Machine learning intelligent question-answering analysis method and question-answering system based on government services

Abstract

The present invention relates to the field of artificial intelligence technology, and in particular to a machine learning intelligent question-answering analysis method and question-answering system based on government services. The method comprises the following steps: collecting government service data based on a multi-channel government service data collection engine to obtain government service data; designing a combination matrix of government service question-answer associations based on the government service data to generate a government service question-answering combination matrix; performing hierarchical topic relationship analysis of question-answer combinations on the government service question-answering combination matrix to generate hierarchical topic relationship data of question-answer combinations; designing a government service intelligent question-answering model based on the hierarchical topic relationship data of question-answer combinations; receiving real-time government service user question data; transmitting the real-time government service user question data to the government service intelligent question-answering model to perform government service multi-answer intelligent feedback processing to obtain government service multi-answer intelligent ranking data. The present invention realizes efficient and accurate government service question-answering intelligent feedback.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a machine learning intelligent question-answering analysis method and a question-answering system based on government services. Background Art

[0002] With the comprehensive digital transformation of government services, increasing demand for government services is shifting from offline to online. Government affairs are often manually dependent, resulting in heavy workloads and high repetitiveness. With the development of artificial intelligence (AI), intelligent question-and-answer (Q&A) has been widely used in intelligent customer service Q&A systems for government services. However, existing Q&A analysis methods for government services often rely on diverse and complex data sources, including historical question-and-answer data and government service knowledge base data. When dealing with content with strong logic and multi-level connections within government services, they lack a comprehensive solution. This results in insufficient understanding depth and feedback accuracy in Q&A models. Given the diverse nature of user questions, they are unable to automatically generate targeted, intelligent ranking and feedback for multiple answers. Summary of the Invention

[0003] Based on this, the present invention provides a machine learning intelligent question and answer analysis method and question and answer system based on government services to solve at least one of the above technical problems.

[0004] To achieve the above objectives, a machine learning intelligent question-answering analysis method based on government services includes the following steps:

[0005] Step S1: deploying a multi-channel government service data collection engine, and collecting government service data based on the multi-channel government service data collection engine to obtain government service data, wherein the government service data includes historical government service question data, historical government service answer data, and government service knowledge base data;

[0006] Step S2: performing bidirectional word segmentation analysis on the government service data to generate government service word segmentation analysis data; performing government service semantic trunk triple analysis based on the government service word segmentation analysis data to generate government service semantic trunk triple data; performing a combination matrix design of government service question-answer association on the government service semantic trunk triple data to generate a government service question-answer combination matrix;

[0007] Step S3: performing a topic distribution probability analysis of the question-answer combination on the government service question-answer combination matrix to generate topic distribution probability data of the government service question-answer combination; performing a hierarchical topic relationship analysis of the question-answer combination based on the topic distribution probability data of the government service question-answer combination to generate hierarchical topic relationship data of the question-answer combination;

[0008] Step S4: Establish an intelligent feedback mapping relationship of government service questions and answers through the question-answer combination hierarchical subject relationship data to obtain a government service intelligent question-answer model; receive instant government service user question data; transmit the instant government service user question data to the government service intelligent question-answer model for government service multi-answer intelligent feedback processing to obtain government service multi-answer intelligent sorting data; transmit the government service multi-answer intelligent sorting data to the terminal to execute the government service intelligent question-answer feedback operation.

[0009] Furthermore, step S1 includes the following steps:

[0010] Step S11: Obtain the government service authorization API interface;

[0011] Step S12: performing interface source address identification processing on the government service authorization API interface to obtain an identified government service authorization API interface;

[0012] Step S13: Designing a government service data verification script;

[0013] Step S14: deploying a multi-channel government service data collection engine based on the government service authorization API interface and the government service data verification script;

[0014] Step S15: Collect government service data based on the multi-channel government service data collection engine to obtain government service data.

[0015] Furthermore, the government service data verification script described in step S13 is used to perform government service text compliance verification screening operations and government service text redundancy screening operations.

[0016] Furthermore, step S2 includes the following steps:

[0017] Step S21: performing bidirectional word segmentation analysis on the government service data to generate government service word segmentation analysis data;

[0018] Step S22: performing multi-scale word segmentation window analysis based on the government service word segmentation analysis data to generate multi-scale word segmentation window data;

[0019] Step S23: using the multi-scale word segmentation window data to perform window word segmentation part-of-speech identification analysis on the government service word segmentation parsing data to generate government service word segmentation part-of-speech data;

[0020] Step S24: performing government service part-of-speech entity relationship analysis based on the government service part-of-speech data to generate government service part-of-speech entity relationship data;

[0021] Step S25: performing government service semantic trunk triple analysis based on the government service part-of-speech entity relationship data to generate government service semantic trunk triple data;

[0022] Step S26: performing a government service question-answer correlation analysis based on the government service data to generate government service question-answer correlation data;

[0023] Step S27: Design a combination matrix of government service question and answer associations for the government service semantic backbone triple data using the government service question and answer association data to generate a government service question and answer combination matrix.

[0024] Furthermore, step S25 includes the following steps:

[0025] Step S251: Building a syntactic dependency tree of government services based on the preset BERT algorithm and government service word segmentation part-of-speech data to generate a syntactic dependency tree model of government services;

[0026] Step S252: performing tree model dependency type optimization processing on the government service syntax dependency tree model to generate an optimized government service syntax dependency tree model;

[0027] Step S253: using the optimized government service syntactic dependency tree model to perform government service syntactic dependency trunk feature analysis on the government service part-of-speech entity relationship data to generate government service syntactic dependency trunk feature data;

[0028] Step S254: Perform government service semantic trunk triple analysis based on the government service syntactic dependency trunk feature data to generate government service semantic trunk triple data.

[0029] Furthermore, step S3 includes the following steps:

[0030] Step S31: Using the LDA topic model to perform topic distribution probability analysis on the government service question and answer combination matrix, and generate government service question and answer combination topic distribution probability data;

[0031] Step S32: performing question-answer combination topic identification based on the government service question-answer combination topic distribution probability data to generate question-answer combination topic data;

[0032] Step S33: performing topic distribution probability identification processing on subsets within clusters of question-answer combination clusters based on the government service question-answer combination matrix and the government service question-answer combination topic distribution probability data, and generating topic distribution probability data on subsets within clusters of question-answer combination clusters;

[0033] Step S34: performing sub-topic analysis and identification processing within the question-answer combination cluster based on the subset topic distribution probability data within the question-answer combination cluster to generate sub-topic data within the question-answer combination cluster;

[0034] Step S35: Perform hierarchical topic relationship analysis of the question-answer combination based on the question-answer combination topic data and the sub-topic data within the question-answer combination cluster to generate question-answer combination hierarchical topic relationship data.

[0035] Furthermore, step S33 includes the following steps:

[0036] Step S331: performing a question-answer combination similarity analysis on the government service question-answer combination matrix to generate government service question-answer combination similarity data; performing a government service question-answer combination cluster analysis based on the government service question-answer combination similarity data to generate government service question-answer combination cluster data;

[0037] Step S332: Map the government service question and answer combination topic distribution probability data to the government service question and answer combination cluster data to perform intra-cluster subset topic distribution probability identification processing to generate the question and answer combination intra-cluster subset topic distribution probability data.

[0038] Furthermore, step S34 includes the following steps:

[0039] Step S341: performing topic feature analysis on the subset topic distribution probability data within the question-answer combination cluster to generate topic feature data of the subset topic within the question-answer combination cluster;

[0040] Step S342: reconstructing the topic of the subset within the question-answer combination cluster based on the topic feature data of the subset within the question-answer combination cluster to generate the topic reconstructed data of the subset within the question-answer combination cluster;

[0041] Step S343: performing sub-topic identification processing within the question-answer combination cluster based on the subset topic reconstruction data within the question-answer combination cluster to generate sub-topic data within the question-answer combination cluster.

[0042] Furthermore, step S4 includes the following steps:

[0043] Step S41: Establishing an intelligent feedback mapping relationship of government service questions and answers through question-answer combination hierarchical topic relationship data to obtain a government service intelligent question-answering model;

[0044] Step S42: receiving user question data of instant government service;

[0045] Step S43: transmitting the instant government service user question data to the government service intelligent question and answer model for intelligent feedback processing of multiple answers to the government service, including: matching the user question data with the question and answer combination hierarchical topic relationship data through the government service intelligent question and answer model to generate government service multiple matching answer data; performing government service multiple matching answer confidence analysis based on the government service multiple matching answer data to generate government service multiple matching answer confidence data; using the government service multiple matching answer confidence data as a government service multiple answer ranking mechanism, and performing multiple answer intelligent ranking and confidence output on the government service multiple matching answer data through the government service multiple answer ranking mechanism to obtain government service multiple matching answer intelligent ranking data;

[0046] Step S44: Transmitting the government service multi-answer intelligent sorting data to the terminal to execute the government service intelligent question and answer feedback task.

[0047] This specification provides a machine learning intelligent question-answering system based on government services, which is used to perform the machine learning intelligent question-answering analysis method based on government services as described above. The machine learning intelligent question-answering system based on government services includes:

[0048] A data collection module, configured to deploy a multi-channel government service data collection engine, and collect government service data based on the multi-channel government service data collection engine to obtain government service data, wherein the government service data includes historical government service question data, historical government service answer data, and government service knowledge base data;

[0049] The question-answer combination standardization module is used to perform bidirectional word segmentation and parsing on government service data to generate government service word segmentation and parsing data; perform government service semantic backbone triple analysis based on government service word segmentation and parsing data to generate government service semantic backbone triple data; and design a combination matrix for government service question-answer association based on government service semantic backbone triple data to generate a government service question-answer combination matrix.

[0050] The question-answer combination topic identification module is used to perform topic distribution probability analysis of the government service question-answer combination matrix to generate government service question-answer combination topic distribution probability data; based on the government service question-answer combination topic distribution probability data, it performs hierarchical topic relationship analysis of the question-answer combination to generate hierarchical topic relationship data of the question-answer combination;

[0051] The intelligent question and answer feedback module is used to establish an intelligent feedback mapping relationship of government service questions and answers through the question and answer combination hierarchical subject relationship data to obtain a government service intelligent question and answer model; receive real-time government service user question data; transmit the real-time government service user question data to the government service intelligent question and answer model for government service multi-answer intelligent feedback processing to obtain government service multi-answer intelligent sorting data; transmit the government service multi-answer intelligent sorting data to the terminal to execute the government service intelligent question and answer feedback operation.

[0052] The beneficial effect of this application is that the deployment of a multi-channel government service data collection engine and the collection of government service data have a significant effect on improving data quality and the accuracy of the question-and-answer system. By obtaining the government service authorization API interface and performing interface source address identification processing, it is possible to ensure that government service data is obtained legally and securely from official channels, ensuring the reliability of the data source and avoiding interference with system performance by external untrusted data. The designed government service data verification script performs compliance verification screening operations and text redundancy screening operations, further enhancing the accuracy of data processing. Text compliance verification can effectively eliminate questions and answers that do not comply with government service specifications, while text redundancy screening can remove redundant data and duplicate content, thereby ensuring the quality and uniqueness of the data. Through this layer of screening, the system can focus more on high-quality data and improve the efficiency and accuracy of subsequent processing. By combining the government service authorization API interface identification and the multi-channel data collection engine deployed with the verification script, it is possible to automatically and efficiently collect government service-related data from multiple channels. Through two-way word segmentation parsing and semantic backbone triple analysis technology, the semantic understanding ability of government service data is significantly improved, enabling the question-and-answer system to better handle complex and diverse natural language questions. Bidirectional word segmentation and parsing solves the problem of traditional unidirectional word segmentation inaccurately capturing contextual information by analyzing the contextual relationships between words from both the left and right perspectives. By generating high-quality word segmentation and parsing data, the system can more accurately understand keywords and important information in user questions, improving the accuracy of semantic analysis. Multi-scale word segmentation window analysis based on the generated word segmentation and parsing data helps the system capture semantic information at different levels, further enhancing its understanding of diverse language structures. Multi-scale word segmentation window analysis resolves ambiguity in government service texts by performing word combination and semantic analysis at different scales, improving the system's accuracy when processing long and complex sentences and polysemous words. Part-of-speech tag analysis based on word segmentation data further refines the depth of semantic understanding. By assigning grammatical tags to each word (such as noun, verb, adjective, etc.), the system can distinguish different types of information and correctly understand the overall structure of the sentence, which helps improve the accuracy of subsequent semantic analysis. Through the syntactic dependency tree model and semantic backbone triple analysis, the system uses the BERT algorithm and optimized syntactic dependency tree model to deeply explore the core grammatical and semantic structure of the text, thereby extracting key information from government Q&A. This syntactic tree structure not only solves the problem of grammatical dependency but also effectively identifies important components of the Q&A, such as the subject, predicate, and object, providing more accurate semantic support for subsequent Q&A analysis and matching. The LDA topic model is used to perform topic distribution probability analysis on the government service Q&A combination matrix, revealing the underlying topic structure within the government service Q&A.The application of the Latent Dirichlet Allocation (LDA) topic model effectively mines the distribution of different topics from large amounts of question-and-answer data and generates topic distribution probability data for government service question-and-answer combinations. This automatically identifies and differentiates different categories of government service questions, enabling the system to accurately identify and classify diverse government service needs. Further analysis and processing of this topic distribution probability data allows the system to assign a clear topic identifier to each question-and-answer combination, making subsequent question-and-answer processing clearer and more standardized. Each question-and-answer combination has a corresponding topic label, helping the system quickly locate relevant reference information when processing user questions, avoiding inefficient irrelevant searches and duplicate processing, thereby improving response speed and accuracy. Similarity analysis and cluster analysis of question-and-answer combinations further enhance the system's ability to handle complex questions. Similarity analysis helps the system understand the connections between different government service questions, while cluster analysis groups question-and-answer combinations with similar topics, effectively reducing computational overhead and improving search efficiency. By mapping topic distribution probabilities to clusters, the system can identify subsets within each cluster, thereby improving the system's ability to accurately respond to multiple similar question scenarios. Feature analysis and reconstruction of the topic distribution probability data for cluster subsets generates more refined sub-topic data for each question-answer combination cluster. This refined processing ensures that each question-answer combination can be further analyzed and responded to based on its sub-topics, enabling more detailed and diverse government service responses. By establishing intelligent feedback mappings based on hierarchical topic relationship data for question-answer combinations, the system accurately matches users' real-time questions with relevant information in the government service knowledge base, ensuring highly relevant and accurate answers. This mapping not only improves the matching accuracy of user questions but also reduces incorrect responses due to semantic misunderstanding. Receiving real-time government service user question data and transmitting it to the government service intelligent question-answering model for processing ensures that every user question receives a timely and accurate response. During the government service multi-answer intelligent feedback processing phase, the system analyzes the confidence of multiple matching answers, prioritizing the most relevant and correct answers based on their credibility. This phase further enhanced the accuracy and robustness of the intelligent question-answering system by generating confidence data for multiple matching answers to government services. The application of a multi-answer ranking mechanism for government services enables the system to intelligently sort multiple answers based on confidence, and through confidence output, ensures that users receive the most reliable answers. The introduction of this mechanism significantly improves the reliability of the question-answering system and the user experience, as users receive the answer that best meets their needs, rather than a barrage of redundant and uncertain options. Transmitting this intelligent ranking data for multiple government service answers to the terminal ensures that the terminal can successfully execute the government service intelligent question-answering feedback process.In this way, users can not only obtain high-quality answers, but also accelerate the feedback of questions and answers through the intelligent sorting mechanism, significantly improving the response efficiency of government services.

[0053] Therefore, the machine learning intelligent question and answer analysis method based on government services of the present invention can efficiently integrate historical question data, answer data and government service knowledge base data through the deployment of a multi-channel government service data collection engine, and use data verification scripts to screen the collected data to ensure the accuracy and compliance of the data. By adopting two-way word segmentation analysis and multi-scale word segmentation window technology, the system can deeply explore the semantic characteristics of government service data, and further optimize the semantic parsing process through the syntactic dependency tree model based on the BERT algorithm, and effectively improve the understanding of the semantics of government services through the semantic backbone triple analysis technology. In particular, the optimization of the syntactic dependency tree model based on BERT can accurately capture the logical associations and hierarchical structures in the government service text, so that the processing depth of the question and answer model for complex natural language is significantly enhanced, solving the shortcomings of the existing technology in semantic parsing and logical reasoning. By designing a question and answer combination matrix and a topic distribution probability analysis based on the LDA topic model, the hierarchical topic relationship construction and intelligent feedback mapping mechanism of the question and answer combination are realized, which can provide diversified and highly accurate intelligent feedback according to the user's immediate question content, further improve the rationality of the answer priority, and significantly improve the user experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 This is a flowchart of the steps of a machine learning intelligent question-answering analysis method based on government services of the present invention;

[0055] Figure 2 for Figure 1 Detailed implementation steps of step S2 in FIG.

[0056] Figure 3 for Figure 1 Detailed implementation steps of step S3 in FIG.

[0057] The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings and in conjunction with the embodiments. DETAILED DESCRIPTION

[0058] The following is a clear and complete description of the technical method of the present invention in conjunction with the accompanying drawings. It is obvious that the embodiments described are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making any creative efforts are within the scope of protection of the present invention.

[0059] In addition, the accompanying drawings are merely schematic illustrations of the present invention and are not necessarily drawn to scale. Identical reference numerals in the figures denote identical or similar parts, and thus repetitive descriptions thereof will be omitted. Some of the block diagrams shown in the accompanying drawings are functional entities that do not necessarily correspond to physically or logically separate entities. These functional entities may be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor and / or microcontroller approaches.

[0060] It should be understood that although the terms "first," "second," and the like are used herein to describe various elements, these elements should not be limited by these terms. These terms are used solely to distinguish one element from another. For example, a first element may be referred to as a second element, and similarly, a second element may be referred to as a first element, without departing from the scope of the exemplary embodiments. The term "and / or" as used herein includes any and all combinations of one or more of the listed associated items.

[0061] To achieve this, please refer to Figures 1 to 3 The present invention provides a machine learning intelligent question-answering analysis method based on government services. In the embodiments of the present invention, please refer to Figure 1 FIG. 1 is a flow chart of a method for intelligent question-answering analysis based on machine learning for government services according to the present invention. The method comprises the following steps:

[0062] Step S1: deploying a multi-channel government service data collection engine, and collecting government service data based on the multi-channel government service data collection engine to obtain government service data, wherein the government service data includes historical government service question data, historical government service answer data, and government service knowledge base data;

[0063] In the implementation of the government service system, according to an embodiment of the present invention, corresponding API interface permissions are obtained by accessing the government service provider's authorization management platform. During specific implementation, a catalog of API interfaces representing the source of government service data must be clearly defined, such as the "User Basic Information Query Interface," the "Approval Process Progress Interface," and the "Government Content Interpretation Interface." During implementation, an interface authorization application must first be submitted to the government service provider, including a description of the intended interface's purpose, access frequency requirements, and the security authentication protocol. Upon review, the government service provider generates a corresponding API interface key (such as an OAuth 2.0 key) and returns an interface authorization file containing the API call documentation. For the obtained government service API interface, source address identification is performed to ensure accurate invocation in a multi-interface environment. The source address (i.e., request URL) of each API interface is resolved into a unique identifier. Each time an interface is called, the system dynamically retrieves the corresponding source address based on the identifier to ensure accuracy and consistency. A validation script must be designed for the collected government service data to ensure data integrity and validity. The core logic of the validation script includes text compliance verification and redundant data filtering. Text compliance verification involves checking the collected data against government service standards. A multi-channel government service data collection engine is constructed based on the previously identified API interface information and verification scripts. In implementation, the open-source framework Scrapy is used to build a distributed crawling module, combined with Celery for task scheduling. First, the interface identifier and its corresponding URL are configured as input parameters for the collection task. Second, the verification script serves as a post-collection processing module, filtering out noncompliant data while retaining valid data. This data is then stored in a local or cloud database in a structured format (such as JSON or CSV). The deployed data collection engine is launched, and automated calls to various API interfaces are executed by setting a scheduling frequency and execution cycle. For example, for the "Approval Process Progress Interface," the call frequency is set to once per hour. A scheduling script is used to initiate HTTP GET requests and store the returned data locally. After each collection task is completed, the task ID, execution time, return status code, and number of collected data items are logged. If the collected data contains an error status code or an empty value, the exception alert module is immediately triggered to send an email or SMS notification to the administrator for troubleshooting. Through this step, a structured government service data set is finally obtained, providing high-quality input for subsequent question-and-answer analysis.

[0064] Step S2: performing bidirectional word segmentation analysis on the government service data to generate government service word segmentation analysis data; performing government service semantic trunk triple analysis based on the government service word segmentation analysis data to generate government service semantic trunk triple data; performing a combination matrix design of government service question-answer association on the government service semantic trunk triple data to generate a government service question-answer combination matrix;

[0065] In an embodiment of the present invention, the text of government service data is first parsed using bidirectional word segmentation technology. Using open-source word segmentation tools such as Jieba or HanLP, Chinese word segmentation is first performed, breaking the input text (e.g., "Government affairs execution status has been released to the public") into word sequences. Bidirectional word segmentation is then performed, identifying words both left-to-right and right-to-left, ensuring that semantic connections between words are captured during the processing. During this process, proper noun recognition is performed for specific named entities, such as "Government affairs execution status," to ensure the accuracy of the word segmentation results. The results are then converted into word sequences and stored as structured data. Multi-scale word segmentation window analysis aims to capture semantic hierarchical relationships through lexical aggregation analysis at different scales (window sizes). In implementation, using word segmentation parsing data, analysis windows of varying sizes are first set, typically 1, 2, or 3 words. A sliding window approach is then used to generate window word sequences of varying sizes. Window word segmentation and part-of-speech tagging analysis is performed on government service word segmentation and parsing data using multi-scale word segmentation window data. Based on the multi-scale word segmentation window data, a part-of-speech tagging tool (such as Stanza or HanLP) is used to tag each word with its part of speech. Specifically, the word segmentation results are input and the tool performs part-of-speech tagging to determine each word's grammatical role (e.g., noun, verb, adjective, etc.). This process includes: parsing the word segmentation results; tagging each word using an existing part-of-speech tagging model; recording the tagging results in the word segmentation and parsing data, and generating part-of-speech tagging data. This tagging data includes each word's part-of-speech tag, part-of-speech category (e.g., verb, noun), and position identifier to facilitate subsequent analysis. Based on government service part-of-speech data, we conduct part-of-speech entity relationship analysis. Using the annotated part-of-speech data, we identify and analyze entity relationships between words. Using dependency parsing methods (such as Spacy or StanfordNLP), we perform syntactic dependency analysis on the part-of-speech data, constructing a relationship graph for all related words. Each entity and its relationship is stored as a triple of "entity 1-relationship-entity 2" (e.g., "government affairs-execution-situation"). Based on the part-of-speech entity relationship data, we perform semantic backbone triple analysis to extract core semantic information from the text. We apply NLP tools (such as BERT or RoBERTa) for deep semantic understanding. By further processing the entity relationship data, we identify key triples in sentences. Deep learning models are used to optimize the semantic structure of triples to ensure the accuracy of the analysis results. The generated triple data is stored in a structured database table for subsequent query and analysis. We conduct association analysis on questions and answers in government service data. Using statistical methods or machine learning techniques, we identify the correlation between questions and answers. By analyzing historical question and answer data, we can extract effective association rules from the text.Historical question-and-answer data is classified and clustered. Pattern recognition methods are then used to generate association information based on historical questions and answers, as well as corresponding similar questions and answers. This information is then stored and labeled using data association. A combination matrix for government service question-and-answer associations is designed using the government service semantic backbone triple data using the government service question-and-answer association data to generate a government service question-and-answer combination matrix. Relationship data between users' historical questions and answers is extracted from the government service question-and-answer association data. This data typically includes the keywords and intent of the user's questions, as well as the system-provided answers. By analyzing the semantic backbone triple data, representative semantic backbone information, such as the relationship between the subject, verb, and object, is extracted. A combination matrix is then designed based on the aforementioned government service question-and-answer association data and the semantic backbone triple data. Specifically, each row represents a key triple for a question-and-answer scenario, while each column represents a different question-and-answer topic or intent. By comparing the semantic similarity and correlation between different question-and-answer scenarios, the system generates a question-and-answer combination matrix, in which each element represents the strength of association or degree of match between different question-and-answer scenarios. Through the design of this combined matrix, the system can identify which question and answer content is relevant, thereby effectively improving the accuracy and intelligence level of question and answer matching.

[0066] Step S3: performing a topic distribution probability analysis of the question-answer combination on the government service question-answer combination matrix to generate topic distribution probability data of the government service question-answer combination; performing a hierarchical topic relationship analysis of the question-answer combination based on the topic distribution probability data of the government service question-answer combination to generate hierarchical topic relationship data of the question-answer combination;

[0067] In this embodiment of the present invention, a Latent Dirichlet Allocation (LDA) topic model is used to perform topic distribution probability analysis on a government service question-answer combination matrix. Using the government service question-answer combination matrix as input, the text in the matrix is first preprocessed, including stop word removal, word segmentation, and word frequency statistics. The preprocessed text data is then vectorized into a bag-of-words model and input into the LDA model for topic modeling analysis. The LDA model iteratively calculates a set of topic distribution probability data for each question-answer combination. The output is stored in a matrix format, with rows representing question-answer combinations, columns representing topics, and cell values representing topic probabilities. Based on the topic distribution probability data generated by the LDA model, a topic identification process is performed on each question-answer combination, and the dominant topic for each question-answer combination is determined based on the topic distribution probabilities. The dominant topic is determined based on the highest probability value. The generated question-answer combination topic data is stored in a structured table format, with each record containing the question-answer combination, the dominant topic number, and the corresponding keyword set. Based on the question-answer combination matrix and topic distribution probability data, question-answer combination clustering and cluster subset topic distribution probability analysis are performed. A hierarchical clustering algorithm is used to cluster question-answer combinations. The input data is the question-answer combination matrix and its topic distribution probability. Cosine similarity is used to calculate the similarity between question-answer combinations. A cluster tree is constructed based on the similarity, and a threshold is set for clustering. For each cluster, the subset topic distribution probability is further calculated for the question-answer combination within it. The output within-cluster subset topic distribution probability data includes cluster number, topic probability distribution, and dominant topic. The subset topic distribution probability data within the question-answer combination cluster is further analyzed to identify subtopics within each cluster. Subtopics with significant differences are first extracted based on the within-cluster subset topic distribution probability data. Subtopics are identified by calculating the distribution differences between topics (such as KL divergence). For subsets with significant topic differences, topic feature extraction methods (such as TF-IDF) are used to extract keywords to identify the semantic content of the subtopics. The output sub-topic data within the cluster is stored in a structured form, including the cluster number, sub-topic number, and its keyword set. A hierarchical topic relationship of the question-answer combination is constructed based on the question-answer combination topic data and the sub-topic data within the cluster. A hierarchical structure is constructed according to the subordinate relationship between topics and sub-topics, using a tree data structure. Each node represents a topic or sub-topic, and an edge represents a subordinate relationship. Through relationship analysis, the association path between each question-answer combination and its corresponding topic is determined. The generated hierarchical topic relationship data is stored in a tree structure or nested table format, with each record containing the topic hierarchical path and question-answer combination association information.

[0068] Step S4: Establish an intelligent feedback mapping relationship of government service questions and answers through the question-answer combination hierarchical subject relationship data to obtain a government service intelligent question-answer model; receive instant government service user question data; transmit the instant government service user question data to the government service intelligent question-answer model for government service multi-answer intelligent feedback processing to obtain government service multi-answer intelligent sorting data; transmit the government service multi-answer intelligent sorting data to the terminal to execute the government service intelligent question-answer feedback operation.

[0069] In this embodiment of the present invention, a graph neural network (GNN) model is used to construct intelligent feedback mapping relationships using hierarchical topic relationship data from question-answer combinations as input. First, the hierarchical topic relationship data from question-answer combinations is mapped into a graph structure, with nodes representing topics and edges representing hierarchical relationships between topics. Edge weights are determined by the strength of the topic relationships. The GNN model then extracts features and models the relationships within this graph. Training employs a supervised learning approach, with the cross-entropy loss as the objective function and a learning rate set to 0.001. The model then outputs an intelligent government service question-answering model capable of efficiently mapping user questions to corresponding answers based on hierarchical topic relationships. User-input government service question data is received in real time via an API interface in the form of a JSON object. The system uses regular expressions to validate the input data to ensure field integrity and character validity. Once validated, the question text is extracted and normalized, including removing extra spaces, converting to lowercase, and removing stop words. The processed question text is then stored in vector form, for example, using TF-IDF to represent the question as a vector. The real-time question data is then matched against the hierarchical topic relationship data from question-answer combinations in the intelligent question-answering model. Using a BERT-based sentence embedding method, user questions and hierarchical topic questions are vectorized and cosine similarity is calculated. Matching and association are performed based on the similarity analysis results to generate multi-match answer data for government services. The similarity scores of the multi-match answer data are normalized, for example, by using the Softmax function to convert similarity into a confidence distribution. The generated multi-match answer confidence data includes matching answers and their corresponding confidence levels. A sorting mechanism is constructed based on the confidence data. A descending sorting strategy is used to prioritize answers with high confidence levels. Confidence annotations are added to the sorted answers to generate intelligently sorted multi-answer data for government services. After receiving the results, the user terminal calls the display module to generate user-friendly feedback results, such as displaying the answer content and confidence level in the form of cards.

[0070] Furthermore, step S1 includes the following steps:

[0071] Step S11: Obtain the government service authorization API interface;

[0072] Step S12: performing interface source address identification processing on the government service authorization API interface to obtain an identified government service authorization API interface;

[0073] Step S13: Designing a government service data verification script;

[0074] Step S14: deploying a multi-channel government service data collection engine based on the government service authorization API interface and the government service data verification script;

[0075] Step S15: Collect government service data based on the multi-channel government service data collection engine to obtain government service data.

[0076] In an embodiment of the present invention, in the implementation of the government service system, the corresponding API interface permissions are obtained by accessing the authorization management platform of the government service provider. In the specific implementation process, it is necessary to clarify the API interface catalog list of the government service data source, for example, "User Basic Information Query Interface", "Approval Process Progress Interface" and "Government Interpretation Content Interface". In the implementation process, it is first necessary to submit an interface authorization application to the government service provider, including a description of the purpose of the application interface, access frequency requirements and security authentication protocol content. After the review, the government service provider generates the corresponding API interface key (such as an OAuth 2.0 key) and returns an interface authorization file containing the API call document. Taking the interface "Government Interpretation Content Interface" as an example, the authorization file includes the HTTP request URL of the interface, the authentication fields required for the request header (such as Authorization: Bearer <token>), supported HTTP methods (such as GET or POST), and required parameter fields (such as policy_id and request_time). For obtained government service API interfaces, source address identification is required to ensure accurate invocation in a multi-interface environment. Specifically, each API interface's source address (i.e., request URL) is parsed into a unique identifier. For example, for the URL of the "Government Affairs Interpretation Content Interface," the identifier GOV_API_POLICY_DETAILS is generated and stored in the local database table api_interface_mapping. The table contains fields such as the interface name (such as "Government Affairs Interpretation Content Interface"), the interface URL, the interface identifier (such as GOV_API_POLICY_DETAILS), and the authentication key. Each time an interface is called, the system dynamically retrieves the corresponding source address from the api_interface_mapping table based on the identifier to ensure call accuracy and consistency. For collected government service data, a validation script must be designed to ensure data integrity and validity. The core logic of the validation script consists of text compliance verification and redundant data screening. Text compliance verification involves checking the collected data against government service standards. For example, in approval progress data, the "status" field must be verified to contain predefined values (such as "Accepted," "Under Review," and "Completed"). The "apply_date" time field must be formatted in YYYY-MM-DD format. Redundant data filtering is performed based on MD5 hash calculations, comparing newly collected data with historical data to ensure duplicates. If a data entry's hash value already exists in the historical data table, it is marked as redundant and filtered out. The verification script is implemented in Python, combining regular expressions for content matching and SQL stored procedures for redundancy detection. Based on the identified API interface information and the verification script, a multi-channel government service data collection engine is constructed. In implementation, the open source framework Scrapy is used to build a distributed crawling module, combined with Celery for task scheduling. First, the interface identifier and its corresponding URL are configured as input parameters for the collection task. For example, in the task configuration file api_tasks.json, the following is recorded: {"task_id":"1","api_id":"GOV_API_POLICY_DETAILS","frequency":"hourly"}. Secondly, verification scripts are used as post-collection processing modules to filter out non-compliant data, retain valid data, and store the valid data in a structured format (such as JSON or CSV) in a local or cloud database.For example, collected government interpretation data is stored in the policy_details table according to the fields policy_id, title, content, and publish_date to ensure data quality. The deployed data collection engine is launched, and automated calls to various API interfaces are made by setting a scheduling frequency and execution cycle. For example, for the "Approval Process Progress Interface," the call frequency is set to once per hour. A scheduling script is used to initiate HTTP GET requests, and the returned data is stored locally. After each collection task is completed, the task ID, execution time, return status code (such as 200 or 500), and number of collected data items are recorded in the log. If the collected data contains an error status code or a null value, the exception alert module is immediately triggered to send an email or SMS notification to the administrator for troubleshooting. This step ultimately results in a structured collection of government service data, providing high-quality input for subsequent Q&A analysis.

[0077] Furthermore, the government service data verification script described in step S13 is used to perform government service text compliance verification screening operations and government service text redundancy screening operations.

[0078] In this embodiment of the present invention, the compliance verification and screening job for government service text checks the validity of data formats and field contents based on predefined rules. Taking data from the "Government Service Interpretation Content Interface" as an example, the compliance verification script checks the fields in the returned government service interpretation data. The script first verifies that the policy_id field is a string type and meets a specific length range (for example, 8 to 12 characters). Next, it verifies that the title field is a non-empty string and does not exceed 100 characters. It also verifies that the content field contains valid government service interpretation content and is between 500 and 5000 characters long. Finally, the script verifies that the publish_date field conforms to the YYYY-MM-DD date format. The government service text redundancy screening job aims to remove duplicate data, improve data quality, and avoid unnecessary duplicate storage. This process calculates a hash value of the text content and compares the newly collected data with the data already stored in the database. If the hash values are identical, the data is considered redundant. The MD5 algorithm is used to generate a hash value for the data content, for example, using the hashlib library to calculate the MD5 value of the content field. The data already stored in the database will be compared with the hash value of the new data. If data with the same hash value already exists, the new data will not be stored.

[0079] Further, as an embodiment of the present invention, refer to Figure 2 As shown, Figure 1 Detailed step flow diagram of step S2 in the embodiment, step S2 includes the following steps:

[0080] Step S21: performing bidirectional word segmentation analysis on the government service data to generate government service word segmentation analysis data;

[0081] In an embodiment of the present invention, the text of the government service data is first parsed using a bidirectional word segmentation technology. Using an open source word segmentation tool such as Jieba or HanLP, Chinese word segmentation is first performed to divide the input text (for example, "The implementation status of government affairs has been released to the public") into a sequence of words. Then, bidirectional word segmentation is performed, that is, word recognition is performed from left to right and from right to left at the same time to ensure that the semantic connection between words can be captured during the processing. In this process, specific named entities, such as "government implementation status", are recognized as proper nouns to ensure the accuracy of the word segmentation results. Finally, the word segmentation results are converted into a word sequence (such as "government affairs", "implementation", "situation", "already", "to", "public", "released") and stored as structured data, such as JSON format, recording the starting position and ending position of each word. The generated word segmentation parsing data serves as the basis for subsequent steps.

[0082] Step S22: performing multi-scale word segmentation window analysis based on the government service word segmentation analysis data to generate multi-scale word segmentation window data;

[0083] In this embodiment of the present invention, multi-scale word segmentation window analysis aims to capture semantic hierarchical relationships through lexical aggregation analysis at different scales (window sizes). In implementation, using word segmentation parsing data, analysis windows of varying sizes are first set, typically 1, 2, or 3 words. Using a sliding window approach, window word sequences of varying sizes are generated. For example, for the word segmentation result "government affairs execution status has been released to the public," window data of scales 1, 2, and 3 are generated: a 1-word window: ["government affairs", "execution", "situation", "has", "to", "public", "released"]; a 2-word window: ["government affairs execution", "execution status", "situation has", "has been released", "released to the public", "released to the public"]; and a 3-word window: ["government affairs execution status", "execution status", "situation has", "situation has been released", "has been released to the public", "released to the public"]. This analysis provides a deeper understanding of the potential semantics of words in different contexts. The generated window data is stored in a structured format with multiple fields for subsequent processing.

[0084] Step S23: using the multi-scale word segmentation window data to perform window word segmentation part-of-speech identification analysis on the government service word segmentation parsing data to generate government service word segmentation part-of-speech data;

[0085] In an embodiment of the present invention, when performing part-of-speech tagging analysis, a part-of-speech tagging tool (such as Stanza or HanLP) is used to tag the part of speech for each word based on the multi-scale word segmentation window data. During the specific operation, the word segmentation results are input, and the tool is used to perform part-of-speech tagging to determine the grammatical role of each word (such as noun, verb, adjective, etc.). For example, for the window "Government Affairs Execution Situation", the part-of-speech tagging result is: "Government Affairs (noun), Execution (verb), Situation (noun)". The process includes: parsing the word segmentation results; tagging each word using an existing part-of-speech tagging model; recording the tagging results in the word segmentation parsing data, and generating part-of-speech tagging data. The identification data includes the part-of-speech tag, part-of-speech category (such as verb, noun), and position identifier of each word to facilitate subsequent analysis.

[0086] Step S24: performing government service part-of-speech entity relationship analysis based on the government service part-of-speech data to generate government service part-of-speech entity relationship data;

[0087] In an embodiment of the present invention, when performing part-of-speech entity relationship analysis, the entity relationships between words are identified and analyzed using the labeled part-of-speech data, and a dependency syntactic analysis method (such as Spacy or StanfordNLP) is used to perform syntactic dependency analysis on the part-of-speech data. For example, for the sentence "The implementation status of government affairs has been released to the public," "government affairs" and "implementation" are first identified as subject-predicate relationships, and a relationship graph is constructed: "government affairs (subject) to implementation (predicate)." Through dependency analysis, the relationship between "government affairs" and "implementation" is generated, and combined with other part-of-speech identifiers, a relationship graph of all related words is constructed. Ultimately, each entity and its relationship is stored as a triple of "entity 1-relationship-entity 2" (such as "government affairs-implementation-situation").

[0088] Step S25: performing government service semantic trunk triple analysis based on the government service part-of-speech entity relationship data to generate government service semantic trunk triple data;

[0089] In an embodiment of the present invention, semantic backbone triple analysis is performed based on part-of-speech entity relationship data to extract core semantic information from the text. In specific implementations, NLP tools (such as BERT or RoBERTa) are used for deep semantic understanding, and key triples in sentences are identified through further processing of entity relationship data. For example, for the sentence "The implementation status of government affairs has been released to the public," the analysis process can identify: "government affairs" as the subject, "execution" as the action, and "situation" as the object, ultimately generating the triple data "government affairs-execution-situation." In addition, the semantic structure of the triples is optimized through a deep learning model to ensure the accuracy of the analysis results. The generated triple data is stored as a structured database table to facilitate subsequent query and analysis.

[0090] Step S26: performing a government service question-answer correlation analysis based on the government service data to generate government service question-answer correlation data;

[0091] In this embodiment of the present invention, association analysis is performed on questions and answers in government service data. Using statistical methods or machine learning techniques, the correlation between questions and answers is identified. By analyzing historical question and answer data, effective association rules can be extracted from the text. For example, the question "What is government interpretation?" is associated with the answer "Government interpretation refers to a detailed explanation of government content." The historical question and answer data is classified and clustered, and then pattern recognition methods are used to generate association information based on historical questions and answers, as well as corresponding similar questions and similar answers. This information is then stored and labeled using data association.

[0092] Step S27: Design a combination matrix of government service question and answer associations for the government service semantic backbone triple data using the government service question and answer association data to generate a government service question and answer combination matrix.

[0093] In an embodiment of the present invention, a combination matrix is designed to associate government service questions and answers with government service semantic backbone triple data using government service question and answer association data, generating a government service question and answer combination matrix. In specific implementation, the relationship data between users' historical questions and answers is first extracted from the government service question and answer association data. This data typically includes the keywords and intent of the user's questions, as well as the answers provided by the system. Next, representative semantic backbone information, such as the relationship between the subject, verb, and object, is extracted by analyzing the semantic backbone triple data. Then, a combination matrix is designed based on the government service question and answer association data and the semantic backbone triple data. In the specific design, each row represents a key triple for a question and answer scenario, while each column represents a different question and answer topic or intent. By comparing the semantic similarity and correlation between different question and answer scenarios, the system generates a question and answer combination matrix, in which each element represents the strength of association or matching between different question and answer scenarios. This combination matrix design enables the system to identify which question and answer content is relevant, thereby effectively improving the accuracy and intelligence of question and answer matching.

[0094] Furthermore, step S25 includes the following steps:

[0095] Step S251: Building a syntactic dependency tree of government services based on the preset BERT algorithm and government service word segmentation part-of-speech data to generate a syntactic dependency tree model of government services;

[0096] Step S252: performing tree model dependency type optimization processing on the government service syntax dependency tree model to generate an optimized government service syntax dependency tree model;

[0097] Step S253: using the optimized government service syntactic dependency tree model to perform government service syntactic dependency trunk feature analysis on the government service part-of-speech entity relationship data to generate government service syntactic dependency trunk feature data;

[0098] Step S254: Perform government service semantic trunk triple analysis based on the government service syntactic dependency trunk feature data to generate government service semantic trunk triple data.

[0099] In this embodiment of the present invention, a syntactic dependency tree for government services is constructed based on the preset BERT (Bidirectional Encoder Representations from Transformers) algorithm and part-of-speech data for government service word segmentation. The part-of-speech data for government service word segmentation is input into the BERT model. The BERT model utilizes a bidirectional Transformer architecture, which can capture the deep semantics of words through contextual information. After encoding each word, a dependency parsing algorithm (such as the dependency parser in Stanford NLP) is used to combine each word's part-of-speech tag to construct a dependency tree. The BERT model extracts the dependencies between words from the grammatical structure of each sentence and represents them as a tree structure. For example, the subject-verb relationship in the sentence "Government affairs execution status has been released to the public" is "government affairs (subject) to execution (verb)." This generates a complete syntactic dependency tree model, with nodes representing words and edges representing dependencies between words. The generated dependency tree model is saved as structured data, containing the parent node, child nodes, and relationship type for each word. The dependency types of the syntactic dependency tree model are further optimized. Based on the characteristics of government service data, the types and weights of dependency relationships are defined and optimized. The optimization process involves first analyzing whether each dependency type in the dependency tree (such as "subject-predicate relationship" and "verb-object relationship") conforms to the structural characteristics of government service data. For example, questions in government texts have specific dependency types, such as the "question-action-answer" structure. Then, a rule-based approach is used to adjust the dependency types in the tree, especially for complex syntactic structures, to better reflect semantic relationships. For example, the dependency type of the fixed phrase "government interpretation" is set to a "fixed collocation" type rather than a simple "verb-object relationship." Furthermore, by reweighting the dependency paths of words in the tree, key information nodes such as "government" or "execution" are given higher weights in the tree model. The optimized dependency tree model is output in a standardized format (such as JSON or XML) to ensure that grammatical structure information can be accurately captured and utilized in subsequent processing. The optimized syntactic dependency tree model is used to analyze the syntactic dependency backbone features of government service part-of-speech entity relationship data. Identify the core features in the syntactic dependency tree, which are the core grammatical structure of a sentence, typically including subject-verb-object relationships and key information points. By analyzing the optimized syntactic dependency tree, we extract the core feature data. For example, for the sentence "Government affairs execution status has been released to the public," the core relationship in the tree model is "Government affairs - Execution - Status." During analysis, we select the nodes and edges most closely related to the syntactic structure and extract them as the core features. For complex government service questions with nested clauses or modifiers, the optimized tree model can help accurately identify the core features.For example, in the sentence "The implementation status of government affairs has been released to the public", the main features of the sentence are "government affairs - implementation - status", while "has been" is an additional component. The generated syntactic dependency main feature data is stored in the form of triples, and each triple includes two entities and the relationship between them. The syntactic dependency main feature data of government affairs services is used for semantic main triple analysis. By extracting the main part of the syntactic dependency relationship tree (i.e., the most important grammatical relationship), it is transformed into structured triple data. Natural language processing techniques (such as BERT, Word2Vec, or rule - based semantic analysis methods) are used to perform semantic analysis on the extracted main features to ensure the integrity and accuracy of semantic information. For example, for the sentence "The implementation status of government affairs has been released to the public", the obtained triple is: "government affairs - implementation - status". For sentences containing multiple verbs or predicates, such as "The progress of government affairs implementation has been released on the website", context analysis is required to ensure the generation of correct triples, such as: "government affairs - implementation - progress" and "progress - release - website". On this basis, the optimization algorithm further confirms the correctness and semantic hierarchy of the triples through semantic relevance analysis, and sorts or annotates them to improve the response quality of the question - answering system. The generated semantic main triple data serves as the basis for subsequent question - answering analysis, enabling the system to more accurately understand user questions and provide correct answers.

[0100] Further, as an embodiment of the present invention, refer to Figure 3 shown in Figure 1 the detailed step - by - step process schematic diagram of step S3 in

[0101] Step S31: Use the LDA topic model to analyze the topic distribution probability of the question - answering combination matrix of government affairs services, and generate the topic distribution probability data of the government affairs services question - answering combination;

[0102] In this embodiment of the present invention, a Latent Dirichlet Allocation (LDA) topic model is used to perform topic distribution probability analysis on a combination matrix of government service questions and answers. The combination matrix of government service questions and answers is used as input data. The text in the matrix is first preprocessed, including stop word removal, word segmentation, and word frequency statistics. For example, a preset automated text processing engine is used to remove stop words from the matrix text, removing common terms such as "how" and "handle," while retaining domain keywords. The Jieba word segmentation tool is used to perform bidirectional word segmentation on questions and answers such as "How to apply for a residence permit" and "What materials are required for residence permit application?" to generate a word list. For example, the word "residence permit application conditions" is segmented into ["residence permit," "handle," "conditions"], and word frequency is counted (e.g., "residence permit" appears 300 times, and "handle" appears 250 times). The preprocessed text data is then vectorized into a bag-of-words model and fed into an LDA model for topic modeling analysis. The LDA model iteratively calculates a set of topic distribution probability data for each question-answer combination. For example, for the question-answer combination "How to apply for a residence permit" and "Residence permit application requirements," the Scikit-learn library's LDA model was used, with a set of topics K = 3 (e.g., "residence permit application," "household registration transfer," and "certificate documents"), 500 iterations, and a learning rate of 0.01. After inputting the bag-of-words model matrix, the model generated a topic distribution probability vector for each question-answer combination. For example, the topic distribution for the question-answer combination "How to apply for a residence permit" was: Topic 1 (residence permit application) 0.7, Topic 2 (household registration transfer) 0.2, and Topic 3 (certificate documents) 0.1; the topic distribution for the question-answer combination "What documents are required for household registration transfer" was: Topic 1 (residence permit application) 0.3, Topic 2 (household registration transfer) 0.6, and Topic 3 (certificate documents) 0.1. The output is stored as a 500*3 matrix, with each row corresponding to a question-answer combination and each column corresponding to the probability value of a topic. The output is stored in matrix format, with rows representing question-answer combinations, columns representing topics, and cell values representing the probability values of the topics.

[0103] Step S32: performing question-answer combination topic identification based on the government service question-answer combination topic distribution probability data to generate question-answer combination topic data;

[0104] In the embodiment of the present invention, topic identification processing is performed on each question-answer combination based on the topic distribution probability data generated by the LDA model. First, the dominant topic of each question-answer combination is determined based on the topic distribution probability. The dominant topic is determined based on the one with the largest probability value, and the calculation formula is: , is the probability of the i-th question-answer combination on the k-th topic. For example, the topic distribution for the question-answer combination "Residence Permit Application Conditions" is [0.7, 0.2, 0.1], with Topic 1 (Residence Permit Application) as the dominant topic; the topic distribution for the question-answer combination "Household Registration Transfer Process" is [0.2, 0.6, 0.2], with Topic 2 (Household Registration Transfer) as the dominant topic. For each topic, extract the top five terms with the highest probability as keywords. Use the TF-IDF algorithm to calculate the relevance between terms and topics using the following formula: in, is the frequency of word w in topic k, is the inverse document frequency of word w. For example, the keywords for Topic 1 (residence permit application) are "residence permit" (TF-IDF=0.8), "application" (TF-IDF=0.7), and "conditions" (TF-IDF=0.6). By calculating the keyword distribution of each topic, we further extract representative keywords for each topic. For example, the keywords for Topic 1 are "residence permit," "application," and "conditions." The generated question-answer combination topic data is stored in a structured table format, with each record containing a question-answer combination, the leading topic number, and the corresponding keyword set.

[0105] Step S33: performing topic distribution probability identification processing on subsets within clusters of question-answer combination clusters based on the government service question-answer combination matrix and the government service question-answer combination topic distribution probability data, and generating topic distribution probability data on subsets within clusters of question-answer combination clusters;

[0106] In this embodiment of the present invention, question-answer combination clustering and topic distribution probability analysis of subsets within clusters are performed based on the question-answer combination matrix and topic distribution probability data. A hierarchical clustering algorithm is used to perform cluster analysis on question-answer combinations. The input data is the question-answer combination matrix and its topic distribution probability. Cosine similarity is used to calculate the similarity between question-answer combinations. The formula is: ,in, and is the topic distribution vector for the i-th and j-th question-answer combination. Setting the similarity threshold to 0.6, we divide the 500 questions and answers into two clusters using agglomerative hierarchical clustering: cluster C1 (300 questions, with the dominant topic being "residence permit application") and cluster C2 (200 questions, with the dominant topic being "household registration transfer"). For the 300 question-answer combinations in cluster C1, we calculate the mean vector of the topic distribution: , the dominant topic is Topic 1 (Residence Permit Application, 0.65), and the secondary topic is Topic 2 (Household Registration Migration, 0.2). Similarly, the mean vector of cluster C2 is [0.2, 0.6, 0.2], and the dominant topic is Topic 2 (Household Registration Migration, 0.6). A cluster tree is constructed based on similarity and a threshold is set for cluster division. For each cluster, the subset topic distribution probability is further calculated for its internal question and answer combinations. The output of the cluster subset topic distribution probability data includes the cluster number, topic probability distribution, and dominant topic.

[0107] Step S34: performing sub-topic analysis and identification processing within the question-answer combination cluster based on the subset topic distribution probability data within the question-answer combination cluster to generate sub-topic data within the question-answer combination cluster;

[0108] In the embodiment of the present invention, the probability data of the subset topic distribution within the question-answer combination cluster is further analyzed to identify the subtopics within each cluster. First, subtopics with significant differences are extracted based on the probability data of the subset topic distribution within the cluster. Subtopics are identified by calculating the distribution differences between topics (such as KL divergence). The formula is: ,when If the divergence exceeds a certain threshold, Topic 1 and Topic 2 are considered significantly different and need to be divided into different subtopics. For subsets with significant topic differences, topic feature extraction methods (such as TF-IDF) are used to extract keywords to identify the semantic content of the subtopics. For example, for the question and answer text within cluster C1, the TextRank algorithm is used to extract subtopic keywords. Taking the question "What materials are required to apply for a residence permit?" as an example, after word segmentation, the word sequence ["residence permit", "application", "submission", "materials"] is generated. Through iterative calculation (damping coefficient = 0.85, number of iterations = 10), the keyword weights are: "materials" (1.3), "submission" (1.1), "residence permit" (1.0). The final output subtopic data within the cluster is stored in a structured format, including the cluster number, subtopic number, and keyword set.

[0109] Step S35: Perform hierarchical topic relationship analysis of the question-answer combination based on the question-answer combination topic data and the sub-topic data within the question-answer combination cluster to generate question-answer combination hierarchical topic relationship data.

[0110] In an embodiment of the present invention, a hierarchical topic relationship of a question-answer combination is constructed based on the topic data of the question-answer combination and the sub-topic data within the cluster. In a specific implementation, a hierarchical structure is first constructed according to the subordinate relationship between the topic and the sub-topic. For example, for the dominant topic "application for residence permit", its subordinate sub-topics include "application conditions", "application process", etc. When constructing the hierarchical relationship, a tree data structure is adopted, in which each node represents a topic or sub-topic, and the edge represents the subordinate relationship. Through relationship analysis, the association path between each question-answer combination and its corresponding topic is determined. For example, the path of "how to apply for a residence permit" is from "application for residence permit" to "application conditions". The hierarchical topic relationship data finally generated is stored in the form of a tree structure or a nested table, and each record contains the topic hierarchical path and question-answer combination association information.

[0111] Furthermore, step S33 includes the following steps:

[0112] Step S331: performing a question-answer combination similarity analysis on the government service question-answer combination matrix to generate government service question-answer combination similarity data; performing a government service question-answer combination cluster analysis based on the government service question-answer combination similarity data to generate government service question-answer combination cluster data;

[0113] Step S332: Map the government service question and answer combination topic distribution probability data to the government service question and answer combination cluster data to perform intra-cluster subset topic distribution probability identification processing to generate the question and answer combination intra-cluster subset topic distribution probability data.

[0114] In an embodiment of the present invention, when performing question-answer combination similarity analysis on a government service question-answer combination matrix, the matrix is normalized to ensure that the data in different dimensions of the matrix have the same dimensions. Then, the similarity between each pair of question-answer combinations in the matrix is calculated using cosine similarity or the Pearson Correlation Coefficient. Cosine similarity is used to calculate the semantic relevance between question-answer combinations, and the formula is: Taking question-answer combination A "How to apply for a passport" (vector 1, 0.3, 0.1) and question-answer combination B "What materials are needed to apply for a passport" (vector 0.9, 0.4, 0.2) as examples, we can calculate This shows that the semantics of the two are highly similar. Finally, the similarity values between all question and answer combinations are stored as a symmetric matrix to generate the similarity data of government service question and answer combinations. This data is used for subsequent clustering analysis. Based on the generated question and answer combination similarity data, hierarchical clustering or K-means clustering algorithm is used to cluster the question and answer combinations. For hierarchical clustering, a bottom-up merging method is used to gradually merge similar question and answer combinations according to the similarity values in the similarity matrix until the preset number of clusters is reached. An agglomerative hierarchical clustering algorithm is used with the cosine similarity matrix as input and the merging threshold is set to 0.85. Initially, each question and answer is an independent cluster, and clusters with similarities greater than the threshold are gradually merged. For example: Question and Answer A (0.95) and Question and Answer B are merged into cluster C1; Question and Answer C "Hong Kong and Macau Pass Renewal Process" (vector 0.2, 0.8, 0.1) and Question and Answer D "How to renew the Hong Kong and Macau Pass" (vector 0.3, 0.7, 0.1) are merged into cluster C2 with a similarity of 0.92. Finally, the cluster cluster data is generated, including the cluster number, member list and inter-cluster distance. The topic distribution probability data of the government service question and answer combination is used to map it with the question and answer combination cluster cluster data. For each question and answer combination in each cluster, the corresponding topic distribution probability is extracted, and the topic distribution probability of all question and answer combinations in the cluster is weighted averaged to generate the overall topic distribution probability within the cluster. For example, the topic distribution probability of each question and answer in cluster C1 is weighted averaged, and the weight is the similarity ranking of the question and answer in the cluster (the higher the ranking, the higher the weight, ranging from 0.1 to 1.0). The calculation formula is: .in, is the weight of the i-th question and answer (e.g., question and answer A has a weight of 1.0, question and answer B has a weight of 0.9), is the probability of the i-th question and answer in topic k. Assume that the average topic distribution of cluster C1 is: μ = [0.75 (topic 1), 0.2 (topic 2), 0.05 (topic 3)]. For cluster 1, which contains question-answer combinations A and B, the topic distribution probabilities are topic 1 (0.7), topic 2 (0.3) and topic 1 (0.6), topic 2 (0.4), respectively. Based on the overall topic distribution probability within the cluster, we further analyze the subset topic distribution characteristics. By setting a topic contribution threshold (e.g., 0.5), we filter out subsets of question-answer combinations with significant topic contributions within the cluster. For example, for cluster 1, if the distribution probability of topic 1 exceeds 0.5, we identify this topic as the dominant topic, and classify all question-answer combinations with a topic 1 probability above 0.5 as the topic 1 subset. Simultaneously, we statistically analyze the topic distribution probabilities within the subset to generate subset-level topic distribution probability data. The resulting subset topic distribution probability data for the question-answer combination cluster includes the cluster number, subset number, topic number, and their probability distribution. For example, cluster 1 contains the subset of topic 1, whose topic distribution probabilities are topic 1 (0.8) and topic 2 (0.2).

[0115] Furthermore, step S34 includes the following steps:

[0116] Step S341: performing topic feature analysis on the subset topic distribution probability data within the question-answer combination cluster to generate topic feature data of the subset topic within the question-answer combination cluster;

[0117] Step S342: reconstructing the topic of the subset within the question-answer combination cluster based on the topic feature data of the subset within the question-answer combination cluster to generate the topic reconstructed data of the subset within the question-answer combination cluster;

[0118] Step S343: performing sub-topic identification processing within the question-answer combination cluster based on the subset topic reconstruction data within the question-answer combination cluster to generate sub-topic data within the question-answer combination cluster.

[0119] In an embodiment of the present invention, a topic feature analysis of the subsets within the question-answer combination cluster is performed on the topic distribution probability data of the subsets within the question-answer combination cluster, and the topic distribution probability data of the subsets within the question-answer combination cluster is input, which includes the topic number of each subset and its distribution probability. For example, for subset 1, the topic distribution probability is topic 1 (0.6), topic 2 (0.3), and topic 3 (0.1). A statistical analysis is performed on the topic distribution probability data of each subset, including the calculation of the importance weight of the topic and the evaluation of the synergistic relationship between the topics. The importance weight is obtained through normalization processing, the synergistic relationship is calculated through the correlation matrix, and the mutual information (MI) is used to calculate the correlation between topic 1 and topic 2. The formula is: , where x=1 means the question and answer includes topic 1, and y=1 means it includes topic 2. Statistics show that there are 30 questions and answers that include both topics 1 and 2, p(1,1)=30 / 120=0.25, 60 questions and answers that only include topic 1, p(1,0=60 / 120=0.5, 24 questions and answers that only include topic 2, p(0,1)=24 / 120=0.2, and 6 questions and answers that do not include either topic, p(0,0)=6 / 120=0.05. The value of -0.34 and the mutual information is negative, indicating that topics 1 and 2 are weakly correlated and need to be processed independently. The generated subset topic feature data within the question-answer combination cluster contains structured information such as subset number, dominant topic, secondary topic, and topic synergy. Input the subset topic feature data within the question-answer combination cluster, and perform feature analysis on the dominant topic and secondary topic of each subset. For example, the dominant topic of subset 1 is topic 1, and the secondary topic is topic 2. The question-answer combination within the subset is optimized and reconstructed based on the weight distribution of the dominant topic and secondary topic. The reconstruction operation includes redistributing the topic weight of the question-answer combination and adjusting the topic distribution. Set the dominant topic weight enhancement coefficient α=1.3 and the secondary topic attenuation coefficient β=0.7. The changed topic 1 is 0.78, topic 2 is 0.21, and topic 3 remains unchanged. Then normalization is performed again to enhance the focus of the dominant topic. During the reconstruction process, feature optimization models, such as the Kullback-Leibler divergence, are introduced to calculate the change in topic distribution before and after reconstruction, and the optimization strategy is adjusted until the optimal equilibrium is achieved. The reconstructed topic distribution data is ultimately output. The generated topic reconstruction data for subsets within the Q&A cluster includes the subset number, the reconstructed topic distribution probability, and the adjusted weights. For example, the reconstructed data for subset 1 is Topic 1 (0.72), Topic 2 (0.19), and Topic 3 (0.09). The reconstructed topic data for subsets within the Q&A cluster is input. For each subset, the reconstructed topic distribution data is extracted and the dominant topic is identified based on a preset topic identification threshold (e.g., 0.6). The dominant topic after reconstruction is identified based on the topic distribution probability in the reconstructed data. For example, for subset 1, the weight of Topic 1 is 0.72, exceeding the threshold of 0.6, so Topic 1 is identified as the dominant subtopic of the subset. Furthermore, the feature contributions of secondary subtopics are calculated. For example, Topic 2 has a weight of 0.19 and is identified as a secondary subtopic. If the weight of a newly added topic is less than the secondary subtopic threshold (e.g., 0.1), it is marked as a background topic and removed from subsequent analysis. The generated subtopic data within the Q&A cluster includes the subset number, dominant subtopic number, secondary subtopic number, and background topic number. For example, the subtopic data for subset 1 is dominant subtopic 1, secondary subtopic 2, and background topic 3.

[0120] Furthermore, step S4 includes the following steps:

[0121] Step S41: Establishing an intelligent feedback mapping relationship of government service questions and answers through question-answer combination hierarchical topic relationship data to obtain a government service intelligent question-answering model;

[0122] Step S42: receiving user question data of instant government service;

[0123] Step S43: transmitting the instant government service user question data to the government service intelligent question and answer model for intelligent feedback processing of multiple answers to the government service, including: matching the user question data with the question and answer combination hierarchical topic relationship data through the government service intelligent question and answer model to generate government service multiple matching answer data; performing government service multiple matching answer confidence analysis based on the government service multiple matching answer data to generate government service multiple matching answer confidence data; using the government service multiple matching answer confidence data as a government service multiple answer ranking mechanism, and performing multiple answer intelligent ranking and confidence output on the government service multiple matching answer data through the government service multiple answer ranking mechanism to obtain government service multiple matching answer intelligent ranking data;

[0124] Step S44: Transmitting the government service multi-answer intelligent sorting data to the terminal to execute the government service intelligent question and answer feedback task.

[0125] In this embodiment of the present invention, the hierarchical topic relationship data of question-answer combinations is used as input, and a graph neural network (GNN) model is used to construct an intelligent feedback mapping relationship. First, the hierarchical topic relationship data of question-answer combinations is mapped into a graph structure, where nodes represent topics and edges represent the hierarchical relationships between topics. Edge weights are determined by the strength of the topic relationship (for example, the edge weight between topic A and topic B is 0.8).

[0126] A GNN model was used to extract features and model relationships within the graph structure. The GraphSAGE algorithm was then used to construct an intelligent question-answering model. Node features were input as TF-IDF vectors of topic keywords (e.g., [0.8, 0.3, 0.1] for T1), and edge features were weights. The training process was as follows: the node embeddings were initialized to 128 dimensions. A two-layer aggregation layer was used: the first layer aggregated features of neighboring nodes, and the second layer generated the final embedding. The objective function was a cross-entropy loss, the optimizer was Adam, with a learning rate of 0.001, and 200 training epochs. The model was fed with labeled question-answer pairs (e.g., the question "What materials are required to apply for a residence permit?" corresponds to the answer path T1→T1-1). The model then learned the mapping between question vectors and topic nodes. After training, the model outputted an intelligent question-answering model for government services. This model accepts user question vectors and outputs association probabilities for each topic node. For example, for the question "Residence permit application requirements," the model outputs an association probability of 0.95 with T1-1 and a probability of 0.1 with T1-2. Government service question data is received in real time via an API interface in the form of JSON objects. The system uses regular expressions to validate the input data to ensure field integrity and character validity. Once validated, the question text is extracted and normalized, including removing extra spaces, converting to lowercase, and removing stop words. The processed question text is stored in vector form, for example, using TF-IDF to represent the question as a vector. The real-time question data is matched with the hierarchical topic relationship data of the question and answer combination in the intelligent question-answering model. Using the BERT-based sentence embedding method, the user question and the hierarchical topic question are vectorized and cosine similarity is calculated. For example, the vector similarity between the user question vector and topic A is 0.98, and the vector similarity with topic B is 0.76. Topics with similarities above a preset threshold (e.g., 0.7) are considered preliminary matches and generate multiple matching answers for government services. The similarity scores of the multiple matching answers are normalized, for example, by using the Softmax function to convert the similarities into a confidence distribution. The matching results are a confidence of 0.85 for Topic A and a confidence of 0.15 for Topic B. The generated multi-match answer confidence data contains the matching answers and their corresponding confidences, for example, {"Topic A": 0.85,"Topic B": 0.15}. A sorting mechanism is constructed based on the confidence data. A descending sorting strategy is adopted to output answers with high confidence first, for example, the sorting result is that Topic A is greater than Topic B. Confidence annotations are added to the sorted answers to generate multi-answer intelligent sorting data for government services. The multi-answer intelligent sorting data is transmitted to the user terminal in standard JSON format via the HTTP protocol. The output data includes the sorting results and confidence. After receiving the data, the user terminal calls the display module to generate user-friendly feedback results, such as displaying the answer content and confidence in the form of cards.

[0127] This specification provides a machine learning intelligent question-answering system based on government services, which is used to perform the machine learning intelligent question-answering analysis method based on government services as described above. The machine learning intelligent question-answering system based on government services includes:

[0128] A data collection module, configured to deploy a multi-channel government service data collection engine, and collect government service data based on the multi-channel government service data collection engine to obtain government service data, wherein the government service data includes historical government service question data, historical government service answer data, and government service knowledge base data;

[0129] The question-answer combination standardization module is used to perform bidirectional word segmentation and parsing on government service data to generate government service word segmentation and parsing data; perform government service semantic backbone triple analysis based on government service word segmentation and parsing data to generate government service semantic backbone triple data; and design a combination matrix for government service question-answer association based on government service semantic backbone triple data to generate a government service question-answer combination matrix.

[0130] The question-answer combination topic identification module is used to perform topic distribution probability analysis of the government service question-answer combination matrix to generate government service question-answer combination topic distribution probability data; based on the government service question-answer combination topic distribution probability data, it performs hierarchical topic relationship analysis of the question-answer combination to generate hierarchical topic relationship data of the question-answer combination;

[0131] The intelligent question and answer feedback module is used to establish an intelligent feedback mapping relationship of government service questions and answers through the question and answer combination hierarchical subject relationship data to obtain a government service intelligent question and answer model; receive real-time government service user question data; transmit the real-time government service user question data to the government service intelligent question and answer model for government service multi-answer intelligent feedback processing to obtain government service multi-answer intelligent sorting data; transmit the government service multi-answer intelligent sorting data to the terminal to execute the government service intelligent question and answer feedback operation.

[0132] The present invention is therefore intended to be illustrative and non-restrictive in all respects, with the scope of the invention being defined by the appended claims rather than the foregoing description, and all changes that come within the meaning and range of equivalents of the application documents are intended to be embraced therein.

[0133] The foregoing description is intended only to provide specific embodiments of the present invention, which will enable those skilled in the art to understand and implement the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not intended to be limited to the embodiments shown herein, but is to be construed in the widest possible manner consistent with the principles and novel features disclosed herein.< / token>

Claims

1. A machine learning intelligent question-answering analysis method based on government services, characterized in that: The following steps are involved: Step S1: deploying a multi-channel government service data collection engine, and collecting government service data based on the multi-channel government service data collection engine to obtain government service data, wherein the government service data includes historical government service question data, historical government service answer data, and government service knowledge base data; Step S2: performing bidirectional word segmentation and parsing processing on the government service data to generate government service word segmentation and parsing data; performing government service semantic trunk triple analysis based on the government service word segmentation and parsing data to generate government service semantic trunk triple data; Design a combination matrix of government service question-answer associations based on the government service semantic backbone triple data to generate a government service question-answer combination matrix; Step S3: performing a topic distribution probability analysis of the question-answer combination on the government service question-answer combination matrix to generate topic distribution probability data of the government service question-answer combination; performing a hierarchical topic relationship analysis of the question-answer combination based on the topic distribution probability data of the government service question-answer combination to generate hierarchical topic relationship data of the question-answer combination; Wherein, step S3 includes: Step S31: Using the LDA topic model to perform topic distribution probability analysis on the government service question and answer combination matrix, and generate government service question and answer combination topic distribution probability data; Step S32: performing question-answer combination topic identification based on the government service question-answer combination topic distribution probability data to generate question-answer combination topic data; Step S33: performing question-answer combination similarity analysis on the government service question-answer combination matrix to generate government service question-answer combination similarity data; performing government service question-answer combination cluster analysis based on the government service question-answer combination similarity data to generate government service question-answer combination cluster data; mapping the government service question-answer combination topic distribution probability data to the government service question-answer combination cluster data to perform intra-cluster subset topic distribution probability identification processing to generate intra-cluster subset topic distribution probability data of the question-answer combination cluster; Step S34: performing a topic feature analysis on the subset topic distribution probability data within the question and answer combination cluster to generate topic feature data on the subset topic within the question and answer combination cluster; performing topic reconstruction processing on the subset topic within the question and answer combination cluster based on the topic feature data within the subset topic within the question and answer combination cluster to generate topic reconstruction data on the subset topic within the question and answer combination cluster; performing sub-topic identification processing on the subset topic within the question and answer combination cluster based on the topic reconstruction data within the subset topic within the question and answer combination cluster to generate sub-topic data within the question and answer combination cluster; Step S35: performing hierarchical topic relationship analysis of the question-answer combination based on the question-answer combination topic data and the sub-topic data within the question-answer combination cluster to generate question-answer combination hierarchical topic relationship data; Step S4: Establish an intelligent feedback mapping relationship of government service questions and answers through the question-answer combination hierarchical subject relationship data to obtain a government service intelligent question-answer model; receive instant government service user question data; transmit the instant government service user question data to the government service intelligent question-answer model for government service multi-answer intelligent feedback processing to obtain government service multi-answer intelligent sorting data; transmit the government service multi-answer intelligent sorting data to the terminal to execute the government service intelligent question-answer feedback operation.

2. The machine learning intelligent question-answering analysis method based on government services according to claim 1 is characterized in that: Step S1 includes the following steps: Step S11: Obtain the government service authorization API interface; Step S12: performing interface source address identification processing on the government service authorization API interface to obtain an identified government service authorization API interface; Step S13: Designing a government service data verification script; Step S14: deploying a multi-channel government service data collection engine based on the government service authorization API interface and the government service data verification script; Step S15: Collect government service data based on the multi-channel government service data collection engine to obtain government service data.

3. The machine learning intelligent question-answering analysis method based on government services according to claim 2 is characterized in that: The government service data verification script described in step S13 is used to perform government service text compliance verification and screening operations and government service text redundancy screening operations.

4. The machine learning intelligent question-answering analysis method based on government services according to claim 1 is characterized in that: Step S2 includes the following steps: Step S21: performing bidirectional word segmentation analysis on the government service data to generate government service word segmentation analysis data; Step S22: performing multi-scale word segmentation window analysis based on the government service word segmentation analysis data to generate multi-scale word segmentation window data; Step S23: using the multi-scale word segmentation window data to perform window word segmentation part-of-speech identification analysis on the government service word segmentation parsing data to generate government service word segmentation part-of-speech data; Step S24: performing government service part-of-speech entity relationship analysis based on the government service part-of-speech data to generate government service part-of-speech entity relationship data; Step S25: performing government service semantic trunk triple analysis based on the government service part-of-speech entity relationship data to generate government service semantic trunk triple data; Step S26: performing a government service question-answer correlation analysis based on the government service data to generate government service question-answer correlation data; Step S27: Design a combination matrix of government service question and answer associations for the government service semantic backbone triple data using the government service question and answer association data to generate a government service question and answer combination matrix.

5. The machine learning intelligent question-answering analysis method based on government services according to claim 4 is characterized in that: Step S25 includes the following steps: Step S251: Building a syntactic dependency tree of government services based on the preset BERT algorithm and government service word segmentation part-of-speech data to generate a syntactic dependency tree model of government services; Step S252: performing tree model dependency type optimization processing on the government service syntax dependency tree model to generate an optimized government service syntax dependency tree model, wherein the tree model dependency type optimization processing is to define and optimize the types and weights of dependency relationships based on the characteristics of the government service data; Step S253: using the optimized government service syntactic dependency tree model to perform government service syntactic dependency trunk feature analysis on the government service part-of-speech entity relationship data to generate government service syntactic dependency trunk feature data; Step S254: Perform government service semantic trunk triple analysis based on the government service syntactic dependency trunk feature data to generate government service semantic trunk triple data.

6. The machine learning intelligent question-answering analysis method based on government services according to claim 1 is characterized in that: Step S4 includes the following steps: Step S41: Establishing an intelligent feedback mapping relationship of government service questions and answers through question-answer combination hierarchical topic relationship data to obtain a government service intelligent question-answering model; Step S42: receiving user question data of instant government service; Step S43: transmitting the instant government service user question data to the government service intelligent question and answer model for intelligent feedback processing of multiple answers to the government service, including: matching the user question data with the question and answer combination hierarchical topic relationship data through the government service intelligent question and answer model to generate government service multiple matching answer data; performing government service multiple matching answer confidence analysis based on the government service multiple matching answer data to generate government service multiple matching answer confidence data; using the government service multiple matching answer confidence data as a government service multiple answer ranking mechanism, and performing multiple answer intelligent ranking and confidence output on the government service multiple matching answer data through the government service multiple answer ranking mechanism to obtain government service multiple matching answer intelligent ranking data; Step S44: Transmitting the government service multi-answer intelligent sorting data to the terminal to execute the government service intelligent question and answer feedback task.

7. A machine learning intelligent question-answering system based on government services, characterized in that: For executing the machine learning intelligent question-answering analysis method based on government services according to claim 1, the machine learning intelligent question-answering system based on government services comprises: A data collection module, configured to deploy a multi-channel government service data collection engine, and collect government service data based on the multi-channel government service data collection engine to obtain government service data, wherein the government service data includes historical government service question data, historical government service answer data, and government service knowledge base data; The question-answer combination standardization module is used to perform bidirectional word segmentation and parsing on government service data to generate government service word segmentation and parsing data; perform government service semantic backbone triple analysis based on government service word segmentation and parsing data to generate government service semantic backbone triple data; and design a combination matrix for government service question-answer association based on government service semantic backbone triple data to generate a government service question-answer combination matrix. The question-answer combination topic identification module is used to perform topic distribution probability analysis of the government service question-answer combination matrix to generate government service question-answer combination topic distribution probability data; based on the government service question-answer combination topic distribution probability data, it performs hierarchical topic relationship analysis of the question-answer combination to generate hierarchical topic relationship data of the question-answer combination; The intelligent question and answer feedback module is used to establish an intelligent feedback mapping relationship of government service questions and answers through the question and answer combination hierarchical subject relationship data to obtain a government service intelligent question and answer model; receive real-time government service user question data; transmit the real-time government service user question data to the government service intelligent question and answer model for government service multi-answer intelligent feedback processing to obtain government service multi-answer intelligent sorting data; transmit the government service multi-answer intelligent sorting data to the terminal to execute the government service intelligent question and answer feedback operation.

Images