Method and system for generating enhanced knowledge questions and answers for mixed retrieval of heterogeneous database

Abstract

The invention discloses a method and system for generating enhanced knowledge questions and answers for heterogeneous database mixed retrieval, and the method comprises the following steps: (1) processing text data, and constructing a heterogeneous database; (2) receiving user query, and performing mixed retrieval in the heterogeneous database; the retrieval results are merged, and documents and triple information related to user query are obtained; (3) reordering the retrieved documents and triple information by using a reordering model, and selecting K retrieval results with the highest relevancy with user query; and (4) combining user query and retrieval results, constructing cue words, inputting the cue words into the large language model, and generating coherent and accurate natural language answers. By means of the method and device, more comprehensive and accurate information retrieval can be achieved, and the accuracy and reliability of knowledge question and answer results are improved.

Description

Technical Field

[0001] The present invention belongs to the field of natural language processing, and in particular, relates to a knowledge question answering method and system for hybrid retrieval generation enhancement of heterogeneous databases. Background Art

[0002] Traditional knowledge question answering methods mainly rely on structured data and static knowledge bases to parse user queries through predefined rules and templates. These systems usually adopt a retrieval-based strategy to find relevant answers from a pre-built knowledge base. However, traditional methods often face many challenges when dealing with complex natural language queries, such as insufficient information accuracy, inability to cope with changes in dynamic knowledge, and poor adaptability to new domain knowledge, resulting in poor user experience.

[0003] With the rapid development of artificial intelligence technology, large language models have made breakthrough progress in the field of natural language processing. Large language models are pre-trained with a large amount of text data to learn rich language patterns and knowledge information, so that they can generate fluent and natural answers and show good versatility and adaptability in various language tasks. With its powerful text generation and context understanding capabilities, large language models have been widely used in many fields such as knowledge question and answer, educational guidance, content generation, etc., greatly improving the intelligence of user interaction.

[0004] However, large language models also face some challenges and limitations. First, large language models may generate wrong or even completely fictitious answers when generating responses, which is the famous large language model hallucination problem. Second, large language models acquire general language capabilities through pre-training, but do not have knowledge in professional fields, so they cannot give accurate answers to certain professional questions. In addition, the knowledge base of large language models is static, and the models are usually loaded into the system after offline pre-training, so the information generated by large language models is not timely enough.

[0005] In order to solve the above problems, retrieval-enhanced generation technology has been proposed and widely studied and applied. Retrieval-enhanced generation combines the traditional knowledge retrieval mechanism with the generation capability of large language models. It retrieves the external knowledge base before generating answers, extracts query-related information from the knowledge base, and then inputs this information as context into the large language model to generate more accurate answers. Retrieval-enhanced generation technology can not only rely on the language knowledge learned during training, but also dynamically obtain the latest and most relevant external information to generate more accurate and rich answers, reducing the illusion of large language models.

[0006] Retrieval techniques are mainly divided into two categories: dense retrieval and sparse retrieval. Dense retrieval uses a pre-trained deep learning model to encode queries and documents into high-dimensional vectors, and retrieves relevant information by calculating the similarity between these vectors. The advantage of dense retrieval is its powerful semantic understanding ability, which can capture the deep semantic relationship between queries and documents. Even if the document does not contain the exact words in the query, it can find relevant content based on semantic similarity. Therefore, dense retrieval performs well in natural language processing tasks, especially for solving word synonyms, implicit semantics, and complex natural language queries. Sparse retrieval is a traditional information retrieval method, and the most typical example is keyword retrieval based on inverted index. In sparse retrieval, queries and documents are retrieved by keyword matching, and the retrieval results are based on specific words that appear in the document. The advantage of sparse retrieval is that it is fast and has good results for exact matching, but when faced with complex or ambiguous queries, it is easily limited by the surface form of the vocabulary and has difficulty capturing semantic relevance.

[0007] The mainstream method in the existing retrieval enhancement generation retrieval link is vector retrieval based on vector database. Although vector retrieval has powerful semantic processing capabilities, it still has some limitations. First of all, an obvious disadvantage of vector retrieval is the lack of accurate query capability. Since it relies on semantic similarity to match queries and documents, rather than directly matching specific keywords or phrases, the relevance of retrieval results is low in short text search or exact matching scenarios. In addition, when faced with complex reasoning problems or information that requires multi-hop associations, vector retrieval has difficulty in mining deep semantic associations, limiting its performance in complex reasoning tasks. Summary of the invention

[0008] The present invention provides a knowledge question answering method and system for generating enhanced hybrid retrieval of heterogeneous databases, which can improve the accuracy and reliability of knowledge question answering results.

[0009] A knowledge question answering method for generating enhanced hybrid retrieval of heterogeneous databases includes the following steps:

[0010] (1) Process text data and build heterogeneous databases;

[0011] (2) Receive user queries and perform hybrid retrieval in heterogeneous databases; merge retrieval results to obtain documents and triple information related to the user query;

[0012] (3) Use the re-ranking model to re-rank the retrieved documents and triple information, and select the K retrieval results with the highest relevance to the user query;

[0013] (4) Combine user queries with search results, construct prompt words and input them into a large language model to generate coherent and accurate natural language answers.

[0014] The method proposed in the present invention first constructs a heterogeneous database composed of a document database, a vector database and a graph database, providing a flexible data model and query method. Subsequently, in the hybrid retrieval stage, keyword retrieval, vector retrieval and knowledge graph retrieval are performed on the heterogeneous database based on the user query to comprehensively capture information related to the user query. Through the combination of heterogeneous databases and hybrid retrieval methods, the present invention can fully utilize the advantages of various retrieval methods, making up for the shortcomings of a single retrieval method in accurate query and deep understanding of related information, thereby achieving more comprehensive and accurate information acquisition and improving the accuracy and reliability of knowledge question and answer results.

[0015] In step (1), the text data is processed in blocks to construct a heterogeneous database consisting of a document database, a vector database, and a graph database.

[0016] The specific process of step (1) is as follows:

[0017] (1-1) Use a fixed-size chunking method to break text data into smaller chunks while maintaining some overlap between chunks to ensure that semantic context is not lost between chunks;

[0018] (1-2) Perform word segmentation on each text block and remove stop words; record the document number corresponding to each word and perform word frequency statistics on each word; create an inverted index and store the inverted index in the document database;

[0019] (1-3) Encode the text block using the pre-trained language model to generate the corresponding high-dimensional semantic vector representation; store the vector representation and text block number of each text block in the vector database;

[0020] (1-4) Use a large language model to extract entity-relationship-entity structured triples from text blocks; perform deduplication and entity disambiguation on the extracted triples to remove duplicate triples and merge entities with similar semantics; store the processed triples in a graph database.

[0021] In step (2), based on the user query, keyword search, vector search and knowledge graph search are performed in the document database, vector database and graph database constructed in step (1) respectively to obtain documents and triple information related to the user query.

[0022] The specific process of step (2) is as follows:

[0023] (2-1) Perform keyword search: Segment the user query and remove stop words; use the query to search the inverted index in the text database to find all documents containing the query; use the BM25 algorithm to calculate the match between the query and each document, sort the documents according to the match, and select the documents with the highest match as the keyword search results;

[0024] (2-2) Perform vector retrieval: Input the user query into the pre-trained language model used in step (1-3) to generate a high-dimensional semantic vector representation of the query; use the query vector to search the vector database to find documents with similar semantics to the query; measure the semantic similarity between the query and the document by calculating the cosine similarity between the vectors, and select several documents with the highest semantic similarity to the query as the result of the vector retrieval;

[0025] (2-3) Perform knowledge graph retrieval: Use a large language model to perform entity recognition on user queries and extract entities in the query; search the graph database based on the extracted entity names to obtain triple information related to these entities; merge all retrieved triple information as the result of the knowledge graph retrieval;

[0026] (2-4) Merge the search results from keyword search, vector search, and knowledge graph search, remove duplicate documents, and obtain documents and triple information related to the user query.

[0027] In step (3), in order to further improve the relevance and accuracy of the retrieval results, all retrieval results and user queries are input into the re-ranking model, which outputs a similarity score for each retrieval result to indicate the degree of match between the retrieval result and the user query; all retrieval results are sorted according to their similarity scores with the query, and the K retrieval results with the highest similarity scores are selected to obtain a set of documents and triples that are highly relevant to the user query.

[0028] In step (4), the user query is combined with the re-ranked search results to construct an effective prompt word which is input into the large language model; the large language model combines the input context to generate a coherent and accurate natural language answer.

[0029] A knowledge question and answer system for hybrid retrieval generation enhancement for heterogeneous databases, characterized in that it includes a memory and one or more processors, wherein the memory stores executable code, and when the one or more processors execute the executable code, they are used to implement the above-mentioned knowledge question and answer method for hybrid retrieval generation enhancement for heterogeneous databases.

[0030] A knowledge question answering system for hybrid retrieval generation enhancement of heterogeneous databases, comprising:

[0031] User interaction module: responsible for receiving user query requests and returning the final generated answers to the user;

[0032] Heterogeneous database management module: responsible for the construction and maintenance of heterogeneous databases, including data processing, storage and updating of document databases, vector databases and graph databases;

[0033] Hybrid retrieval module: used to perform keyword retrieval, vector retrieval and knowledge graph retrieval in the document database, vector database and graph database according to the user query, and obtain documents and triple information related to the user query;

[0034] Re-ranking module: used to re-rank the documents and triple information obtained by the hybrid retrieval module, and select the K retrieval results most relevant to the user query for answer generation;

[0035] Answer generation module: It is used to combine user queries with re-ranked search results, construct effective prompt words and input them into the large language model to generate coherent and accurate natural language answers.

[0036] Compared with the prior art, the present invention has the following beneficial effects:

[0037] 1. The present invention integrates keyword retrieval, vector retrieval and knowledge graph retrieval, and uses a re-ranking model to re-rank the retrieval results, which makes up for the shortcomings of the existing single retrieval method in accurate query and deep understanding of related information, and improves the accuracy and comprehensiveness of the retrieval results.

[0038] 2. The present invention uses the search results of hybrid search as context input into the large language model, giving full play to the powerful capabilities of the large language model in context understanding and text generation, enhancing the ability to understand and process complex queries, and improving the accuracy and reliability of knowledge question and answer results.

[0039] 3. The present invention has strong applicability and scalability. Heterogeneous databases provide flexible data models and query methods, which can effectively adapt to changing query requirements and data characteristics. At the same time, heterogeneous databases support the integration of data from different sources, which can form a more complete knowledge system, thereby improving the efficiency of data acquisition. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 A flow chart of a knowledge question-answering method for generating enhanced hybrid retrieval of heterogeneous databases according to the present invention;

[0041] Figure 2 Constructing a flow chart for the heterogeneous database in the present invention;

[0042] Figure 3This is a flow chart of the mixed search in the present invention;

[0043] Figure 4 This is an architecture diagram of a knowledge question-answering system that generates enhanced hybrid retrieval of heterogeneous databases according to the present invention. DETAILED DESCRIPTION

[0044] The present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be pointed out that the embodiments described below are intended to facilitate the understanding of the present invention and do not have any limiting effect on the present invention.

[0045] like Figure 1 As shown, a knowledge question answering method for generating enhanced hybrid retrieval of heterogeneous databases includes the following steps:

[0046] Step (1) divides the text data into blocks and constructs a heterogeneous database consisting of a document database, a vector database, and a graph database. The document database is responsible for storing a large amount of unstructured text content. By establishing an inverted index in the document database, the retrieval requirements for specific keywords or phrases can be met. The vector database is used to store the vector representation of text data. The text is encoded through a deep learning model and converted into a high-dimensional vector to represent the semantic information of the text, so as to quickly find documents with semantics similar to the query vector through vector distance. The graph database is used to store the relationship between entities and represent them through a knowledge graph.

[0047] like Figure 2 As shown, the specific process of step (1) is as follows:

[0048] (1-1) Use a fixed-size chunking method to break up large blocks of text into smaller paragraphs, making the text data easier to process and more accurately matching user queries. When chunking text, it is necessary to maintain some overlap between text chunks to ensure that the semantic context is not lost between chunks.

[0049] (1-2) Perform word segmentation on each text block and remove stop words. Then record the document number corresponding to each word and perform word frequency statistics on each word. Finally, create an inverted index based on the vocabulary information, the document list corresponding to the vocabulary, and the word frequency of the vocabulary in each document, and store the inverted index in the document database.

[0050] (1-3) Use the pre-trained language model to encode the text block and generate the corresponding high-dimensional semantic vector representation, and then store the vector representation and text block number of each text block in the vector database.

[0051] (1-4) The prompt word and the text block are input into the large language model. The large language model extracts the entity-relationship-entity structured triples from the text block according to the prompt word requirements. Then, the extracted triples are deduplicated and entity disambiguated to remove duplicate triples and merge entities with similar semantics. Finally, the processed triples are stored in the graph database.

[0052] Step (2) receives a user query and performs keyword search, vector search and knowledge graph search in the document database, vector database and graph database constructed in step (1) respectively. Keyword search can quickly return documents that match the keywords in the query by searching the inverted index in the text database. Vector search uses a pre-trained language model to perform semantic search in the vector database and uses the cosine similarity algorithm to find the documents that are most semantically relevant to the query. Knowledge graph search extracts entities from the query, searches in the graph database, and returns triples related to the entities mentioned in the query. By merging the search results from keyword search, vector search and knowledge graph search, document and triple information related to the user query can be obtained.

[0053] like Figure 3 As shown, the specific process of step (2) is as follows:

[0054] (2-1) Perform keyword search: First, perform word segmentation on the user query and remove stop words. Then use the query words to search the inverted index in the text database to find all documents containing the query words. Then use the BM25 algorithm to calculate the matching degree between the query and each document. The BM25 algorithm calculates the relevance score between the query and the document based on the word frequency, inverse document frequency, and document length. The score represents their matching degree. The relevance score calculation method is shown in formula (1).

[0055]

[0056] Where D is the document, Q is the query, score(D,Q) is the relevance score between document D and query Q, and q i is the i-th word, f(q i , D) is the frequency of the i-th word in document D, |D| is the length of document D, avg(dl) is the average length of all documents, k1 and b are adjustable parameters, and the general range is k1∈[1.2,2], b=0.75. IDF(q i ) is the word q i The inverse document frequency is an indicator that measures the uniqueness or information content of a word in the entire document collection. It is determined by the number of documents in the entire document collection that contain the word. Its calculation method is shown in formula (2).

[0057]

[0058] Where N is the total number of documents, n(q i ) contains the word q i The number of documents.

[0059] Finally, the documents are sorted according to the relevance scores, and several documents with the highest scores are selected as the results of keyword retrieval.

[0060] (2-2) Perform vector retrieval: First, input the user query into the pre-trained language model used in step (1-3) to generate a high-dimensional semantic vector representation of the query. Then use the query vector to search the vector database to find documents that are semantically similar to the query. The present invention measures the semantic similarity between the query and the document by calculating the cosine similarity between the query vector and the document vector. The cosine similarity calculation formula is shown in formula (3).

[0061]

[0062] Where q is the query vector, d is the document vector, q·d is the dot product of vectors q and d, and ‖q‖ and ‖d‖ are the moduli of vectors q and d, respectively.

[0063] Finally, the documents are sorted according to cosine similarity, and several documents with the highest semantic similarity to the query are selected as the results of vector retrieval.

[0064] (2-3) Perform knowledge graph retrieval: First, use the large language model to perform entity recognition on the user query and extract the entities in the query. Then, based on the extracted entity names, search the graph database to obtain triple information related to these entities. Finally, merge all the retrieved triple information as the result of the knowledge graph retrieval;

[0065] (2-4) Merge the search results from keyword search, vector search, and knowledge graph search, and remove duplicate documents to obtain documents and triple information related to the user query.

[0066] Step (3) uses the re-ranking model to re-rank the search results, thereby further improving the relevance and accuracy of the search results. The re-ranking model takes the query and text as input and directly outputs a similarity score, which is used to represent the degree of match between the query and the text, providing a more detailed relevance measure than the vector embedding-based model.

[0067] First, all search results and user queries are input into the re-ranking model, which outputs a similarity score for each search result. Then, all search results are sorted according to their similarity scores with the query, and the K search results with the highest similarity scores are selected, and finally a set of documents and triples that are highly relevant to the user query is obtained.

[0068] Step (4) combines the user query with the text block and triples obtained in step (3) to construct a valid prompt word and input it into the large language model. The large language model combines the input context to generate a coherent and accurate natural language answer.

[0069] like Figure 4 As shown, a knowledge question answering system for hybrid retrieval generation and enhancement of heterogeneous databases includes the following modules:

[0070] User interaction module: responsible for receiving user query requests and returning the final generated answers to the user.

[0071] Heterogeneous database management module: responsible for the construction and maintenance of heterogeneous databases, including data processing, storage and updating of document databases, vector databases and graph databases, in order to support efficient hybrid retrieval.

[0072] Hybrid retrieval module: Based on user queries, keyword retrieval, vector retrieval and knowledge graph retrieval are performed in the document database, vector database and graph database respectively to obtain documents and triple information related to the user query.

[0073] Re-ranking module: re-rank the documents and triple information obtained by hybrid retrieval, and select the K retrieval results most relevant to the user query for answer generation.

[0074] Answer generation module: Combines user queries with re-ranked search results, constructs effective prompt words and inputs them into the large language model to generate coherent and accurate natural language answers.

[0075] The embodiments described above provide a detailed description of the technical solutions and beneficial effects of the present invention. It should be understood that the above are only specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, supplements and equivalent substitutions made within the scope of the principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A knowledge question answering method for hybrid retrieval generation enhancement of heterogeneous databases, characterized in that: The following steps are involved: (1) Process text data and build heterogeneous databases; (2) Receive user queries and perform hybrid retrieval in heterogeneous databases; merge retrieval results to obtain documents and triple information related to the user query; (3) Use the re-ranking model to re-rank the retrieved documents and triple information, and select the K retrieval results with the highest relevance to the user query; (4) Combine user queries with search results, construct prompt words and input them into a large language model to generate coherent and accurate natural language answers.

2. The knowledge question answering method for generating enhanced hybrid retrieval of heterogeneous databases according to claim 1 is characterized in that: In step (1), the text data is processed in blocks to construct a heterogeneous database consisting of a document database, a vector database, and a graph database.

3. The knowledge question answering method for generating enhanced hybrid retrieval of heterogeneous databases according to claim 2 is characterized in that: The specific process of step (1) is as follows: (1-1) Use a fixed-size chunking method to break text data into smaller chunks while maintaining some overlap between chunks to ensure that semantic context is not lost between chunks; (1-2) Perform word segmentation on each text block and remove stop words; record the document number corresponding to each word and perform word frequency statistics on each word; create an inverted index and store the inverted index in the document database; (1-3) Encode the text block using the pre-trained language model to generate the corresponding high-dimensional semantic vector representation; store the vector representation and text block number of each text block in the vector database; (1-4) Use a large language model to extract structured entity-relation-entity triples from text blocks; The extracted triples are deduplicated and entity disambiguated to remove duplicate triples and merge entities with similar semantics; the processed triples are stored in a graph database.

4. The knowledge question answering method for generating enhanced hybrid retrieval of heterogeneous databases according to claim 1 is characterized in that: In step (2), based on the user query, keyword search, vector search and knowledge graph search are performed in the document database, vector database and graph database constructed in step (1) respectively to obtain documents and triple information related to the user query.

5. The knowledge question answering method for generating enhanced hybrid retrieval of heterogeneous databases according to claim 3 is characterized in that: The specific process of step (2) is as follows: (2-1) Perform keyword search: Segment the user query and remove stop words; use the query to search the inverted index in the text database to find all documents containing the query; use the BM25 algorithm to calculate the match between the query and each document, sort the documents according to the match, and select the documents with the highest match as the keyword search results; (2-2) Perform vector retrieval: Input the user query into the pre-trained language model used in step (1-3) to generate a high-dimensional semantic vector representation of the query; use the query vector to search the vector database to find documents with similar semantics to the query; measure the semantic similarity between the query and the document by calculating the cosine similarity between the vectors, and select several documents with the highest semantic similarity to the query as the result of the vector retrieval; (2-3) Perform knowledge graph retrieval: Use a large language model to perform entity recognition on user queries and extract entities in the query; search the graph database based on the extracted entity names to obtain triple information related to these entities; merge all retrieved triple information as the result of the knowledge graph retrieval; (2-4) Merge the search results from keyword search, vector search, and knowledge graph search, and remove duplicate documents to obtain documents and triple information related to the user query.

6. The knowledge question answering method for generating enhanced hybrid retrieval of heterogeneous databases according to claim 1 is characterized in that: In step (3), all search results and user queries are input into the re-ranking model, which outputs a similarity score for each search result to indicate the degree of match between the search result and the user query. All search results are sorted according to their similarity scores with the query, and the K search results with the highest similarity scores are selected to obtain a set of documents and triples that are highly relevant to the user query.

7. The knowledge question answering method for generating enhanced hybrid retrieval of heterogeneous databases according to claim 1 is characterized in that: In step (4), the user query is combined with the re-ranked search results to construct an effective prompt word which is input into the large language model; the large language model combines the input context to generate a coherent and accurate natural language answer.

8. A knowledge question answering system for hybrid retrieval generation and enhancement of heterogeneous databases, characterized in that: It includes a memory and one or more processors, wherein the memory stores executable code, and when the one or more processors execute the executable code, it is used to implement the knowledge question answering method for generating enhanced hybrid retrieval of heterogeneous databases as described in any one of claims 1-7.

9. A knowledge question answering system for hybrid retrieval generation and enhancement of heterogeneous databases, characterized in that: include: User interaction module: responsible for receiving user query requests and returning the final generated answers to the user; Heterogeneous database management module: responsible for the construction and maintenance of heterogeneous databases, including data processing, storage and updating of document databases, vector databases and graph databases; Hybrid retrieval module: used to perform keyword retrieval, vector retrieval and knowledge graph retrieval in the document database, vector database and graph database according to the user query, and obtain documents and triple information related to the user query; Re-ranking module: used to re-rank the documents and triple information obtained by the hybrid retrieval module, and select the K retrieval results most relevant to the user query for answer generation; Answer generation module: It is used to combine user queries with re-ranked search results, construct effective prompt words and input them into the large language model to generate coherent and accurate natural language answers.

Images