in units of characters; α and β are adjustable parameters used to increase the importance of low-frequency keywords. A rule engine is introduced to identify comparison relationships, temporal and conditional logic, and a hierarchical recurrent neural network is used to perform multi-level parsing of nested structures.

3. The intelligent question-and-answer system for enterprise regulations based on semantic understanding according to claim 1, characterized in that: The knowledge reasoning layer, when generating multi-hop reasoning paths, further includes: A graph traversal algorithm is used to generate candidate inference chains, and the path relevance score R is calculated. p The formula is as follows: ； Among them W i V represents the node weight, indicating the importance of a node in the inference path, with a value ranging from 0 to 1. i The node verification score represents the authenticity and accuracy of the node information, with a value ranging from 0 to 1; m is the total number of nodes in the path. Limit the maximum search depth to avoid circular reasoning and use a state transition probability model to optimize path selection.

4. The intelligent question-and-answer system for enterprise regulations based on semantic understanding according to claim 1, characterized in that, The knowledge reasoning layer further includes the following during answer verification: The evidence fusion matrix is ​​used, as shown below: ； Measuring the correlation between different reasoning paths, where e i This represents the weight of the i-th piece of evidence in the reasoning path, with a value ranging from 0 to 1; Based on conditional confidence c i Relevance score r of reasoning steps i Calculate the overall verification score V: ； Where c i r represents the conditional confidence level at step i, with a value ranging from 0 to 1; i represents the relevance score of the reasoning step i, with a value ranging from 0 to 1; n is the total number of reasoning steps. A threshold θ is set to control the answer acceptance standard, with a value ranging from 0.7 to 1, and is combined with the logical consistency formula: ； Perform deductive verification, where R represents the validity of the current reasoning step, and A and B are the correlation judgment conditions between the previous and subsequent steps.

5. The intelligent question-and-answer system for enterprise regulations based on semantic understanding according to claim 1, characterized in that, The feedback optimization layer further includes: Constructing the user intent evolution graph M: ； Where V is the set of nodes, representing different intentions and semantic states; f is the set of edges, representing the transition relationships between states; and f is the state transition function, which determines how the system transitions from one intention to another. Using the learning rate adjustment function η t : ； The self-learning speed of the dynamic control model is determined by the following: t is the current iteration number, τ is the decay period (usually set to several thousand iterations), and η0 is the initial learning rate (range 0.01-0.1).

6. The intelligent question-and-answer system for enterprise regulations based on semantic understanding according to claim 1, characterized in that, The output layer, when generating the answer, further includes: A sequence-to-sequence text generation model is used to generate natural language responses; Based on a multi-dimensional confidence fusion mechanism, the four dimensions of semantic similarity, logical consistency, temporal rationality, and regulatory authority are combined for weighted calculation to form the final confidence score. Output the top Z best answers, where Z is an integer, usually set to 3-5, along with an index of source regulations and a summary of the reasoning path.

7. The intelligent question-and-answer system for enterprise regulations based on semantic understanding according to claim 1, characterized in that, The knowledge reasoning layer, during semantic encoding, further includes: Use the LSTM model to extract node embedding vectors; Construct a semantic similarity function S match : ； Measuring the consistency between nodes and regulatory provisions, where V node V represents the node embedding vector. rule This represents the embedding vector of the rule matching item; A global semantic representation is generated by fusing features from multiple nodes through a contextual attention module.

8. The intelligent question-and-answer system for enterprise regulations based on semantic understanding according to claim 1, characterized in that, When processing long texts, the system further includes: Block processing technology is used to divide long paragraphs of regulations into multiple semantic blocks according to their themes; An independent graph structure is constructed within each semantic block, and then merged into a complete relational network through cross-block connections. The sentence length is adaptively adjusted using dynamic windowing technology. Window length W: ； Where K is the base window length, ranging from 5 to 10; len(text) represents the length of the current sentence, in characters; W max This represents the maximum window length.

9. The intelligent question-and-answer system for enterprise regulations based on semantic understanding according to claim 1, characterized in that, The dynamic regulatory timeliness perception module further includes: Real-time monitoring of updates to the regulatory database; Automatically adjust node weights and reasoning path priorities in the knowledge graph; Mark and exclude expired terms to ensure the timeliness of answers; The semantic interpretability visualization interface further includes: Provides a graphical interface to display semantic relationship graphs and multi-hop reasoning paths; Highlight key semantic matching nodes and logical dependencies; It supports interactive user queries and feedback input.

10. A semantic understanding-based intelligent question-answering method for enterprise regulations, implemented using the semantic understanding-based intelligent question-answering system for enterprise regulations as described in any one of claims 1-9, characterized in that, Includes the following steps: Step S1: Obtain the user's input of a multi-level complex rule query statement; Step S2: Perform word segmentation, dependency parsing and semantic role labeling on the query statement to construct a semantic structure graph containing roles, conditions, comparison relations and implicit intentions; Step S3: Use a rule engine to identify comparison relationships, tenses and conditional logic in the query, and use a hierarchical recurrent neural network to parse nested structure statements; Step S4: Based on the enterprise rules and regulations knowledge graph and multi-level semantic parsing model, a multi-hop reasoning path is generated through a graph traversal algorithm, and the path relevance score is calculated. Step S5 involves multi-step evidence fusion and answer verification through a logical reasoning engine, using the evidence fusion matrix and threshold to control answer adoption; Step S6: Based on deep semantic parsing, perform fine-grained semantic encoding and intent recognition on the inference nodes; Step S7: Based on the multi-dimensional confidence fusion mechanism of semantic similarity, logical consistency, temporal rationality and authority, output the structured answer and reference basis; Step S8: Construct a user intent evolution graph and dynamically optimize semantic parsing and reasoning strategies based on user feedback; Step S9: Automatically update the knowledge graph by combining the dynamic legal timeliness perception module.

Images