Government affair question and answer method based on knowledge graph and large language model

By constructing a government affairs question-and-answer system based on knowledge graphs and large language models, the problems of high cost of manual responses and difficulty of intelligent assistants in handling complex questions in government affairs question-and-answer systems have been solved, achieving more accurate and comprehensive government affairs question-and-answer.

CN118939761APending Publication Date: 2024-11-12KUNSHAN XUAN LIFE INFORMATION TECH CO LTD
View PDF 0 Cites 5 Cited by

Patent Information

Application Number
CN202410520405.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-04-28
Publication Date
2024-11-12

AI Technical Summary

Technical Problem

Existing government Q&A systems suffer from high costs and low efficiency in manual responses, and intelligent assistants struggle to handle complex questions, lack in-depth reasoning, and may provide answers that are outdated.

Method used

This paper proposes a government affairs question-answering method based on knowledge graphs and large language models. Through data collection and sorting, a government affairs knowledge graph is constructed, and the knowledge triples are expanded into five-tuples, including preconditions and regional dimensions. The large language model is then used to generate the answer text.

Benefits of technology

It improves the accuracy and explainability of government affairs Q&A, and can effectively mine relevant knowledge to generate more accurate and comprehensive answers.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118939761A_ABST
    Figure CN118939761A_ABST
Patent Text Reader

Abstract

The invention relates to the technical field of language processing, and discloses a government affair question and answer method based on a knowledge graph and a large language model, which aims at a government affair knowledge graph, expands a knowledge triple into a quintuple and adds two dimensions of preconditions and regions, and compared with other methods, the government affair question and answer method based on the knowledge graph and the large language model has the advantages that the knowledge triple is expanded into the quintuple; according to the method, government affair knowledge of different preconditions and different regions can be well distinguished, accurate government affair knowledge is provided for the large language model by constructing the government affair knowledge graph, the accuracy and interpretability of large language model answering are improved, expansion is carried out along the upstream and downstream of initial knowledge, more related knowledge can be effectively mined, and the method is suitable for popularization and application. And the answering accuracy and comprehensiveness are improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of language processing technology, and specifically to a government affairs question and answer method based on a knowledge graph and a large language model. Background Art

[0002] In recent years, with the advancement of "Internet + Government Affairs" services, channels such as the 12345 citizen hotline, the mayor's mailbox, and public opinion message boards have become important communication bridges for government agencies to listen to public voices, understand public sentiment, and serve the people. However, manual replies have problems such as difficulty in responding immediately and consuming manpower costs. Building a government affairs question-and-answer system that automatically answers public demands has become an urgent need for the innovative development of social governance.

[0003] With the continuous breakthroughs in natural language processing and large language model technology, the field of intelligent question answering has received widespread attention. Intelligent question answering is an important application in the field of artificial intelligence. It can automatically reply to questions raised by users and help users quickly obtain the information or answers they need.

[0004] Existing government Q&A services generally use manual services, which requires manual customer service to have a certain amount of government knowledge reserves. Since government knowledge is regional, policies in different cities, districts, counties, and even towns may differ, so knowledge is not universal. Manual services not only have high labor costs, but also very low service efficiency. Some regions use intelligent assistants based on question-and-answer library matching to automatically reply to users' questions. Although efficiency has been improved, these intelligent assistants have difficulty dealing with complex problems in specific situations and cannot perform deep reasoning. Although the question-and-answer method based on the large language model has good reasoning ability, it may produce answers that are no longer valid and it is difficult to solve the "AI hallucination" problem. Summary of the invention

[0005] The purpose of the present invention is to provide a government question answering method based on knowledge graph and large language model to solve the problems raised in the above background technology.

[0006] To achieve the above object, the present invention provides the following technical solution: a government affairs question-answering method based on knowledge graph and large language model, the method comprising the following steps:

[0007] S1, data collection and sorting;

[0008] S2. Build a government affairs knowledge graph;

[0009] S3, extracting question text entities and linking them to the government knowledge graph;

[0010] S4, expand the knowledge subgraph along the upstream and downstream;

[0011] S5. Large language model text generation.

[0012] Optionally, S1 further includes: collecting unstructured data such as policy documents and regulatory documents related to people's livelihood demands, and structured data such as 12345 consulting work order details based on Internet platforms, government platforms, internal documents, etc. Policy documents or regulatory documents that are not in plain text format need to be standardized.

[0013] Optionally, the S2 further comprises:

[0014] Based on policy documents and regulatory documents, some data are manually annotated, entities are extracted through entity extraction models, and relationships are extracted through entity context and dependency syntax analysis to obtain government knowledge triples (h, r, t). Preconditions and regional information are extracted through rule templates, and the triples are expanded into five-tuples (h, r, t, p, d), where h represents the head entity, r represents the relationship, t represents the tail entity, p represents the prerequisite, and d represents the regional name. A government knowledge graph is constructed based on the government knowledge five-tuple.

[0015] The entity extraction model includes a text input layer, a word embedding layer, a feature extraction layer, a multi-head attention layer, and a feature classification layer.

[0016] Optionally, the S3 further includes:

[0017] For user question text, we extract question text elements including region name, entity, relationship, and premise through entity extraction model, relationship extraction model, and rule template. Then we use entity linking technology to link question entities to existing entities in the government knowledge graph constructed by S2 to extract target quintuples.

[0018] The regional subgraph of the knowledge graph is extracted through the regional name in the question text, and then the entities extracted from the question text are linked to the corresponding correct entity objects in the knowledge base to extract the set of candidate quintuples. The embedding vector of the premise condition in each candidate quintuple is generated by the pre-trained BERT model, and the similarity is calculated with the embedding vector of the premise condition extracted from the question text to obtain the first similarity. If one of the premise conditions is empty, the first similarity value is 1. The embedding vector of the relationship in each candidate quintuple is generated by the pre-trained BERT model, and the similarity is calculated with the embedding vector of the relationship extracted from the question text to obtain the second similarity. The quintuple with the largest average of the first similarity and the second similarity is selected as the target quintuple.

[0019] Optionally, the S4 further includes:

[0020] For the target quintuple, expand along the upstream direction of the head entity to obtain the preceding knowledge quintuple, and expand along the downstream direction of the tail entity to obtain the following knowledge quintuple, and finally obtain the knowledge subgraph;

[0021] The target quintuple is transformed into a sentence-granularity vector and a text embedding vector is obtained. The cosine similarity between the target quintuple and the question text embedding vector is calculated. If the similarity is greater than the preset threshold, it is added to the quintuple set together with the target quintuple. If the similarity is less than the preset threshold, the route is no longer expanded.

[0022] Optionally, the knowledge subgraph includes a set of several quintuples, which are input into a large language model as credible knowledge, and the sentence granularity transformation refers to filling each element in the quintuple into a sentence template.

[0023] Optionally, the S5 further includes:

[0024] Use the fine-tuned large language model as the question-answering reasoning model, fill the quintuple set and question text into the pre-built prompt template, and input the question-answering reasoning model to generate the answer text of the user's question;

[0025] Based on the external knowledge provided by the government knowledge graph, the big model is used, assisted by the endogenous text information within the model, to generate answer text for user questions, thereby enhancing the accuracy of the big model's knowledge questions and answers.

[0026] Optionally, the large language model is a deep learning model trained using a large-scale text corpus.

[0027] Compared with the prior art, the present invention provides a government affairs question-answering method based on knowledge graph and large language model, which has the following beneficial effects:

[0028] 1. This government question-answering method based on knowledge graph and large language model, for the government knowledge graph, expands the knowledge triples into five-tuples, and adds two dimensions of prerequisites and regions. Compared with other methods, the present invention can well distinguish the government knowledge of different prerequisites and different regions. By constructing a government knowledge graph, it provides accurate government knowledge for the large language model, thereby improving the accuracy and interpretability of the large language model's answers.

[0029] 2. This government question-answering method based on knowledge graph and large language model expands along the upstream and downstream of the initial knowledge, can effectively mine more relevant knowledge, and improve the accuracy and comprehensiveness of the answers. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 It is a schematic diagram of the process structure of the present invention;

[0031] Figure 2It is a schematic diagram of the overall architecture of the present invention. DETAILED DESCRIPTION

[0032] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0033] like Figure 1-Figure 2 As shown, the present invention provides a technical solution: a government question answering method based on knowledge graph and large language model, comprising the following steps:

[0034] S1, data collection and sorting;

[0035] Collect non-structured data such as policy documents and regulatory documents related to people's livelihood demands, and structured data such as 12345 consultation work order details through various means such as Internet platforms, government platforms, and internal documents. Policy documents or regulatory documents that are not in plain text format need to be standardized, for example, converting doc, docx, pdf, xlsx and other format files into txt text files.

[0036] S2. Build a government affairs knowledge graph;

[0037] Based on policy documents and regulatory documents, some data are manually annotated, entities are extracted through entity extraction models, and relationships are extracted through entity context and dependency syntax analysis to obtain government knowledge triples (h, r, t). Preconditions and regional information are extracted through rule templates, and the triples are expanded into five-tuples (h, r, t, p, d), where h represents the head entity, r represents the relationship, t represents the tail entity, p represents the prerequisite, and d represents the regional name. A government knowledge graph is constructed based on the government knowledge five-tuple.

[0038] The entity extraction model includes text input layer, word embedding layer, feature extraction layer, multi-head attention layer, and feature classification layer;

[0039] S21. Perform preprocessing operations such as text cleaning and word segmentation on the input text at the text input layer to obtain a text sequence; text cleaning can remove modal particles, punctuation marks, etc. through regular expressions.

[0040] For word segmentation, you can choose HanLP as the word segmentation tool. HanLP is a natural language processing toolkit that provides a series of Chinese text processing tools and models with high processing speed and good performance.

[0041] The text sequence is x={x1,x2,...,x n ,} form;

[0042] The text sequence is obtained by the BERT model of the word embedding layer to obtain word vectors. The BERT model is based on the pre-trained language model of Transformer. The BERT model learns the contextual representation of language by training a large amount of unlabeled text. Considering that government knowledge is based on vocabulary, the Chinese pre-trained model BERT-wwm based on the Whole Word Masking technology is used as the word vector generator for text.

[0043] The BERT model contains multiple Transformer encoders, each of which mainly includes a multi-head Attention module and a forward network. The BERT model uses each character as the original word vector, converts each character in the text into a one-dimensional vector by querying the word vector table, and inputs it into the model. The model outputs the semantic vector representation of the full text after integrating the information of each character.

[0044] The word vector is processed by the feature extraction layer to obtain the corresponding feature vector. The output of the word embedding layer has certain semantic information, but still carries invalid feature information and weakens the position information in the text. Therefore, the word vector generated by BERT needs to be processed by the feature extraction layer to capture important semantic features and reduce computational complexity; the feature extraction layer is composed of a bidirectional long short-term memory network BiLSTM. The BiLSTM structure includes two LSTM networks in opposite directions to obtain bidirectional semantic dependencies. There are three gate structures inside the hidden layer of LSTM, namely the input gate, output gate and forget gate, which jointly control the preservation and forgetting of historical information in the current state.

[0045] The calculation process of the input gate is as follows:

[0046] i t =σ(W i [h t-1 ,x t ]+b i )

[0047]

[0048] The calculation process of the output gate is as follows:

[0049] O t =σ(W o [h t-1 ,x t ]+b o )

[0050] The calculation process of the forget gate is as follows:

[0051] f t =σ(W f [h t-1 ,x t ]+b f )

[0052] The final result of LSTM includes long memory C t and short memory h t , the calculation process is as follows:

[0053]

[0054] h t =O t tanh(C t )

[0055] In the BiLSTM network, the forward LSTM processes the sequence in the forward direction to obtain the output, and the reverse LSTM processes the sequence in the reverse direction to obtain the output. The output of the two is concatenated to get the output The calculation formula is as follows:

[0056]

[0057] By stacking two LSTM layers, the BiLSTM network can effectively utilize the forward and backward dependencies in the sequence, thus overcoming the limitation of unidirectional information flow.

[0058] In the multi-head attention layer, the word-level features are reorganized into sentence-level feature vectors.

[0059] In the multi-head attention mechanism, the linear transformation matrices Q, K, and V perform n attention operations with multiple heads. Each head records the information at a certain position in the sequence. The process of a single head performing attention operation is as follows:

[0060]

[0061]

[0062] By integrating the outputs of all heads, we can get the semantic information of the complete sequence. The calculation formula is as follows:

[0063] Mult(Q,K,V)=W M [head1,head2,...,head n ]

[0064] The text sequence is annotated at the feature classification layer, and the entities contained in the text are identified based on the BIO sequence.

[0065] The feature classification layer uses the CRF model to score the possible annotation type of each position. The higher the score, the greater the possibility that the position annotation is valid. The loss function of the CRF model has two parameters, namely the emission score and the conversion score. The specific calculation process is as follows:

[0066]

[0067] S22, pairing the entities extracted in S21;

[0068] Relationship extraction is performed based on entity context. Within the range of N distance between each pair of entities, the context words are intersected to obtain candidate relationship words. The cosine similarity between each candidate relationship word vector and the overall context vector is calculated, and the candidate relationship words with similarity exceeding the preset threshold or with the highest similarity are selected as the final relationship words. The formal representation of this calculation process is as follows:

[0069] WordsList(h,t)=C(h)∩C(t)

[0070]

[0071] Where: C(·) represents the context word generation function, that is, the context words within the range of N of the input entity distance; V(·) represents the word vector generation function; Sim(·) represents the cosine similarity function, and θ is the preset similarity threshold.

[0072] The calculation formula for cosine similarity is:

[0073]

[0074] After calculation, the relationship between each entity pair can be extracted, such as the example: (pension, [factor, positive correlation], payment period).

[0075] Since the entity relationship triples extracted by the entity context-based method have a lot of redundancy and the entities have semantic missing, they are supplemented by a method based on dependency syntactic analysis.

[0076] First, the HanLP tool is used to implement word segmentation and part-of-speech tagging of Chinese text, and to perform dependency syntactic analysis, such as subject-predicate relationship, parallel relationship, core relationship, etc. If the subject or object is missing in the clause, it is completed based on the entire sentence.

[0077] Subject completion refers to obtaining a subject with complete semantics through the dependency syntactic tree. First, find the subject of the core relational word, and then find out whether there is a missing subject before the parallel relational word. If so, complete the subject of the core relational word in the sentence with the subject corresponding to the parallel relational word and add it to the original sentence. For example, "Personal pensions implement the individual account system, which is used to supplement basic pension insurance." The subject is missing before the parallel relational word "supplement" in "implement". After subject completion, the sentence is: "Personal pensions implement the individual account system, which is used to supplement basic pension insurance for personal pensions";

[0078] In addition, the rules are combined to ensure the integrity of the triple. For parallel subjects or parallel objects.

[0079] S23, manually formulate rule templates to obtain prerequisites;

[0080] Obtain the prerequisites by manually formulating rule templates and add them to the knowledge triples extracted from the corresponding context;

[0081] Some rule templates are as follows:

[0082]

[0083]

[0084] Policies may vary in different regions, so HanLP is used to identify words with the part of speech of place names in file names, file headers, and web page titles, and add them to the knowledge triples extracted from the file. If the file name, file header, and web page title do not contain place names, the place names with the highest number of occurrences are searched in the full text.

[0085] The triple (h, r, t) is expanded into a five-tuple (h, r, t, p, d) to construct a government knowledge graph, where h represents the head entity, r represents the relationship, t represents the tail entity, p represents the prerequisite, and d represents the region name.

[0086] S3, extracting question text entities and linking them to the government knowledge graph;

[0087] For user question text, the elements of the question text, including area name, entity, relationship, prerequisite, etc., are extracted through entity extraction model, relationship extraction model and rule template. Then, the question entity is linked to the existing entity in the government knowledge graph constructed by S2 using entity linking technology to extract the target quintuple.

[0088] Entity linking refers to generating candidate entities for each input entity based on existing entities in the knowledge graph through methods based on string matching and word vector matching, and then selecting the candidate entity that best matches the context information for linking.

[0089] First, the regional subgraph of the knowledge graph is extracted through the regional name in the question text. Then, the entities extracted from the question text are linked to the corresponding correct entity objects in the knowledge base, and the candidate quintuple set is extracted. The embedding vector of the premise in each candidate quintuple is generated through the pre-trained BERT model, and the similarity is calculated with the embedding vector of the premise extracted from the question text to obtain the first similarity. If one premise is empty, the first similarity value is 1. The embedding vector of the relationship in each candidate quintuple is generated through the pre-trained BERT model, and the similarity is calculated with the embedding vector of the relationship extracted from the question text to obtain the second similarity. The quintuple with the largest mean of the first similarity and the second similarity is selected as the target quintuple.

[0090] S4, expand the knowledge subgraph along the upstream and downstream;

[0091] For the target quintuple, expand along the upstream direction of the head entity to obtain the preceding knowledge quintuple, and expand along the downstream direction of the tail entity to obtain the following knowledge quintuple, and finally obtain the knowledge subgraph.

[0092] The target quintuple is transformed into a sentence-granularity vector and a text embedding vector is obtained. The cosine similarity between the target quintuple and the question text embedding vector is calculated. If the similarity is greater than the preset threshold, it is added to the quintuple set together with the target quintuple. If the similarity is less than the preset threshold, the route is no longer expanded.

[0093] The knowledge subgraph, i.e., a set of several quintuples, is input into the large language model as credible knowledge.

[0094] Sentence granularity transformation refers to filling each element in the quintuple into the sentence template.

[0095] S5, large language model text generation;

[0096] The fine-tuned large language model is used as the question-answering reasoning model, the quintuple set and question text are filled into the pre-built prompt template, and the question-answering reasoning model is input to generate the answer text of the user's question.

[0097] Based on the external knowledge provided by the government knowledge graph, the big model is used, assisted by the endogenous text information within the model, to generate answer text for user questions, thereby enhancing the accuracy of the big model's knowledge questions and answers.

[0098] A large language model refers to a deep learning model that is trained using a large-scale text corpus. It is large in scale, contains billions or even more parameters, and can interpret the meaning of language text and generate natural language text.

[0099] Fill the quintuple set and question text into the pre-built prompt template, replace the "{Knowledge}" part in the prompt template with the sentences after sentence granularity transformation of several quintuples in the knowledge subgraph, and replace the "{Question}" part in the prompt template with the question text.

[0100] Optional, prompt template is as follows: "Please answer the question based on the given hint knowledge. If there is no relevant answer, you can return "No answer yet". If there is, the answer is output in list form, and no reasoning process is required. \nThe hint knowledge is as follows: {Knowledge}\nThe question is as follows: {Question}"

[0101] The large language model is the open source ChatGLM2-6B model. The weight parameters of the large language model are adjusted based on LoRA. LoRA is a method for fine-tuning large pre-trained models with a small amount of labeled data. The effect of the fine-tuned model is close to full fine-tuning, which can fully improve the performance of the model on specific tasks.

[0102] First, the LoRA model parameters are split into two parts, including the pre-trained weights w∈R d*d And the fine-tuned incremental weight Δw∈R d*d , w is the frozen pre-trained weight, and Δw is the weight update generated during fine-tuning.

[0103] During the training process, for a fixed pre-training weight w, Δw is represented by two low-rank matrices A and B, and the output is as follows:

[0104] h=wx+BAx,A∈R r*d ,B∈R r*d

[0105] At the beginning of training, A is Gaussian initialized and B is zero initialized to avoid adding extra noise to the model. During the training process, the low-rank matrices A and B are trained, and only the low-rank matrix part is saved as the model weight.

[0106] Manually construct a large language model fine-tuning training set, perform Tokenizer processing on the training set, set training parameters, load the open source ChatGLM2-6B model, and save the model weight with the lowest loss value in the fine-tuning training. The training set includes prompt sentences, question sets, and answer sets.

[0107] The above generally describes the present invention in detail, but it is obvious to a person skilled in the art that some modifications or improvements can be made to the present invention. Therefore, modifications or improvements that do not depart from the spirit of the present invention are within the scope of protection of the present invention.

Claims

1. A government affairs question answering method based on knowledge graph and large language model, characterized by: The method comprises the following steps: S1, data collection and sorting; S2. Build a government affairs knowledge graph; S3, extracting question text entities and linking them to the government knowledge graph; S4, expand the knowledge subgraph along the upstream and downstream; S5. Large language model text generation.

2. According to claim 1, a government question-answering method based on knowledge graph and large language model is characterized by: The S1 further includes: collecting unstructured data such as policy documents and regulatory documents related to people's livelihood demands, and structured data such as 12345 consulting work order details based on various methods such as Internet platforms, government platforms, and internal documents. Policy documents or regulatory documents that are not in plain text format need to be standardized.

3. According to claim 1, a government question-answering method based on knowledge graph and large language model is characterized by: The S2 further comprises: Based on policy documents and regulatory documents, some data are manually annotated, entities are extracted through entity extraction models, and relationships are extracted through entity context and dependency syntax analysis to obtain government knowledge triples (h, r, t). Preconditions and regional information are extracted through rule templates, and the triples are expanded into five-tuples (h, r, t, p, d), where h represents the head entity, r represents the relationship, t represents the tail entity, p represents the prerequisite, and d represents the regional name. A government knowledge graph is constructed based on the government knowledge five-tuple. The entity extraction model includes a text input layer, a word embedding layer, a feature extraction layer, a multi-head attention layer, and a feature classification layer.

4. The government question-answering method based on knowledge graph and large language model according to claim 1, characterized in that: The S3 further comprises: For user question text, we extract question text elements including region name, entity, relationship, and premise through entity extraction model, relationship extraction model, and rule template. Then we use entity linking technology to link question entities to existing entities in the government knowledge graph constructed by S2 to extract target quintuples. The regional subgraph of the knowledge graph is extracted through the regional name in the question text, and then the entities extracted from the question text are linked to the corresponding correct entity objects in the knowledge base to extract the set of candidate quintuples. The embedding vector of the premise condition in each candidate quintuple is generated by the pre-trained BERT model, and the similarity is calculated with the embedding vector of the premise condition extracted from the question text to obtain the first similarity. If one of the premise conditions is empty, the first similarity value is 1. The embedding vector of the relationship in each candidate quintuple is generated by the pre-trained BERT model, and the similarity is calculated with the embedding vector of the relationship extracted from the question text to obtain the second similarity. The quintuple with the largest average of the first similarity and the second similarity is selected as the target quintuple.

5. According to claim 1, a government question-answering method based on knowledge graph and large language model is characterized by: The S4 further comprises: For the target quintuple, expand along the upstream direction of the head entity to obtain the preceding knowledge quintuple, and expand along the downstream direction of the tail entity to obtain the following knowledge quintuple, and finally obtain the knowledge subgraph; The target quintuple is transformed into a sentence-granularity vector and a text embedding vector is obtained. The cosine similarity between the target quintuple and the question text embedding vector is calculated. If the similarity is greater than the preset threshold, it is added to the quintuple set together with the target quintuple. If the similarity is less than the preset threshold, the route is no longer expanded.

6. The government question-answering method based on knowledge graph and large language model according to claim 5 is characterized by: The knowledge subgraph includes a set of several quintuples, which are input into the large language model as credible knowledge. The sentence granularity transformation refers to filling each element in the quintuple into the sentence template.

7. The government affairs question answering method based on knowledge graph and large language model according to claim 1 is characterized by: The S5 further comprises: Use the fine-tuned large language model as the question-answering reasoning model, fill the quintuple set and question text into the pre-built prompt template, and input the question-answering reasoning model to generate the answer text of the user's question; Based on the external knowledge provided by the government knowledge graph, the big model is used, assisted by the endogenous text information within the model, to generate answer text for user questions, thereby enhancing the accuracy of the big model's knowledge questions and answers.

8. The government affairs question answering method based on knowledge graph and large language model according to claim 7 is characterized by: The large language model is a deep learning model trained using a large-scale text corpus.

Citation Information

Cited By

  • Policy matching model construction method and system based on large language model, and medium

    CN120145067A

  • Government affair hotline service knowledge graph construction method and system based on large language model

    CN120354923A

  • Industrial production operation safety knowledge question-answering method and system based on large language model

    CN120744062A

  • Government affair policy question and answer method based on knowledge graph and related equipment

    CN120849569A

  • A government affair policy question and answer method based on a knowledge graph and related equipment

    CN120849569B