Generative question answering method and system based on knowledge graph and document retrieval integration
By integrating knowledge graphs with document retrieval systems and using PageRank algorithm and deep learning model, the problem of insufficient processing capabilities of unstructured data in the existing technology is solved, and more efficient information fusion and deep semantic analysis are achieved, improving The system's answer quality and information coverage capabilities are achieved.
Patent Information
- Application Number
- CN202411555244.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-04
- Publication Date
- 2024-11-29
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
In the prior art, the knowledge graph-based system has limited performance when processing unstructured data, and cannot effectively integrate information from multiple data sources, resulting in low quality of answers to complex queries; while the document-based system has deep semantics. There are technical bottlenecks in understanding and cross-document information integration, resulting in insufficient information coverage and insufficient reasoning capabilities.
Using an integrated method based on knowledge graph and document retrieval, by processing structured and unstructured data in parallel, combining PageRank algorithm and deep learning model, the generative question-and-answer system can extract information from different data sources and efficiently integrate it, improving the system's complexity Performance in inference tasks and in-depth semantic analysis.
It enhances the system's information coverage ability and inference and semantic understanding ability, improves the accuracy and relevance of responses to complex queries, and can better handle diversified queries and cross-domain reasoning tasks.
Smart Images

Figure CN119046448A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a generative question answering method and system, and in particular to a generative question answering method and system based on the integration of knowledge graph and document retrieval. Background Art
[0002] In the development of knowledge graph and document retrieval technology, knowledge graph systems based on structured data and document retrieval systems based on unstructured data are two relatively mature systems in current technology. Among them, knowledge graph is a structured data representation method that represents entities and their relationships through nodes and edges, and is widely used in semantic search, question-answering systems, and recommendation systems. Knowledge graph retrieval systems rely on entity, relationship, and attribute information in knowledge graphs, and find the most relevant results to the query through query semantic parsing, entity matching, path reasoning and other steps. The advantage of this type of system is that it can accurately retrieve and reason about predefined entities and relationships, thereby providing high-precision results. Document retrieval systems rely on unstructured data, such as text, articles, reports, etc., and mainly use keyword matching, vector search, semantic analysis and other technologies to find documents related to the query. Traditional methods based on TF-IDF, BM25 and semantic search technologies based on deep learning (such as BERT, etc.) are common methods in this field. These systems have significant advantages in processing large-scale text data and can effectively retrieve any text, but they usually lack the ability to understand and reason about deep semantic relationships in text.
[0003] In the prior art, the patent with application number CN202410719730.0 discloses a knowledge graph question-answering method and system based on the A2C algorithm and the GCN model, which mainly relies on the graph structure of the knowledge graph and its related algorithms for question-answering, which makes it perform well when processing structured data based on the graph. However, the limitation of this system is that it only relies on the knowledge graph and cannot process information from unstructured data (such as documents), so it has limited performance when responding to diversified queries. The patent with application number CN202410719068.9 discloses an intelligent customer service question-answering method based on knowledge graph and large language model. This intelligent customer service question-answering method mainly relies on the combination of industry knowledge base and large language model, first generates preliminary answers, and then improves the accuracy and professionalism of questions and answers through triple matching and secondary answer optimization. This method performs well in specific fields, but its limitation is that it has limited support for multiple fields and multiple data sources, especially when it is necessary to integrate information from multiple unstructured documents and knowledge graphs, and its ability to handle complex queries is limited. The patent with application number CN202410300934.0 discloses a RAG knowledge question-answering method and device based on fusion vectors and keyword retrieval. Although the patent can process unstructured data, it lacks the ability to understand and reason about deep semantic relationships in documents. In addition, the document retrieval process usually has difficulty in handling the task of cross-document information integration, and cannot achieve effective fusion and reasoning of multiple document information, which may affect the quality of answers to complex questions. Application number CN202410121013.8 is a question tuning method based on the large language model RAG framework. In terms of context processing, the patent performs multiple rounds of screening and optimization on the retrieved context to select the most relevant content for reasoning, but its context processing capabilities are limited and it is impossible to effectively integrate diverse information from different data sources. This application adopts Reciprocal Rank Fusion technology and PageRank algorithm, which can effectively retrieve context information from different sources and perform efficient fusion processing, surpassing the limitations of the prior art. In terms of the combination of retrieval and generation, the existing technology gradually optimizes question answers through iterative context retrieval and prompt word generation, focusing on the internal reasoning process of a single model and lacking the integration of information breadth and depth.
[0004] It can be seen that the existing knowledge graph-based retrieval system mainly relies on the entity relationships in the knowledge graph to answer user queries. This method has significant advantages in structured data retrieval, but it has technical disadvantages in the following aspects: Limited coverage: The information in the knowledge graph is usually structured and existing knowledge with limited coverage. It is difficult to process unstructured data beyond the scope of the knowledge graph, which makes it difficult to respond effectively to questions that are not included in the knowledge graph. Limited reasoning capabilities: Knowledge graph systems rely on pre-defined entities and relationships for reasoning and are unable to flexibly handle complex, cross-domain reasoning tasks, especially when faced with ambiguous or vague questions. The system's ability to respond is insufficient.
[0005] Traditional document retrieval systems return relevant documents by performing a hybrid search of keyword matching and semantics on a large number of documents. However, this approach also has shortcomings in the following technical aspects: The relevance of search results is not high: keyword matching search methods usually cannot accurately capture user intent, resulting in the returned documents containing keywords but having low relevance to the user's question. At the same time, semantic search may make misjudgments when dealing with complex semantic relationships, affecting the accuracy of search results; Lack of deep understanding: Document retrieval systems are usually unable to perform deep content understanding and cannot effectively handle semantic relationships across documents, especially when it is necessary to integrate information from multiple documents to form an accurate answer. The system performs poorly.
[0006] In summary, the technical problems in the fields of knowledge graph and document retrieval can be summarized as follows: Insufficient information coverage: Due to the limitations of structured data, knowledge graphs cannot meet the retrieval needs of large amounts of unstructured data; document retrieval has difficulty providing highly relevant results due to its lack of deep semantic understanding.
[0007] Insufficient reasoning and semantic understanding capabilities: The reasoning capabilities of knowledge graphs are limited to predefined relationships and entities, while document retrieval has technical bottlenecks in semantic analysis and cross-document information integration. Summary of the invention
[0008] The present invention aims to address the shortcomings of the single retrieval system in the prior art, and proposes a generative question-answering method and system based on the integration of knowledge graph and document retrieval, which combines the knowledge graph and the document retrieval system to improve the system's performance in complex reasoning tasks and deep semantic analysis, and enhances the reasoning and semantic understanding capabilities.
[0009] In order to achieve the above-mentioned object of the invention, the technical solution of the present invention is as follows: A generative question answering method based on the integration of knowledge graph and document retrieval includes the following steps: Step a: Receive the query input by the user and determine whether a search is required. If the query is simple or a clear answer already exists, the system directly generates a predefined response; otherwise, the search process is entered; Step b: Processing knowledge graph retrieval and document retrieval in parallel, and generating different candidate responses based on the respective retrieval structures of the two retrieval systems; Step c: Use a deep learning model to evaluate the quality of the generated candidate responses to obtain a final response; Step d: Return the final response to the user.
[0010] Furthermore, in step a, the large model fine-tuned by the Query parsing data is used to determine whether the input query needs to be retrieved, and the large model is trained by the Lora plus supervised fine-tuning method.
[0011] Furthermore, in step b, the knowledge graph retrieval includes: Entity extraction: The knowledge graph retrieval system extracts key entities and relationships from the query and finds the corresponding nodes in the knowledge graph; Node retrieval and similarity reasoning: Find the nodes most relevant to the query through the connection relationship of the nodes in the knowledge graph, and extract the corresponding triple set through the PageRank algorithm; Triple-based response generation: Generate responses based on the extracted triples and output the generated responses in the form of natural language.
[0012] Furthermore, in step b, document retrieval includes: Similar query generation: Generate multiple similar queries for the user's original query, use the large language model to process the query, and generate similar queries; Vector search and keyword search: The original query and similar queries are searched in the document collection through vector search and keyword search methods; Response generation: The ranking information of the two is integrated through the inverse sort fusion technology to find the most relevant documents for the corresponding query.
[0013] Furthermore, in step c, a deep learning model is used to perform quality assessment on the generated candidate responses to obtain a final response, including: Train the deep learning model so that it can evaluate and score the newly generated responses. Use multiple different deep learning models to perform weighted averaging or voting on the evaluation scores of the same candidate response, and select the response with the highest score as the final output. Alternatively, continuously update the training data and fine-tune the model based on user feedback and annotated data to improve the evaluation performance.
[0014] Furthermore, the training process of the deep learning model includes: Data collection and annotation: Collect enough candidate responses and corresponding quality evaluation indicators; Deep learning model selection: Select a pre-trained model or other Transformer-based model for fine-tuning.
[0015] Data processing and feature extraction: Take candidate responses and questions / context as input, use appropriate feature representations to represent the quality characteristics of each candidate response, then compare each candidate response with the reference answer or context, and extract semantic, grammatical, and logical features through a deep model; Supervised learning and fine-tuning: Train the deep model to output a quality score, enabling it to evaluate and score newly generated responses; Model validation and testing: Use validation data to evaluate the model’s ability to score the quality of different types of responses and adjust model parameters. Use actual test sets to observe the model’s performance in actual scenarios to ensure that it can correctly evaluate responses.
[0016] The present invention also proposes a generative question-answering system based on the integration of knowledge graph and document retrieval, comprising: Query parsing module, which receives the user's natural language query and determines whether the query needs further retrieval; A retrieval judgment execution module, which directly generates a predefined response and returns it to the user when the query is very simple or a clear answer already exists; otherwise, it enters the knowledge graph and document retrieval parallel retrieval process; A parallel retrieval system, wherein the parallel retrieval system uses a knowledge graph retrieval system and a document retrieval system to process retrieval in parallel and generates different candidate responses based on the retrieval structures of the two retrieval systems; A response evaluation module, wherein the response evaluation module performs a quality evaluation on the generated candidate responses through a deep learning model to obtain a final response; An output module feeds back the final response to the user.
[0017] Furthermore, the query parsing module determines whether the input query needs to be retrieved through a knowledge graph or a document retrieval system based on a large model fine-tuned through query parsing data.
[0018] Furthermore, the knowledge graph retrieval system extracts key entities and relationships from the query, searches for corresponding nodes in the knowledge graph, and searches for nodes most relevant to the query through the connection relationships between the nodes in the knowledge graph. It then extracts the corresponding triple set through the PageRank algorithm, generates a response based on the extracted triples, and finally outputs it in the form of natural language.
[0019] Furthermore, the document retrieval system generates multiple similar queries for the user's original query; retrieves the original query and similar queries in the document collection through vector search and keyword search methods; and then fuses the ranking information of the two through the inverse sort fusion technology to find the document most relevant to the corresponding query.
[0020] In summary, the present invention has the following advantages: 1. The present invention improves the system's performance in complex reasoning tasks and deep semantic analysis by processing the structured data in the knowledge graph and the unstructured data in the document in parallel, thereby enhancing the reasoning and semantic understanding capabilities.
[0021] 2. The present invention combines the knowledge graph and the document retrieval system to make up for the shortcomings of a single system in information coverage, expand the scope of information coverage, and ensure a wide range of response capabilities to user queries.
[0022] 3. The present invention uses deep learning technology to reduce dependence on a single system, reduce the risk of hallucinations caused by large models, improve the accuracy of retrieval and response, and optimize the overall accuracy of the system.
[0023] 4. The present invention can extract information from different angles and data sources by integrating the knowledge graph retrieval system and the document retrieval system, thereby comprehensively generating a more accurate response.
[0024] 5. The present invention utilizes the inverse sorting fusion technology to fuse the retrieval results in the document retrieval system, which significantly improves the relevance and accuracy of the retrieval results.
[0025] 6. The present invention uses the PageRank algorithm in knowledge graph retrieval, which can quickly find subsets related to the query in complex large-scale knowledge graphs, thereby realizing multi-hop reasoning and integration of knowledge, significantly improving efficiency and accuracy.
[0026] 7. In document retrieval, the present invention can more flexibly understand and retrieve the user's query content by generating multiple similar queries and performing vectorization processing, thereby enhancing the adaptability to different query scenarios. BRIEF DESCRIPTION OF THE DRAWINGS
[0027] Figure 1 A schematic diagram of an embodiment of the present invention. DETAILED DESCRIPTION
[0028] In order to explain the present invention more clearly, the present invention is further described below in conjunction with preferred embodiments and drawings. It should be understood by those skilled in the art that the content described below is illustrative rather than restrictive, and should not be used to limit the scope of protection of the present invention.
[0029] Example 1 This embodiment provides a generative question answering method based on the integration of knowledge graph and document retrieval, such as Figure 1 As shown, the following steps are included: Step 1: Receive query: Receive the query input by the user and determine whether a search is required; Specifically, this step is based on the large model fine-tuned by query parsing data (large model and artificial synthesis), to determine whether the input query needs to be retrieved through the knowledge graph or document retrieval system. If the query is very simple, or there is a clear answer, the system directly generates a predefined response; otherwise, it enters the retrieval process.
[0030] In this step, since the large model only needs to complete a simple classification task, the Lora plus supervised fine-tuning method can be used for training. This fine-tuning method can enable the model to master the ability to distinguish whether retrieval is needed with low computational cost and small data required.
[0031] Step 2: Knowledge graph retrieval and document retrieval: Process knowledge graph retrieval and document retrieval in parallel, perform entity recognition and reasoning through the knowledge graph retrieval system, and at the same time, find relevant documents in the document library through similar query generation and vector search methods, and generate different candidate responses based on the retrieval structure of each system.
[0032] Specifically, knowledge graph retrieval includes: S1. Entity extraction: The knowledge graph retrieval system extracts key entities and relationships from the query and searches for corresponding nodes in the knowledge graph; S2. Node retrieval and similarity reasoning: Find the nodes most relevant to the query through the connection relationship of the nodes in the knowledge graph, and extract the corresponding triple set through the PageRank algorithm; S3. Triple-based response generation: The system generates responses based on the extracted triples (entity-relationship-entity), and the generated responses are then output in the form of natural language.
[0033] Among them, the specific description of the PageRank algorithm is as follows: The PageRank algorithm used in the present invention starts from a specific seed node (key entity in the query) and calculates the importance of the nodes in the graph. In this way, the most relevant subgraph can be found from a large number of triples, thereby extracting more relevant knowledge. After using the PageRank algorithm, the triples are scored according to the importance of the nodes, and the triples with higher scores are considered to be more relevant to the query. Finally, the triple set with the highest score is extracted. In this way, it is possible to quickly find a subset related to the query in a complex large-scale knowledge graph, thereby realizing multi-hop reasoning and integration of knowledge. Unlike traditional step-by-step retrieval, PageRank can complete multi-hop reasoning in one go, significantly improving efficiency and accuracy.
[0034] Specifically, document retrieval includes: S1. Similar query generation: Generate multiple similar queries for the user's original query; In this embodiment, a large language model is used to process the query and generate similar queries.
[0035] S2. Vector search and keyword search: The original query and similar queries are searched in the document collection through vector search and keyword search methods; S3. Response generation: Vector search determines relevance by calculating the similarity between the query and the document; keyword search selects relevant documents based on text matching, and then fuses the ranking information of the two through the inverse sort fusion technology to find the most relevant document to the corresponding query.
[0036] This step generates multiple similar queries and performs vectorized processing, which can more flexibly understand and retrieve the user's query content and enhance the adaptability to different query scenarios.
[0037] Step 3: Response evaluation: Based on the response results of different retrieval systems, the system evaluates the final response.
[0038] This step uses a deep learning model to score the generated responses. The model considers multiple dimensions of indicators, including accuracy, coherence, completeness, etc., and then selects the response with the highest score as the final output result. Using a deep learning model to evaluate the quality of each candidate response can ensure that the response output to the user meets high standards in terms of accuracy, coherence, and completeness.
[0039] The specific implementation steps of using deep learning models for response evaluation include: weighted averaging or voting the evaluation scores of multiple different deep learning models (such as BERT, T5) on the same candidate response, and selecting the response with the highest score as the final output. It is also possible to continuously update training data and fine-tune models based on user feedback and annotated data to improve evaluation performance.
[0040] In this embodiment, the training method of the deep learning model is as follows: Data collection and annotation: Collect enough candidate responses and corresponding quality evaluation indicators (such as accuracy, coherence, completeness, etc.). These evaluation indicators can be manually annotated or extracted from existing data sets.
[0041] Deep learning model selection: Choose a pre-trained model, such as BERT, T5, or other Transformer-based models for fine-tuning. BERT is good at understanding the semantics of text and is suitable for evaluating coherence and accuracy. T5 is good at both text generation and quality assessment tasks, and can learn to score through fine-tuning.
[0042] Data processing and feature extraction: Take candidate responses and questions / context as input, use appropriate feature representations (such as BERT's [CLS] tag vector or T5's output encoding) to represent the quality characteristics of each candidate response, then compare each candidate response with the reference answer or context, and extract semantic, grammatical, and logical features through a deep model.
[0043] Supervised learning and fine-tuning: Train the deep model to output a quality score (rating). This process is similar to a regression problem or a multi-classification problem, and the goal is to maximize the accuracy of the model's assessment of the response quality. If it is a regression problem, the MSE (mean square error) loss function is used to optimize the model's scoring. If it is a classification problem, the quality score is discretized into several levels (such as 1-5 points), and the cross entropy loss is used to optimize the model. Use the labeled training data to fine-tune the model so that it can evaluate and score the newly generated responses.
[0044] Model validation and testing: Use unseen validation data to evaluate the model’s ability to score the quality of different types of responses and adjust the model’s parameters. Use the actual test set to observe the model’s performance in actual scenarios to ensure that it can correctly evaluate responses.
[0045] Step 4: Output response: The final response is returned to the user through the output module.
[0046] This step includes two parts: one is result display: the final response result is presented to the user in text form through the output module. The other is log recording: recording user queries and system responses for subsequent analysis and model optimization.
[0047] The present invention combines the dual retrieval methods of structured knowledge and unstructured documents, and can simultaneously process diverse queries, whether they are queries based on entity relationships or queries involving a large number of text documents, and can provide high-quality responses. By combining knowledge graphs and document retrieval, it can process various types of data sources, and by generating multiple similar queries and performing vectorized processing, it can more flexibly understand and retrieve the user's query content, enhancing the adaptability to different query scenarios.
[0048] Example 2 The present invention provides a generative question-answering system based on the integration of knowledge graph and document retrieval. The system integrates a knowledge graph retrieval system and a document retrieval system, and realizes efficient processing and response generation of user queries through complementary advantages.
[0049] Specifically, the generative question answering system includes: a query parsing module, a retrieval judgment execution module, a parallel retrieval system, a response evaluation module and an output module.
[0050] Among them, the Query parsing module is the entry module of the generative question-answering system, which is responsible for receiving the user's natural language query and determining whether the query requires further retrieval.
[0051] The retrieval judgment execution module directly generates a predefined response and returns it to the user when the query is very simple or a clear answer already exists; otherwise, it enters the parallel retrieval process of knowledge graph and document retrieval.
[0052] The parallel retrieval system uses the knowledge graph retrieval system and the document retrieval system to process retrieval in parallel, and generates different candidate responses based on the respective retrieval structures of the two retrieval systems.
[0053] Among them, the knowledge graph retrieval system uses the knowledge graph for reasoning and retrieval, and is suitable for queries containing entities and relationships. Specifically, the knowledge graph retrieval system extracts key entities and relationships from the query, searches for corresponding nodes in the knowledge graph, and searches for the nodes most relevant to the query through the connection relationship between the nodes in the knowledge graph. It then extracts the corresponding triple set through the PageRank algorithm, generates a response based on the extracted triples (entity-relationship-entity), and finally outputs the generated response in the form of natural language.
[0054] The document retrieval system mainly processes queries based on unstructured data (such as text documents) and uses hybrid retrieval technology to extract information from massive documents. Specifically, the document retrieval system uses a large prediction model to process the user's original query and generates multiple similar queries, so that the original query and similar queries are retrieved in the document collection through vector search and keyword search methods. The vector search determines the relevance by calculating the similarity between the query and the document, and the keyword search selects relevant documents based on the text matching degree, and then fuses the ranking information of the two through the reciprocal ranking fusion technology (RRF) to find the most relevant documents to the corresponding query.
[0055] In this embodiment, through the integration of the knowledge graph retrieval system and the document retrieval system, the generative question-answering system can extract information from different angles and data sources and comprehensively generate more accurate responses.
[0056] The response evaluation module uses a deep learning model (BERT, T5, etc.) to evaluate the quality of candidate responses generated by different retrieval systems to ensure their correctness and relevance, and then selects the response with the highest score as the final output result. Specifically, the method of using a deep learning model to evaluate responses is the same as the specific implementation steps of evaluating responses in step 3 of Example 1.
[0057] The output module is responsible for displaying the final generated response to the user. Specifically, the output module presents the final response result to the user in text form and records the user query and system response for subsequent analysis and model optimization.
[0058] In this embodiment, the function of each functional module corresponds to each step in Embodiment 1, and the undescribed parts will not be repeated.
[0059] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any form. Any simple modification or equivalent change made to the above embodiment based on the technical essence of the present invention shall fall within the protection scope of the present invention.
Claims
1. A generative question answering method based on the integration of knowledge graph and document retrieval, It is characterized in that The steps include: Step a: Receive the query input by the user and determine whether a search is required. If the query is simple or a clear answer already exists, the system directly generates a predefined response; otherwise, the search process is entered; Step b: Processing knowledge graph retrieval and document retrieval in parallel, and generating different candidate responses based on the respective retrieval structures of the two retrieval systems; Step c: Use a deep learning model to evaluate the quality of the generated candidate responses to obtain a final response; Step d: Return the final response to the user.
2. A generative question answering method based on the integration of knowledge graph and document retrieval as claimed in claim 1, It is characterized in that In step a, the large model fine-tuned by the query parsing data is used to determine whether the input query needs to be retrieved. The large model is trained using the Lora plus supervised fine-tuning method.
3. A generative question answering method based on the integration of knowledge graph and document retrieval as claimed in claim 1, It is characterized in that In step b, knowledge graph retrieval includes: Entity extraction: The knowledge graph retrieval system extracts key entities and relationships from the query and finds the corresponding nodes in the knowledge graph; Node retrieval and similarity reasoning: Find the nodes most relevant to the query through the connection relationship of the nodes in the knowledge graph, and extract the corresponding triple set through the PageRank algorithm; Triple-based response generation: Generate responses based on the extracted triples and output the generated responses in the form of natural language.
4. A generative question answering method based on integration of knowledge graph and document retrieval as claimed in claim 1, It is characterized in that In step b, document retrieval includes: Similar query generation: Generate multiple similar queries for the user's original query, use the large language model to process the query, and generate similar queries; Vector search and keyword search: The original query and similar queries are searched in the document collection through vector search and keyword search methods; Response generation: The ranking information of the two is integrated through the inverse sort fusion technology to find the most relevant documents for the corresponding query.
5. A generative question answering method based on integration of knowledge graph and document retrieval as claimed in claim 1, It is characterized in that In step c, a deep learning model is used to evaluate the quality of the generated candidate responses to obtain the final response, including: Train the deep learning model so that it can evaluate and score the newly generated responses. Use multiple different deep learning models to perform weighted averaging or voting on the evaluation scores of the same candidate response, and select the response with the highest score as the final output. Alternatively, continuously update the training data and fine-tune the model based on user feedback and annotated data to improve the evaluation performance.
6. A generative question answering method based on the integration of knowledge graph and document retrieval as claimed in claim 1 or 5, It is characterized in that The training process of a deep learning model includes: Data collection and annotation: Collect enough candidate responses and corresponding quality evaluation indicators; Deep learning model selection: select a pre-trained model or other Transformer-based model for fine-tuning; Data processing and feature extraction: Take candidate responses and questions / context as input, use appropriate feature representations to represent the quality characteristics of each candidate response, then compare each candidate response with the reference answer or context, and extract semantic, grammatical, and logical features through a deep model; Supervised learning and fine-tuning: Train the deep model to output a quality score, enabling it to evaluate and score newly generated responses; Model validation and testing: Use validation data to evaluate the model’s ability to score the quality of different types of responses and adjust model parameters. Use actual test sets to observe the model’s performance in actual scenarios to ensure that it can correctly evaluate responses.
7. A generative question answering system based on the integration of knowledge graph and document retrieval, It is characterized in that include: Query parsing module, which receives the user's natural language query and determines whether the query needs further retrieval; A retrieval judgment execution module, which directly generates a predefined response and returns it to the user when the query is very simple or a clear answer already exists; otherwise, it enters the knowledge graph and document retrieval parallel retrieval process; A parallel retrieval system, wherein the parallel retrieval system uses a knowledge graph retrieval system and a document retrieval system to process retrieval in parallel and generates different candidate responses based on the retrieval structures of the two retrieval systems; A response evaluation module, wherein the response evaluation module performs a quality evaluation on the generated candidate responses through a deep learning model to obtain a final response; An output module feeds back the final response to the user.
8. A generative question answering system based on the integration of knowledge graph and document retrieval as claimed in claim 7, It is characterized in that The query parsing module determines whether the input query needs to be retrieved through the knowledge graph or document retrieval system based on the large model fine-tuned by the query parsing data.
9. A generative question answering system based on the integration of knowledge graph and document retrieval as claimed in claim 7, It is characterized in that The knowledge graph retrieval system extracts key entities and relationships from the query, searches for corresponding nodes in the knowledge graph, and searches for nodes most relevant to the query through the connection relationships between the nodes in the knowledge graph. It then extracts the corresponding triple set through the PageRank algorithm, generates a response based on the extracted triples, and finally outputs it in the form of natural language.
10. A generative question answering system based on the integration of knowledge graph and document retrieval as claimed in claim 7, It is characterized in that The document retrieval system generates multiple similar queries for the user's original query; retrieves the original query and similar queries in a document collection through vector search and keyword search methods; and then fuses the ranking information of the two through a reciprocal sorting fusion technology to find the document most relevant to the corresponding query.
Citation Information
Patent Citations
Word sense based local file searching method
CN101059806A
Multi-information-source intelligent education question and answer method based on multi-relational graph convolutional neural network
CN114168704A
Knowledge question and answer method based on large language model and knowledge graph
CN117874180A
Knowledge intensive question reasoning and generating method based on LLM
CN118798367A
Leveraging entity relations to discover answers using a knowledge graph
US20210012217A1
Cited By
Coal industry large model retrieval enhancement generation method and system based on knowledge graph
CN119988600A
Intelligent retrieval method and system for operation and maintenance service of station-city complex based on large model
CN119988692A
Knowledge graph generation method and system based on RAG deduction
CN120633809A
Question and answer method and device, computer readable storage medium and electronic equipment
CN120910195A
Retrieval system for bidding law document large language model
CN120994814A
