Intelligent question answering method and system based on tax knowledge graph

Abstract

The invention relates to the technical field of intelligent question answering, in particular to an intelligent question answering method and system based on a tax knowledge graph, and the system comprises a knowledge graph construction module, a user question processing module, an answer reasoning generation module and a system iteration module. According to the method, an intention recognition engine comprising a contradictory recognition unit and an anomaly recognition unit is constructed, contradiction detection is realized by utilizing an entity conflict measurement model, a tense interval algebraic contradiction model and the like, anomaly analysis is completed in combination with a chaos theory-Lyapunov index fusion model, and the recognition rate of contradictory problems and anomaly problems is improved; according to the method, a reasoning channel selection mechanism is designed, single-layer retrieval, multi-path reasoning and manual switching are dynamically switched according to contradictory and anomaly analysis results, and a knowledge feedback mechanism of a system iteration module is matched, so that the manual intervention cost of complex question answering is reduced, and the answer accuracy is improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent question answering technology, specifically to an intelligent question answering method and system based on a tax knowledge graph. Background Technology

[0002] Intelligent question-answering systems are a typical application of natural language processing, based on massive amounts of data from the internet and deep semantic understanding technology. They are now widely used in various fields closely related to people's lives, such as technology, education, shopping, daily life, and healthcare, greatly improving the efficiency of information acquisition. In modern society, the status of tax law is increasingly important, and tax authorities, as one of the national functional departments, have a significant responsibility to popularize tax laws, regulations, and related policies. Therefore, helping taxpayers quickly and conveniently understand tax knowledge is particularly important for tax authorities. Intelligent question-answering systems, developed and applied based on various technologies such as natural language understanding, machine learning, and information extraction, have achieved good results. However, current intelligent question-answering systems based on tax knowledge graphs often suffer from problems when processing user queries. These include insufficient recognition of metaphorical or colloquial expressions in the queries, or queries containing contradictory or complex information, leading to failed reasoning and insufficient explanatory power in the answers, resulting in a poor user experience. Summary of the Invention

[0003] This invention provides an intelligent question-answering method and system based on tax knowledge graphs to solve the aforementioned technical problems.

[0004] The first aspect of this invention provides an intelligent question-answering system based on a tax knowledge graph, including a knowledge graph construction module, a user question processing module, an answer reasoning and generation module, and a system iteration module.

[0005] The knowledge graph construction module is used to send data collection instructions and data fusion instructions. When the pre-deployed data crawling unit detects the data collection instruction, it performs data collection and execution. The data crawling unit obtains tax-related data through data collection and execution and sends the tax-related data to the pre-deployed data fusion unit. When the data fusion unit detects the data fusion instruction, it updates the tax-related data into the knowledge graph.

[0006] The data collection process includes, but is not limited to, crawling publicly available data sources such as the State Taxation Administration's official website and legal databases through the data crawling unit to obtain policy documents, cases, and frequently asked questions and answers regarding tax rates and filing deadlines.

[0007] The user question processing module collects user input information through a built-in multimodal input unit, transmits the user input information to a pre-deployed intent recognition engine, obtains the user's question intent information through the intent recognition engine, and sends the user's question intent information to the answer reasoning and generation module.

[0008] As a further improvement of the present invention, the intent recognition engine specifically includes a contradiction recognition unit and an anomaly recognition unit. It obtains a contradiction detection data chain and an anomaly detection data chain based on the user's input information, and sends the contradiction detection data chain to the contradiction recognition unit for analysis to obtain contradiction information; and sends the anomaly detection data chain to the anomaly recognition unit for analysis to obtain anomaly information; and combines the contradiction information and the anomaly information to obtain the user's question intent information.

[0009] As a further improvement to the present invention, the contradiction identification unit is further analyzed in the following ways:

[0010] The conflict detection data chain is identified to obtain entity information, rule information, and time verification information; the number of entities is obtained based on the entity information in the user's query, and the graph embedding vector corresponding to each entity is obtained. An entity conflict measurement model is then used. The entity conflict probability ST is calculated; where n represents the number of entities, x i Let x represent the graph embedding vector corresponding to the i-th entity. rel Represented as the embedding vector of a standard entity in a knowledge graph; sin(x) i ,x rel ) represents the cosine similarity between the entity information and the standard entity; rel(x) i ,x rel ) represents the preset entity relationship weight value, and α represents the preset entity importance weight value; when the entity conflict probability is greater than the preset conflict probability threshold, the part of the entity conflict probability that is greater than the conflict probability threshold is marked as the entity conflict value st.

[0011] The number of rules matching the user's question is obtained based on the rule information. The priority coefficient of each rule is obtained based on the knowledge graph, and fuzzy membership functions are used. Obtain the rule matching degree pd; substitute the number of rules, priority coefficient, and rule matching degree into the rule conflict model. The rule conflict value GZ is calculated; where m represents the number of rules, β j Let β be the priority coefficient corresponding to the j-th rule; divide the rule conflict values ​​into multiple conflict value intervals, and give each conflict value interval a corresponding conflict coefficient β. GZ The conflict coefficient is obtained by matching the rule conflict value with each conflict value interval, and then using GZ×β. GZ =gz calculates the rule contradiction value gz.

[0012] Based on the time verification information, obtain the issue time information and the corresponding policy validity period information for the user's issue; based on the issue time information, obtain the issue start time and issue end time, and record them as follows: And obtain the problem time interval. Similarly, the effective date and expiration date corresponding to the policy validity period information are obtained and recorded as follows: And obtain the policy time frame Using the time overlap function The time overlap lap(t) was calculated. w ,t z Using a temporal interval algebraic contradiction model The time discrepancy value T is calculated; where length(t) represents the length of the time interval, calculated in calendar days.

[0013] The three sub-conflict values ​​corresponding to the user problem—entity conflict value, rule conflict value, and time conflict value—are normalized, and their values ​​are then used to determine the comprehensive conflict point model. The calculated contradiction comprehensive value CH k Where k represents the entity contradiction value, rule contradiction value, and time contradiction value, namely st, gz, and T; ω k Let be the weight factor corresponding to each sub-contradiction value; σ(x) represents the sigmoid activation function, mapping the contradiction values ​​to [0,1]; obtain the pre-set minimum value of contradiction synthesis and the maximum value of contradiction synthesis, i.e., CH. k-min and CH k-max Compare the total value of contradictions with the minimum total value of contradictions and the maximum total value of neutralized contradictions.

[0014] When CH k <CH k-min When CH, the corresponding contradictory information is contradictory-free information; when CH k-min ≤CH k <CH k-max When CH, the corresponding contradictory information is contradictory; k ≥CH k-max At that time, the corresponding contradictory information is a contradiction warning signal.

[0015] As a further improvement to the present invention, the anomaly identification unit specifically analyzes the following:

[0016] The anomaly detection data chain is used to identify user behavior patterns, semantic features, and question structure information. Based on the user behavior patterns, the frequency of user questions and the frequency of topic jumps are obtained. When the frequency of user questions exceeds a preset question frequency threshold, the portion of the question frequency exceeding the threshold is marked as a question anomaly. Similarly, the jump frequency threshold corresponding to the topic jump frequency and the corresponding jump anomaly are obtained. The question anomaly and the jump anomaly are added together to obtain the behavior anomaly value.

[0017] Based on semantic features, the semantics of each word in the user question are obtained. Based on the semantics of each word, the user question words are divided into tax words, sensitive words, and sentiment words. The sensitivity coefficient and sentiment coefficient corresponding to each sensitive word and sentiment word are obtained. The number of sensitive words and the number of sentiment words corresponding to the user question are counted. The sensitivity outlier is calculated by multiplying the number of sensitive words with the corresponding sensitivity coefficient. The sentiment outlier is calculated by multiplying the number of sentiment words with the corresponding sentiment coefficient. The semantic outlier is obtained by summing the sensitivity outlier and the sentiment outlier corresponding to the user question.

[0018] Based on the question structure information, the number of statements with fuzzy information entropy and the total number of statements are obtained. The number of statements with fuzzy information entropy is counted to obtain the number of fuzzy statements. The ratio of the number of fuzzy statements to the total number of statements is calculated to obtain the fuzzy statement ratio. A pre-set fuzzy statement threshold is obtained. When the fuzzy statement ratio is greater than the fuzzy statement threshold, the corresponding fuzzy statement ratio is marked as a structural anomaly value. Fuzzy information entropy is the detected structural anomalies in user questions, including but not limited to disordered word order and missing information.

[0019] The three sub-outliers—behavioral outliers, semantic outliers, and structural outliers—are normalized and their values ​​are taken, based on the chaos theory-Lyapunov exponential fusion model. The comprehensive outlier value YC is calculated; where y represents the sub-outlier index, and y = 1, 2, 3 correspond to the sub-outliers of behavioral outliers, semantic outliers, and structural outliers, respectively. and Let exp(x) represent the Lyapunov exponent of the time series of the y-th sub-outlier and the mean Lyapunov exponent of the y-th sub-outlier under normal conditions, respectively; exp(x) represents the natural exponential function; φ y This is represented by the preset weight factor corresponding to the y-th sub-outlier; when the comprehensive outlier is greater than the preset threshold, an anomaly information is generated indicating an anomaly; conversely, when the comprehensive outlier is less than the preset threshold, an anomaly information is generated indicating no anomaly; and when the comprehensive outlier exceeds the preset comprehensive outlier upper limit, an anomaly warning signal is generated.

[0020] The answer reasoning generation module includes a reasoning channel selection unit, a single-layer retrieval channel, and a multi-path reasoning channel. The reasoning channel selection unit receives user question intent information and obtains contradictory and abnormal information based on it. It then performs channel selection analysis based on this contradictory and abnormal information to obtain channel signals and pushes corresponding answers based on the channel signals for each user question. The single-layer retrieval channel directly queries the knowledge graph to return standard answers for simple questions, including but not limited to VAT rates. The path reasoning channel generates a set of candidate answers based on legal chains, case matching, and policy deduction, based on the knowledge graph extraction results.

[0021] As a further improvement of the present invention, channel selection analysis is performed based on contradictory information and abnormal information. Specifically, when contradictory information and abnormal information correspond to non-contradictory information and non-abnormal information, respectively, the corresponding channel signal is generated as a single-layer channel signal.

[0022] When contradictory and abnormal information appear, or when either contradictory or abnormal information appears simultaneously, a corresponding channel signal is generated as a multi-channel signal.

[0023] When contradictory or abnormal information presents either a contradictory warning signal or an abnormal warning signal, or both signals are present simultaneously, a manual processing signal is generated.

[0024] The single-layer retrieval channel performs answer reasoning for user questions whose channel signals are single-layer channel signals; the multi-path reasoning channel performs answer reasoning for user questions whose channel signals are multi-path channel signals.

[0025] The system iteration module acquires the answer content data package for each user's question and uploads it to the knowledge graph for updates and iterations, and stores the answer content data package and the knowledge graph.

[0026] A second aspect of this invention provides an intelligent question-answering method based on a tax knowledge graph, comprising the following steps:

[0027] Step A: Monitor the question-and-answer process in real time. When a signal for manual intervention is detected, generate a manual transfer instruction containing a summary of the question.

[0028] Step B: Push the manual transfer instruction and pre-processed entity recognition, reasoning attempts, and other information to the human agent, who will then take over to supplement and answer complex or unusual questions.

[0029] Step C: Collect manually answered data, label the correct reasoning paths and policy basis, and generate a knowledge graph to feed back into the data package.

[0030] The beneficial effects of the technical solution provided by this invention compared with the prior art are as follows:

[0031] 1. This invention constructs an intent recognition engine containing contradiction recognition units and anomaly recognition units, utilizes entity conflict measurement models and temporal interval algebraic contradiction models to achieve contradiction detection, and combines chaos theory-Lyapunov exponent fusion models to complete anomaly analysis, thereby improving the recognition rate of contradictory and anomaly problems.

[0032] 2. This invention designs a reasoning channel selection mechanism that dynamically switches between single-layer retrieval, multi-path reasoning, and manual transfer based on the results of contradiction and anomaly analysis. Combined with the knowledge feedback mechanism of the system iteration module, this reduces the cost of manual intervention in solving complex problems and improves the accuracy of answers. Attached Figure Description

[0033] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. The following drawings are not deliberately drawn to scale according to the actual size, but are intended to show the main idea of ​​this application.

[0034] Figure 1 This is a schematic diagram of the principle of an intelligent question-answering system based on a tax knowledge graph according to the present invention.

[0035] Figure 2 This is a flowchart of an intelligent question-answering method based on a tax knowledge graph according to the present invention. Detailed Implementation

[0036] 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.

[0037] For ease of understanding, the specific process of the embodiments of the present invention is described below. Please refer to [link / reference]. Figure 1-2 In one embodiment of the present invention, an intelligent question-answering system based on a tax knowledge graph includes:

[0038] The knowledge graph construction module sends data collection and data fusion commands. When the pre-deployed data crawling unit detects the data collection command, it executes the collection. The data crawling unit obtains tax-related data through the collection and execution and sends the tax-related data to the pre-deployed data fusion unit. When the data fusion unit detects the data fusion command, it updates the tax-related data into the knowledge graph. For example, when the 2025 VAT rate adjustment policy is released, the data crawling unit automatically crawls the policy document, parses the "tax rate adjustment" entity and the associated "tax calculation formula" relationship, and updates the knowledge graph synchronously to ensure the timeliness of the knowledge graph and avoid incorrect answers due to policy lag.

[0039] The data collection process includes, but is not limited to, crawling publicly available data sources such as the State Taxation Administration's official website and legal databases through the data crawling unit to obtain policy documents, cases, and frequently asked questions and answers regarding tax rates and filing deadlines.

[0040] The user question processing module collects user input information through a built-in multimodal input unit, transmits the user input information to a pre-deployed intent recognition engine, obtains the user's question intent information through the intent recognition engine, and sends the user's question intent information to the answer reasoning and generation module.

[0041] The intent recognition engine specifically includes a contradiction recognition unit and an anomaly recognition unit. Based on the user's input information, it obtains a contradiction detection data chain and an anomaly detection data chain, and sends the contradiction detection data chain to the contradiction recognition unit for analysis to obtain contradiction information; and sends the anomaly detection data chain to the anomaly recognition unit for analysis to obtain anomaly information; and combines the contradiction information and the anomaly information to obtain the user's question intent information.

[0042] Furthermore, the contradiction identification unit is analyzed in the following ways:

[0043] The conflict detection data chain is identified to obtain entity information, rule information, and time verification information; the number of entities is obtained based on the entity information in the user's query, and the graph embedding vector corresponding to each entity is obtained. An entity conflict measurement model is then used. The entity conflict probability ST is calculated; where n represents the number of entities, x i Let x represent the graph embedding vector corresponding to the i-th entity. rel Represented as the embedding vector of a standard entity in a knowledge graph; sin(x) i ,x rel ) represents the cosine similarity between the entity information and the standard entity; rel(x) i ,x rel) represents the preset entity relationship weight value, and α represents the preset entity importance weight value; when the entity conflict probability is greater than the preset conflict probability threshold, the part of the entity conflict probability that is greater than the conflict probability threshold is marked as the entity conflict value st.

[0044] The number of rules matching the user's question is obtained based on the rule information. The priority coefficient of each rule is obtained based on the knowledge graph, and fuzzy membership functions are used. Obtain the rule matching degree pd; substitute the number of rules, priority coefficient, and rule matching degree into the rule conflict model. The rule conflict value GZ is calculated; where m represents the number of rules, β j Let β be the priority coefficient corresponding to the j-th rule; divide the rule conflict values ​​into multiple conflict value intervals, and give each conflict value interval a corresponding conflict coefficient β. GZ The conflict coefficient is obtained by matching the rule conflict value with each conflict value interval, and then using GZ×β. GZ =gz calculates the rule contradiction value gz.

[0045] Based on the time verification information, obtain the issue time information and the corresponding policy validity period information for the user's issue; based on the issue time information, obtain the issue start time and issue end time, and record them as follows: And obtain the problem time interval. Similarly, the effective date and expiration date corresponding to the policy validity period information are obtained and recorded as follows: And obtain the policy time frame Using the time overlap function The time overlap lap(t) was calculated. w ,t z Using a temporal interval algebraic contradiction model The time discrepancy value T is calculated; where length(t) represents the length of the time interval, calculated in calendar days.

[0046] The three sub-conflict values ​​corresponding to the user problem—entity conflict value, rule conflict value, and time conflict value—are normalized, and their values ​​are then used to determine the comprehensive conflict point model. The calculated contradiction comprehensive value CH k Where k represents the entity contradiction value, rule contradiction value, and time contradiction value, namely st, gz, and T; ω k Let be the weight factor corresponding to each sub-contradiction value; σ(x) represents the sigmoid activation function, mapping the contradiction values ​​to [0,1]; obtain the pre-set minimum value of contradiction synthesis and the maximum value of contradiction synthesis, i.e., CH. k-min and CH k-maxCompare the total value of contradictions with the minimum total value of contradictions and the maximum total value of neutralized contradictions.

[0047] When CH k <CH k-min When the contradictory information is correct, the corresponding contradictory information is correct.

[0048] When CH k-min ≤CH k <CH k-max At that time, the corresponding contradictory information is contradictory;

[0049] When CH k ≥CH k-max At that time, the corresponding contradictory information is a contradiction warning signal.

[0050] Furthermore, the anomaly identification unit analyzes the following:

[0051] The anomaly detection data chain is used to identify user behavior patterns, semantic features, and question structure information. Based on the user behavior patterns, the frequency of user questions and the frequency of topic jumps are obtained. When the frequency of user questions exceeds a preset question frequency threshold, the portion of the question frequency exceeding the threshold is marked as a question anomaly. Similarly, the jump frequency threshold corresponding to the topic jump frequency and the corresponding jump anomaly are obtained. The question anomaly and the jump anomaly are added together to obtain the behavior anomaly value.

[0052] Based on semantic features, the semantics of each word in the user question are obtained. Based on the semantics of each word, the user question words are divided into tax words, sensitive words, and sentiment words. The sensitivity coefficient and sentiment coefficient corresponding to each sensitive word and sentiment word are obtained. The number of sensitive words and the number of sentiment words corresponding to the user question are counted. The sensitivity outlier is calculated by multiplying the number of sensitive words with the corresponding sensitivity coefficient. The sentiment outlier is calculated by multiplying the number of sentiment words with the corresponding sentiment coefficient. The semantic outlier is obtained by summing the sensitivity outlier and the sentiment outlier corresponding to the user question.

[0053] Based on the question structure information, the number of statements with fuzzy information entropy and the total number of statements are obtained. The number of statements with fuzzy information entropy is counted to obtain the number of fuzzy statements. The ratio of the number of fuzzy statements to the total number of statements is calculated to obtain the fuzzy statement ratio. A pre-set fuzzy statement threshold is obtained. When the fuzzy statement ratio is greater than the fuzzy statement threshold, the corresponding fuzzy statement ratio is marked as a structural anomaly value. Fuzzy information entropy is the detected structural anomalies in user questions, including but not limited to disordered word order and missing information.

[0054] The three sub-outliers—behavioral outliers, semantic outliers, and structural outliers—are normalized and their values ​​are taken, based on the chaos theory-Lyapunov exponential fusion model. The comprehensive outlier value YC is calculated; where y represents the sub-outlier index, and y = 1, 2, 3 correspond to the sub-outliers of behavioral outliers, semantic outliers, and structural outliers, respectively. and Let exp(x) represent the Lyapunov exponent of the time series of the y-th sub-outlier and the mean Lyapunov exponent of the y-th sub-outlier under normal conditions, respectively; exp(x) represents the natural exponential function; φ y This is represented by the preset weight factor corresponding to the y-th sub-outlier; when the comprehensive outlier is greater than the preset threshold, an anomaly information is generated indicating an anomaly; conversely, when the comprehensive outlier is less than the preset threshold, an anomaly information is generated indicating no anomaly; and when the comprehensive outlier exceeds the preset comprehensive outlier upper limit, an anomaly warning signal is generated.

[0055] The answer reasoning generation module includes a reasoning channel selection unit, a single-layer retrieval channel, and a multi-path reasoning channel. The reasoning channel selection unit receives user question intent information and obtains contradictory and abnormal information based on this information. It then performs channel selection analysis based on the contradictory and abnormal information to obtain channel signals and pushes corresponding answers based on the channel signals for each user's question. The single-layer retrieval channel directly queries the knowledge graph to return the standard answer for simple questions, including but not limited to VAT rates. The path reasoning channel generates a set of candidate answers based on legal chains, case matching, and policy deduction, based on the knowledge graph extraction results. The reasoning channel selection unit switches channels based on contradictory and abnormal information: the single-layer retrieval channel is used when there are no contradictions or abnormalities, while the multi-path reasoning channel is used when there are contradictions or abnormalities. When a warning signal is triggered, a manual processing signal is generated to achieve hierarchical processing of questions and optimize reasoning efficiency.

[0056] Channel selection analysis is performed based on contradictory and anomalous information. Specifically, when contradictory and anomalous information correspond to contradictory and anomalous information, respectively, the corresponding channel signal is generated as a single-layer channel signal.

[0057] When contradictory and abnormal information appear, or when either contradictory or abnormal information appears simultaneously, a corresponding channel signal is generated as a multi-channel signal.

[0058] When contradictory or abnormal information presents either a contradictory warning signal or an abnormal warning signal, or both signals are present simultaneously, a manual processing signal is generated.

[0059] The single-layer retrieval channel performs answer reasoning for user questions whose channel signals are single-layer channel signals; the multi-path reasoning channel performs answer reasoning for user questions whose channel signals are multi-path channel signals.

[0060] The system iteration module obtains the answer content data package for each user's question and uploads it to the knowledge graph for updating and iteration, and stores the answer content data package and the knowledge graph.

[0061] This invention also provides an intelligent question-answering method based on a tax knowledge graph, comprising the following steps:

[0062] Step A, Monitoring of Manual Handling Signals: The question-and-answer process is monitored in real time. When a manual handling signal is detected, a manual transfer instruction containing a question summary is generated to accurately locate complex questions that require human intervention.

[0063] Step B, Manual Transfer Execution: The manual transfer instruction and pre-processed entity recognition, reasoning attempts, and other information are pushed to the human agent. The human agent takes over to supplement and answer complex or abnormal questions, improving the completeness of the answers to complex questions.

[0064] Step C, Human Feedback Processing: Collect human-answered data, label the correct reasoning paths and policy basis, and generate a knowledge graph to feed back into the data package, transforming artificial intelligence into system knowledge.

[0065] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An intelligent question-answering system based on a tax knowledge graph, characterized in that, include: The user question processing module collects user input information through a built-in multimodal input unit, transmits the user input information to a pre-deployed intent recognition engine, obtains the user's question intent information through the intent recognition engine, and sends the user's question intent information to the answer reasoning and generation module. The intent recognition engine specifically includes a contradiction recognition unit and an anomaly recognition unit. It obtains a contradiction detection data chain and an anomaly detection data chain based on the user's input information, and sends the contradiction detection data chain to the contradiction recognition unit for analysis to obtain contradiction information. And send the anomaly detection data chain to the anomaly identification unit for analysis to obtain anomaly information; By combining contradictory and abnormal information, we can obtain the user's problem intent information; The answer reasoning generation module includes a reasoning channel selection unit, a single-layer retrieval channel, and a multi-path reasoning channel; The reasoning channel selection unit receives user question intention information and obtains contradictory and abnormal information based on the user question intention information. It then performs channel selection analysis based on the contradictory and abnormal information to obtain channel signals and pushes corresponding signals according to the channel signals of each user question.

2. The intelligent question-answering system based on tax knowledge graph according to claim 1, characterized in that, The specific analysis content of the contradiction identification unit is as follows: The contradiction detection data chain is identified to obtain entity information, rule information, and time verification information; the number of entities is obtained based on the entity information in the user's question, the graph embedding vector corresponding to each entity is obtained, and the entity contradiction probability ST is calculated through the entity contradiction measurement model; when the entity contradiction probability is greater than the preset contradiction probability threshold, the part of the entity contradiction probability that is greater than the contradiction probability threshold is marked as the entity contradiction value st. The number of rules matched to the user's question is obtained based on the rule information. The priority coefficient of each rule is obtained based on the knowledge graph. The rule matching degree is obtained using the fuzzy membership function. The rule conflict value GZ is calculated by substituting the number of rules, priority coefficient, and rule matching degree into the rule conflict model. The rule conflict value is then divided into multiple conflict value intervals, and a corresponding conflict coefficient β is given to each conflict value interval. GZ The conflict coefficient is obtained by matching the rule conflict value with each conflict value interval, and then using GZ×β. GZ =gz calculates the rule contradiction value gz; Based on the time verification information, obtain the issue time information and the corresponding policy validity period information for the user's issue; based on the issue time information, obtain the issue start time and issue end time, and record them as follows: And obtain the problem time interval. Similarly, the effective date and expiration date corresponding to the policy validity period information are obtained and recorded as follows: And obtain the policy time frame Using the time overlap function The time overlap lap(t) was calculated. w ,t z Using a temporal interval algebraic contradiction model The time discrepancy value T is calculated; where length(t) represents the length of the time interval. The three sub-conflict values ​​corresponding to the user problem—entity conflict value, rule conflict value, and time conflict value—are normalized, and their values ​​are then used to determine the comprehensive conflict point model. The calculated contradiction comprehensive value CH k Where k represents the entity contradiction value, rule contradiction value, and time contradiction value, namely st, gz, and T; ω k Let be the weight factor corresponding to each sub-contradiction value; σ(x) represents the sigmoid activation function, mapping the contradiction values ​​to [0,1]; obtain the pre-set minimum value of contradiction synthesis and the maximum value of contradiction synthesis, i.e., CH. k-min and CH k-max Compare the total value of contradictions with the minimum total value of contradictions and the maximum total value of neutralized contradictions. When CH k <CH k-min When the contradictory information is correct, the corresponding contradictory information is correct. When CH k-min ≤CH k <CH k-max At that time, the corresponding contradictory information is contradictory; When CH k ≥CH k-max At that time, the corresponding contradictory information is a contradiction warning signal.

3. The intelligent question-answering system based on tax knowledge graph according to claim 1, characterized in that, The anomaly identification unit analyzes the following: The anomaly detection data chain is identified to obtain user behavior pattern information, semantic features, and question structure information; based on the user behavior pattern information, the user question frequency and topic jump frequency are obtained; when the user question frequency is greater than the preset question frequency threshold, the part of the question frequency exceeding the question frequency threshold is marked as a question anomaly value; Similarly, obtain the jump frequency threshold corresponding to the topic jump frequency, and the corresponding jump anomaly value; sum the question anomaly value and the jump anomaly value to obtain the behavior anomaly value; Based on semantic features, the semantics of each word in the user question are obtained. Based on the semantics of each word, the user question words are divided into tax words, sensitive words, and sentiment words. The sensitivity coefficient and sentiment coefficient corresponding to each sensitive word and sentiment word are obtained. The number of sensitive words and sentiment words corresponding to user questions are statistically obtained. The number of sensitive words is multiplied by the corresponding sensitivity coefficient to obtain the sensitive outlier value, and the number of sentiment words is multiplied by the corresponding sentiment coefficient to obtain the sentiment outlier value. The semantic anomaly value is obtained by summing the sensitive outliers and sentiment outliers corresponding to user questions; Based on the problem structure information, obtain the statements with fuzzy information entropy and the total number of statements. Count the number of statements with fuzzy information entropy to obtain the number of fuzzy statements. Calculate the ratio of the number of fuzzy statements to the total number of statements to obtain the fuzzy statement ratio. Obtain a pre-set fuzzy statement threshold. When the fuzzy statement ratio is greater than the fuzzy statement threshold, mark the corresponding fuzzy statement ratio as a structural outlier. The three sub-outliers—behavioral outliers, semantic outliers, and structural outliers—are normalized and their values ​​are taken, based on the chaos theory-Lyapunov exponential fusion model. The comprehensive outlier value YC is calculated; where y represents the sub-outlier index, and y = 1, 2, 3 correspond to the sub-outliers of behavioral outliers, semantic outliers, and structural outliers, respectively. and Let exp(x) represent the Lyapunov exponent of the time series of the y-th sub-outlier and the mean Lyapunov exponent of the y-th sub-outlier under normal conditions, respectively; exp(x) represents the natural exponential function; φ y This is represented as the preset weight factor corresponding to the y-th sub-outlier; When the total outlier value is greater than the preset threshold, an anomaly information is generated indicating that the outlier is present; conversely, when the total outlier value is less than the preset threshold, an anomaly information is generated indicating that the outlier is not present. Furthermore, when the overall abnormal value exceeds the preset overall abnormal upper limit, an abnormal warning signal is generated.

4. The intelligent question-answering system based on tax knowledge graph according to claim 1, characterized in that, The channel selection analysis based on contradictory and abnormal information specifically includes: When contradictory information and abnormal information correspond to contradictory information and abnormal information respectively, the corresponding channel signal generated is a single-layer channel signal; When contradictory and abnormal information appear, or when either contradictory or abnormal information appears simultaneously, a corresponding channel signal is generated as a multi-channel signal. When contradictory or abnormal information presents either a contradictory warning signal or an abnormal warning signal, or both signals are present simultaneously, a manual processing signal is generated. The single-layer retrieval channel performs answer reasoning for user questions whose channel signals are single-layer channel signals; the multi-path reasoning channel performs answer reasoning for user questions whose channel signals are multi-path channel signals.

5. The intelligent question-answering system based on tax knowledge graph according to claim 1, characterized in that, It also includes a knowledge graph construction module, which is used to send data collection instructions and data fusion instructions. When the pre-deployed data crawling unit detects the data collection instruction, it performs data collection and execution. The data crawling unit obtains tax-related data through data collection and execution and sends the tax-related data to the pre-deployed data fusion unit. When the data fusion unit detects the data fusion instruction, it updates the tax-related data to the knowledge graph.

6. The intelligent question-answering system based on tax knowledge graph according to claim 1, characterized in that, It also includes a system iteration module, which acquires the answer content data package of each user's question and uploads it to the knowledge graph for updating and iteration, and stores the answer content data package and the knowledge graph.

7. An intelligent question-answering method based on a tax knowledge graph, characterized in that, An intelligent question-answering system based on a tax knowledge graph, as described in any one of claims 1-6, comprises the following steps: Step A: Monitor the question-and-answer process in real time. When a signal for manual intervention is detected, generate a manual transfer instruction containing a summary of the question. Step B: Push the manual transfer instruction and pre-processed entity recognition, reasoning attempt, and other information to the human agent, who will then take over and respond. Step C: Collect manually answered data, label the correct reasoning paths and policy basis, and generate a knowledge graph to feed back into the data package.

Images