Intelligent question answering system based on heterogeneous graph RAG technology

Abstract

The invention provides an intelligent questioning and answering system based on a heterogeneous graph RAG technology, and relates to the field of intelligent questioning and answering. A heterogeneous graph model is constructed, alternative knowledge nodes are extracted from the heterogeneous graph model, and questions input by a user are obtained; semantic coding is carried out on the alternative knowledge nodes and the questions input by the user by adopting a data processing technology based on deep learning, so that a semantic matching degree is automatically obtained according to principal component significant fusion representation between the semantic features of the alternative knowledge nodes and the semantic features of the input questions; and automatically determining whether to return the alternative knowledge node or not based on comparison between the semantic matching degree and a preset threshold value. The intelligent question answering system based on deep learning and heterogeneous graphs has higher flexibility and stronger generalization ability, and can capture complex structures and meanings in natural languages, thereby improving the accuracy of question analysis.

Description

Technical Field

[0001] The present application relates to the field of intelligent question answering, and more specifically, to an intelligent question answering system based on heterogeneous graph RAG technology. Background Art

[0002] In the era of information explosion, users are facing a world of information overflow, with a large amount of data generated every day. Whether it is the dynamics on social media, the information in emails, the latest information in news reports, or the research results in academic papers, various information sources are emerging like a tide. Users are more eager to get accurate answers to specific questions directly and quickly, which means that they hope that the process of obtaining information can be more efficient and convenient, reducing unnecessary search and screening steps.

[0003] However, traditional question-answering systems mainly rely on hand-written rules and dictionaries to parse questions and find answers, which are usually called rule-based question-answering systems. Specifically, rule-based question-answering systems have difficulty handling unforeseen question forms or complex and changing language structures because they can only parse questions according to preset rules. Once a question is encountered that goes beyond the rule framework, the system often cannot give a correct response. This limitation is particularly evident in practical applications, because the questions asked by users may be diverse, containing various grammatical structures and expressions, and rule-based question-answering systems lack sufficient flexibility to deal with these situations. In addition, users' questions often depend on the content of the previous conversation, the so-called contextual information. In multi-round conversations, there may be implicit relationships or prerequisites between previous and subsequent questions, and rule-based systems often find it difficult to capture this contextual information.

[0004] Therefore, an optimized question answering system is desired. Summary of the invention

[0005] In order to solve the above technical problems, the present application is proposed. The embodiment of the present application provides an intelligent question answering system based on heterogeneous graph RAG technology.

[0006] According to one aspect of the present application, an intelligent question answering system based on heterogeneous graph RAG technology is provided, which includes:

[0007] A heterogeneous graph model construction module, used to construct a heterogeneous graph model, wherein the heterogeneous graph model includes knowledge nodes and relationship edges between knowledge nodes;

[0008] A question data collection module is used to obtain questions input by users;

[0009] A question semantic encoding module, used for semantically encoding the question input by the user to obtain a question semantic encoding feature vector;

[0010] A candidate knowledge node semantic encoding module, used for extracting candidate knowledge nodes from the heterogeneous graph model, and semantically encoding the candidate knowledge nodes to obtain a candidate knowledge node semantic encoding feature vector;

[0011] A question-alternative knowledge point node semantic joint principal component fusion module is used to perform alternative knowledge-question semantic saliency query matching and fusion on the alternative knowledge node semantic encoding feature vector and the question semantic encoding feature vector to obtain a question-alternative knowledge point node semantic saliency fusion representation feature;

[0012] A semantic matching module, used to obtain a semantic matching degree based on the semantically significant fusion representation features of the question-candidate knowledge point node, and return the candidate knowledge node based on the semantic matching degree;

[0013] The question-candidate knowledge point node semantic joint principal component fusion module includes:

[0014] A candidate knowledge node question semantic feature sample covariance calculation unit is used to calculate the sample covariance matrix of the candidate knowledge node semantic encoding feature vector and the question semantic encoding feature vector to obtain a candidate knowledge node semantic sample covariance matrix and a question semantic sample covariance matrix;

[0015] A principal component feature extraction unit for the covariance matrix of candidate knowledge node question semantic samples, used for extracting principal component feature vectors based on matrix decomposition from the covariance matrix of candidate knowledge node semantic samples and the covariance matrix of question semantic samples to obtain a set of principal component feature vectors of candidate knowledge node semantics and a set of principal component feature vectors of question semantics;

[0016] The candidate knowledge node question semantic principal component fusion unit is used to perform approximate query matching semantic association analysis on the set of candidate knowledge node semantic principal component feature vectors and the set of question semantic principal component feature vectors to obtain a question-alternative knowledge point node semantic joint principal component fusion representation vector as the question-alternative knowledge point node semantic significant fusion representation feature.

[0017] In the above-mentioned intelligent question-answering system based on heterogeneous graph RAG technology, the candidate knowledge node question semantic feature sample covariance calculation unit includes: a candidate knowledge node question semantic standardization subunit, which is used to standardize the candidate knowledge node semantic coding feature vector and the question semantic coding feature vector to obtain a standardized candidate knowledge node semantic coding feature vector and a standardized question semantic coding feature vector; a standardized candidate knowledge node question semantic sample covariance matrix calculation subunit, which is used to calculate the sample covariance matrix of the standardized candidate knowledge node semantic coding feature vector and the standardized question semantic coding feature vector to obtain the candidate knowledge node semantic sample covariance matrix and the question semantic sample covariance matrix.

[0018] In the above-mentioned intelligent question-answering system based on heterogeneous graph RAG technology, the candidate knowledge node question semantic standardization subunit is used to: respectively calculate the mean and standard deviation of the semantic coding feature vector of the candidate knowledge node to obtain the mean of the semantic feature of the candidate knowledge node and the standard deviation of the semantic feature of the candidate knowledge node; after positionally subtracting the semantic coding feature vector of the candidate knowledge node from the mean of the semantic feature of the candidate knowledge node, divide the calculated semantic difference vector of the alternative knowledge node by the standard deviation of the semantic feature of the candidate knowledge node to obtain the standardized alternative knowledge node semantic coding feature vector; respectively calculate the mean and standard deviation of the question semantic coding feature vector to obtain the mean of the question semantic feature and the standard deviation of the question semantic feature; after positionally subtracting the question semantic coding feature vector from the mean of the semantic feature of the question, divide the calculated question semantic difference vector by the standard deviation of the semantic feature of the question to obtain the standardized question semantic coding feature vector.

[0019] In the above-mentioned intelligent question-answering system based on heterogeneous graph RAG technology, the standardized candidate knowledge node question semantic sample covariance matrix calculation subunit is used to: after multiplying the transpose vector of the standardized candidate knowledge node semantic coding feature vector with the standardized candidate knowledge node semantic coding feature vector, the obtained standardized candidate knowledge node semantic association matrix is ​​divided by the value obtained by subtracting one from the length of the standardized candidate knowledge node semantic coding feature vector to obtain the candidate knowledge node semantic sample covariance matrix; after multiplying the transpose vector of the standardized question semantic coding feature vector with the standardized question semantic coding feature vector, the obtained standardized question semantic association matrix is ​​divided by the value obtained by subtracting one from the length of the standardized question semantic coding feature vector to obtain the question semantic sample covariance matrix.

[0020] In the above-mentioned intelligent question-answering system based on heterogeneous graph RAG technology, the candidate knowledge node question semantic principal component fusion unit includes: a candidate knowledge node-question semantic maximum approximate query subunit, which is used to input the set of candidate knowledge node semantic principal component feature vectors and the set of question semantic principal component feature vectors into the maximum approximate query matching network to obtain a set of best matching pairs of candidate knowledge node semantic principal component feature vectors and question semantic principal component feature vectors; a candidate knowledge node-question semantic principal component fusion subunit, which is used to input the best matching pairs of each candidate knowledge node semantic principal component feature vector and question semantic principal component feature vector in the set of best matching pairs of candidate knowledge node semantic principal component feature vectors and question semantic principal component feature vectors into the semantic fine-grained gating joint module to obtain a set of candidate knowledge node-question semantic principal component fusion feature vectors; a question-alternative knowledge point node semantic joint principal component fusion representation generation subunit, which is used to cascade the set of candidate knowledge node-question semantic principal component fusion feature vectors to obtain the question-alternative knowledge point node semantic joint principal component fusion representation vector

[0021] In the above-mentioned intelligent question-answering system based on heterogeneous graph RAG technology, the alternative knowledge node-question semantic maximum approximation query sub-unit is used to: extract a predetermined alternative knowledge node semantic main component feature vector from the set of alternative knowledge node semantic main component feature vectors; calculate the cosine similarity between the predetermined alternative knowledge node semantic main component feature vector and each question semantic main component feature vector in the set of question semantic main component feature vectors to obtain a set of matching query similarities; and use the question semantic main component feature vector corresponding to the largest matching query similarity in the set of matching query similarities and the predetermined alternative knowledge node semantic main component feature vector as the best matching pair of the predetermined alternative knowledge node semantic main component feature vector and the question semantic main component feature vector.

[0022] In the above-mentioned intelligent question-answering system based on heterogeneous graph RAG technology, the candidate knowledge node-question semantics principal component fusion subunit is used to: respectively calculate the positional difference, positional dot product and positional addition between the best matching pair of the candidate knowledge node semantics principal component feature vector and the question semantics principal component feature vector to obtain the candidate knowledge node-question semantics principal component pattern differential vector, the candidate knowledge node-question semantics principal component pattern dot product vector and the candidate knowledge node-question semantics principal component pattern sum vector; cascade the candidate knowledge node-question semantics principal component pattern differential vector, the candidate knowledge node-question semantics principal component pattern dot product vector and the candidate knowledge node-question semantics principal component pattern sum vector, and then perform one-dimensional convolution encoding to obtain the candidate knowledge node-question semantics principal component pattern multi-dimensional fusion vector; perform local window-based maximum pooling processing on the candidate knowledge node-question semantics principal component pattern multi-dimensional fusion vector to obtain the candidate knowledge node-question semantics principal component fusion feature vector.

[0023] In the above-mentioned intelligent question-answering system based on heterogeneous graph RAG technology, the semantic matching module includes: a semantic measurement unit, which is used to input the question-alternative knowledge point node semantic joint principal component fusion representation vector into the semantic measurement network to obtain the semantic matching degree; an alternative knowledge node feedback unit, which is used to return the alternative knowledge node in response to the semantic matching degree being greater than or equal to a preset threshold.

[0024] Compared with the prior art, the intelligent question-answering system based on heterogeneous graph RAG technology provided by the present application constructs a heterogeneous graph model, extracts alternative knowledge nodes from the heterogeneous graph model, obtains questions input by users, and uses data processing technology based on deep learning to semantically encode the alternative knowledge nodes and the questions input by users, thereby automatically obtaining the semantic matching degree based on the principal component significant fusion representation between the semantic features of the alternative knowledge nodes and the semantic features of the input questions, and automatically determining whether to return the alternative knowledge nodes based on the comparison of the semantic matching degree with a preset threshold. This intelligent question-answering system based on deep learning and heterogeneous graphs has higher flexibility and stronger generalization ability, and can capture the complex structure and meaning in natural language, thereby improving the accuracy of question parsing. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the prior art descriptions. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work. In the drawings:

[0026] Figure 1 This is a system block diagram of an intelligent question-answering system based on heterogeneous graph RAG technology according to an embodiment of the present application.

[0027] Figure 2 This is a data flow diagram of an intelligent question-answering system based on heterogeneous graph RAG technology according to an embodiment of the present application.

[0028] Figure 3 It is a block diagram of the question-alternative knowledge point node semantic joint principal component fusion module in the intelligent question-answering system based on the heterogeneous graph RAG technology according to an embodiment of the present application.

[0029] Figure 4 It is a block diagram of a covariance calculation unit for semantic feature samples of candidate knowledge node questions in an intelligent question-answering system based on heterogeneous graph RAG technology according to an embodiment of the present application.

[0030] Figure 5 It is a block diagram of the semantic principal component fusion unit of the candidate knowledge node question in the intelligent question-answering system based on the heterogeneous graph RAG technology according to an embodiment of the present application.

[0031] Figure 6 It is a block diagram of the semantic matching module in the intelligent question-answering system based on the heterogeneous graph RAG technology according to an embodiment of the present application. DETAILED DESCRIPTION

[0032] Below, the exemplary embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application, and it should be understood that the present application is not limited to the exemplary embodiments described here.

[0033] In the era of information explosion, users are faced with massive amounts of data and are eager to obtain accurate answers to specific questions efficiently and conveniently. Traditional question-answering systems, which rely on fixed rules and dictionaries, are often unable to cope with questions with novel or complex language structures, making it difficult to provide effective answers. In addition, users' questions are often closely related to the context of the conversation, and rule-based question-answering systems usually find it difficult to capture and utilize this contextual information. This problem is particularly prominent in multi-round conversations.

[0034] It should be understood that heterogeneous graph RAG technology is an artificial intelligence technology that combines heterogeneous graphs with retrieval enhanced generation. A heterogeneous graph is a graph structure containing multiple types of nodes and multiple types of edges. It can construct data from different sources and types into a heterogeneous graph, where nodes can represent different entities (such as documents, knowledge fragments, concepts, etc.), and edges represent the relationships between these entities (such as reference relationships, similarity relationships, etc.). RAG technology can retrieve and integrate information on this graph structure, and then use the retrieved relevant information to generate natural language answers. In applications in the field of intelligent question and answer, heterogeneous graph RAG technology can construct multiple information sources such as knowledge bases, document libraries, structured data, etc. into heterogeneous graphs, and by utilizing the relationships between entities in heterogeneous graphs, it can more accurately understand the context of the question, thereby generating more comprehensive and in-depth answers.

[0035] Based on this, the present application proposes an intelligent question-answering system based on heterogeneous graph RAG technology, which constructs a heterogeneous graph model, extracts alternative knowledge nodes from the heterogeneous graph model, obtains questions input by users, and uses knowledge node analysis and text processing technology based on deep learning to semantically encode the alternative knowledge nodes and the questions input by users, so as to automatically obtain the semantic matching degree based on the principal component significant fusion representation between the semantic features of the alternative knowledge nodes and the semantic features of the input questions, and automatically determines whether to return the alternative knowledge nodes based on the comparison of the semantic matching degree with a preset threshold. In this way, compared with rule-based systems, this system based on deep learning and heterogeneous graphs has higher flexibility and stronger generalization ability. At the same time, it can capture the complex structure and meaning in natural language, thereby improving the accuracy of question parsing.

[0036] Figure 1 This is a system block diagram of an intelligent question-answering system based on heterogeneous graph RAG technology according to an embodiment of the present application. Figure 2 Schematic diagram of data flow of intelligent question-answering system based on heterogeneous graph RAG technology according to an embodiment of the present application. Figure 1 and Figure 2As shown, in the intelligent question answering system 100 based on heterogeneous graph RAG technology, it includes: a heterogeneous graph model construction module 110, which is used to construct a heterogeneous graph model, and the heterogeneous graph model includes knowledge nodes and relationship edges between knowledge nodes; a question data collection module 120, which is used to obtain questions input by users; a question semantic encoding module 130, which is used to semantically encode the questions input by users to obtain question semantic encoding feature vectors; an alternative knowledge node semantic encoding module 140, which is used to extract alternative knowledge nodes from the heterogeneous graph model, and semantically encode the alternative knowledge nodes to obtain alternative knowledge node semantic encoding feature vectors; a question-alternative knowledge point node semantic joint principal component fusion module 150, which is used to perform alternative knowledge-question semantic significance query matching and fusion on the alternative knowledge node semantic encoding feature vector and the question semantic encoding feature vector to obtain question-alternative knowledge point node semantic significance fusion representation features; a semantic matching module 160, which is used to obtain a semantic matching degree based on the question-alternative knowledge point node semantic significance fusion representation features, and return the alternative knowledge node based on the semantic matching degree.

[0037] In an embodiment of the present application, the heterogeneous graph model construction module 110 is used to construct a heterogeneous graph model, which includes knowledge nodes and relationship edges between knowledge nodes. It should be understood that the heterogeneous graph model allows the inclusion of multiple types of nodes, different types of edges can represent different relationships between nodes, and as the knowledge base changes, the heterogeneous graph model can flexibly add, delete and update knowledge nodes and their relationships, which makes the representation of knowledge data more flexible, rich and dynamic. That is, by constructing a heterogeneous graph structure containing knowledge in multiple fields, effective retrieval and integration of cross-domain knowledge can be achieved. In particular, in a specific implementable embodiment of the present application, constructing a heterogeneous graph model includes the following steps: a. Collecting data related to heterogeneous graph construction, including structured data (such as database tables, XML files), semi-structured data (such as JSON, HTML documents) and unstructured data (such as text, images, audio and video), which cover different fields, such as medicine, finance and entertainment, each of which has its own unique entity types and relationships; b. Preprocessing the collected data related to heterogeneous graph construction, including cleaning the data to remove duplicates, fill in missing values, and perform text segmentation, named entity recognition (NER) and relationship extraction, etc.; c. Constructing a heterogeneous graph, which involves defining different node types and edge types, and creating specific node instances, and establishing edge connections between corresponding nodes in response to the relationship information of different node instances to form a complete heterogeneous graph.

[0038] In an embodiment of the present application, the question data acquisition module 120 is used to obtain questions input by the user. It should be understood that the questions input by the user contain the main content or object that the user is concerned about, the relevant information that the user hopes to obtain by asking questions, and the specific restrictions provided by the user in the questions. By deeply analyzing and understanding the user's questions, it is possible to effectively retrieve and match relevant knowledge and provide users with fast and accurate answers in a complex information environment. In other words, the questions input by the user are the starting point for the response of the intelligent answering system. It provides a clear query direction and helps the system quickly locate relevant nodes in a complex knowledge graph, thereby improving the accuracy of the answer.

[0039] In the embodiment of the present application, the question semantic encoding module 130 is used to semantically encode the question input by the user to obtain a question semantic encoding feature vector. Accordingly, considering that the question input by the user reflects the real intention of the user's question, the semantic information of these intentions plays an important role in understanding the core problem of the user. Based on this, in the technical solution of the present application, the question input by the user is semantically encoded to capture and extract the key core semantic information contained in the user's question to obtain a question semantic encoding feature vector.

[0040] In an embodiment of the present application, the alternative knowledge node semantic encoding module 140 is used to extract alternative knowledge nodes from the heterogeneous graph model, and semantically encode the alternative knowledge nodes to obtain a semantic encoding feature vector of the alternative knowledge nodes. It should be understood that the heterogeneous graph model contains different types of knowledge nodes and complex relationships between them, and these nodes may represent different concepts, entities or information fragments. In order to analyze the entities or concepts related to the question more carefully and accurately, the information most relevant to the user query can be screened out to provide an accurate answer. In the technical solution of the present application, alternative knowledge nodes are extracted from the heterogeneous graph model. Furthermore, considering that the alternative knowledge nodes contain rich semantic information and deep meanings of the nodes, in the technical solution of the present application, the alternative knowledge nodes are semantically encoded to obtain a semantic encoding feature vector of the alternative knowledge nodes, so that the semantic information expressed in the alternative nodes can be understood more accurately, thereby improving the accuracy of subsequent answers to user queries.

[0041] In an embodiment of the present application, the question-alternative knowledge point node semantic joint principal component fusion module 150 is used to perform alternative knowledge-question semantic significance query matching fusion on the alternative knowledge node semantic encoding feature vector and the question semantic encoding feature vector to obtain the question-alternative knowledge point node semantic significance fusion representation feature. Accordingly, it is considered that the alternative knowledge node semantic encoding feature vector captures the core semantic information of the question input by the user, including the intent, key information and context of the question. The question semantic encoding feature vector contains the semantic representation of information related to the node, such as entities, concepts, etc. And there is key principal component semantic information between the two. Therefore, in order to ensure that in the subsequent matching process, the system can simultaneously consider and enhance the semantic relevance of the question intent and the alternative knowledge nodes, so as to ensure that the most representative and discriminative features are given priority in the matching process, so that more personalized answers or recommendations can be provided later, in the technical solution of the present application, the alternative knowledge node semantic encoding feature vector and the question semantic encoding feature vector are fused for alternative knowledge-question semantic saliency query matching to obtain the question-alternative knowledge point node semantic joint principal component fusion representation vector as the question-alternative knowledge point node semantic saliency fusion representation feature.

[0042] It is worth mentioning that the alternative knowledge-question semantic saliency query matching fusion is a technology that integrates feature selection, principal component analysis (PCA) and key feature integration, which is used to create a significant correlation representation between feature vectors to highlight key features and reduce unnecessary information. Specifically, firstly, the alternative knowledge node semantic encoding feature vector and the question semantic encoding feature vector are standardized so that different features are compared at the same scale, eliminating the influence of dimension, so that the model will not produce deviations due to different feature scales to obtain standardized alternative knowledge node semantic encoding feature vectors and standardized question semantic encoding feature vectors. Then, the sample covariance matrix of the standardized features is calculated to obtain the alternative knowledge node semantic sample covariance matrix and the question semantic sample covariance matrix, and the two sample covariance matrices are subjected to feature vector extraction based on matrix decomposition to obtain a set of alternative knowledge node semantic principal component feature vectors and a set of question semantic principal component feature vectors. That is, the principal component features often have a more superior representation ability and can effectively capture the core semantic information hidden in the original features. Afterwards, the two sets of principal component feature vectors are matched based on the maximum similarity query of feature vectors. The most relevant semantic information is identified by measuring the similarity between each pair of feature vectors, thereby determining the optimal matching pair. Furthermore, the gating mechanism is used to perform fine-grained semantic integration on the optimal matching pair to enhance and highlight the relevance and importance of the semantics in each pair of matches, thereby obtaining a set of candidate knowledge node-question semantic principal component fusion feature vectors. Finally, the set of candidate knowledge node-question semantic principal component fusion feature vectors is cascaded to achieve a comprehensive and comprehensive in-depth combination of the key information of the candidate nodes and the key information of the user's question, thereby obtaining the question-alternative knowledge point node semantic joint principal component fusion representation vector.

[0043] Specifically, Figure 3 FIG. 1 is a block diagram of a question-candidate knowledge point node semantic joint principal component fusion module in an intelligent question-answering system based on heterogeneous graph RAG technology according to an embodiment of the present application. Figure 3As shown, the question-alternative knowledge point node semantic joint principal component fusion module 150 includes: an alternative knowledge node question semantic feature sample covariance calculation unit 151, which is used to calculate the sample covariance matrix of the alternative knowledge node semantic encoding feature vector and the question semantic encoding feature vector to obtain the alternative knowledge node semantic sample covariance matrix and the question semantic sample covariance matrix; an alternative knowledge node question semantic sample covariance matrix principal component feature extraction unit 152, which is used to perform matrix decomposition-based principal component feature vector extraction on the alternative knowledge node semantic sample covariance matrix and the question semantic sample covariance matrix to obtain a set of alternative knowledge node semantic principal component feature vectors and a set of question semantic principal component feature vectors; an alternative knowledge node question semantic principal component fusion unit 153, which is used to perform approximate query matching semantic association analysis on the set of alternative knowledge node semantic principal component feature vectors and the set of question semantic principal component feature vectors to obtain the question-alternative knowledge point node semantic joint principal component fusion representation vector as the question-alternative knowledge point node semantic significant fusion representation feature.

[0044] More specifically, Figure 4 : is a block diagram of a covariance calculation unit for semantic feature samples of candidate knowledge node questions in an intelligent question-answering system based on heterogeneous graph RAG technology according to an embodiment of the present application. Figure 4 As shown, the candidate knowledge node question semantic feature sample covariance calculation unit 151 includes: a candidate knowledge node question semantic standardization subunit 1511, which is used to standardize the candidate knowledge node semantic coding feature vector and the question semantic coding feature vector to obtain a standardized candidate knowledge node semantic coding feature vector and a standardized question semantic coding feature vector; a standardized candidate knowledge node question semantic sample covariance matrix calculation subunit 1512, which is used to calculate the sample covariance matrix of the standardized candidate knowledge node semantic coding feature vector and the standardized question semantic coding feature vector to obtain the candidate knowledge node semantic sample covariance matrix and the question semantic sample covariance matrix.

[0045] More specifically, in an embodiment of the present application, the alternative knowledge node question semantics standardization subunit is used to: respectively calculate the mean and standard deviation of the alternative knowledge node semantic coding feature vector to obtain the alternative knowledge node semantic feature mean and the alternative knowledge node semantic feature standard deviation; after positionally subtracting the alternative knowledge node semantic coding feature vector from the alternative knowledge node semantic feature mean, divide the calculated alternative knowledge node semantic difference vector by the alternative knowledge node semantic feature standard deviation to obtain the standardized alternative knowledge node semantic coding feature vector; respectively calculate the mean and standard deviation of the question semantic coding feature vector to obtain the question semantic feature mean and the question semantic feature standard deviation; after positionally subtracting the question semantic coding feature vector from the question semantic feature mean, divide the calculated question semantic difference vector by the question semantic feature standard deviation to obtain the standardized question semantic coding feature vector.

[0046] More specifically, in an embodiment of the present application, the standardized candidate knowledge node question semantic sample covariance matrix calculation subunit is used to: after multiplying the transposed vector of the standardized candidate knowledge node semantic coding feature vector with the standardized candidate knowledge node semantic coding feature vector, divide the obtained standardized candidate knowledge node semantic association matrix by the value obtained by subtracting one from the length of the standardized candidate knowledge node semantic coding feature vector by position to obtain the alternative knowledge node semantic sample covariance matrix; after multiplying the transposed vector of the standardized question semantic coding feature vector with the standardized question semantic coding feature vector, divide the obtained standardized question semantic association matrix by the value obtained by subtracting one from the length of the standardized question semantic coding feature vector by position to obtain the question semantic sample covariance matrix.

[0047] More specifically, Figure 5 FIG. 1 is a block diagram of a semantic principal component fusion unit for candidate knowledge node questions in an intelligent question-answering system based on heterogeneous graph RAG technology according to an embodiment of the present application. Figure 5As shown, the candidate knowledge node question semantics principal component fusion unit 153 includes: a candidate knowledge node-question semantics maximum approximate query subunit 1531, which is used to input the set of candidate knowledge node semantics principal component feature vectors and the set of question semantics principal component feature vectors into a maximum approximate query matching network to obtain a set of best matching pairs of candidate knowledge node semantics principal component feature vectors and question semantics principal component feature vectors; an alternative knowledge node-question semantics principal component fusion subunit 1532, which is used to input the best matching pairs of each candidate knowledge node semantics principal component feature vector and question semantics principal component feature vector in the set of best matching pairs of candidate knowledge node semantics principal component feature vectors and question semantics principal component feature vectors into a semantic fine-grained gated joint module to obtain a set of candidate knowledge node-question semantics principal component fusion feature vectors; and a question-alternative knowledge point node semantics joint principal component fusion representation generation subunit 1533, which is used to cascade the set of candidate knowledge node-question semantics principal component fusion feature vectors to obtain the question-alternative knowledge point node semantics joint principal component fusion representation vector.

[0048] More specifically, in an embodiment of the present application, the alternative knowledge node-question semantic maximum approximation query sub-unit is used to: extract a predetermined alternative knowledge node semantic main component feature vector from the set of alternative knowledge node semantic main component feature vectors; calculate the cosine similarity between the predetermined alternative knowledge node semantic main component feature vector and each question semantic main component feature vector in the set of question semantic main component feature vectors to obtain a set of matching query similarities; and use the question semantic main component feature vector corresponding to the largest matching query similarity in the set of matching query similarities and the predetermined alternative knowledge node semantic main component feature vector as the best matching pair of the predetermined alternative knowledge node semantic main component feature vector and the question semantic main component feature vector.

[0049] More specifically, in an embodiment of the present application, the alternative knowledge node-question semantics principal component fusion subunit is used to: respectively calculate the positional difference, positional dot product and positional addition between the best matching pair of the alternative knowledge node semantics principal component feature vector and the question semantics principal component feature vector to obtain the alternative knowledge node-question semantics principal component pattern differential vector, the alternative knowledge node-question semantics principal component pattern dot product vector and the alternative knowledge node-question semantics principal component pattern sum vector; cascade the alternative knowledge node-question semantics principal component pattern differential vector, the alternative knowledge node-question semantics principal component pattern dot product vector and the alternative knowledge node-question semantics principal component pattern sum vector, and then perform one-dimensional convolution encoding to obtain a multi-dimensional fusion vector of the alternative knowledge node-question semantics principal component pattern; perform local window-based maximum pooling processing on the multi-dimensional fusion vector of the alternative knowledge node-question semantics principal component pattern to obtain the alternative knowledge node-question semantics principal component fusion feature vector.

[0050] In the embodiment of the present application, specifically, the question-alternative knowledge point node semantic joint principal component fusion module is used to: perform alternative knowledge-question semantic saliency query matching fusion on the alternative knowledge node semantic encoding feature vector and the question semantic encoding feature vector, and the formula can be expressed as:

[0051]

[0052] C 1 =U 1 Λ 1 U 1 T

[0053] C 2 =U 2 Λ 2 U 2 T

[0054] U 1 =[v 11 ,v 12 ,…,v 1m ]

[0055] Λ 1 =diag(λ 11 ,λ 12 ,…,λ 1m )

[0056] U 2 =[v 21 ,v 22 ,…,v 2m ]

[0057] Λ2 =diag(λ 21 ,λ 22 ,…,λ 2m )

[0058]

[0059] v f =[v p1 ;v p2 ;...v pm ]

[0060] Among them, V 1 and V 2 Respectively represent the semantic encoding feature vector of the candidate knowledge node and the semantic encoding feature vector of the question, μ 1 and μ 2 are the means of the semantic encoding feature vector of the candidate knowledge node and the semantic encoding feature vector of the question, σ 1 and σ 2 are the standard deviations of the semantic encoding feature vector of the candidate knowledge node and the semantic encoding feature vector of the question, respectively, 1norm and V 2norm are the standardized candidate knowledge node semantic encoding feature vector and the standardized question semantic encoding feature vector, respectively, 1norm T and V 2norm T V 1norm and V 2norm The transposed vector, n 1 and n 2 are the number of eigenvalues ​​in the standardized candidate knowledge node semantic encoding feature vector and the standardized question semantic encoding feature vector, respectively, 1 and C 2 are the covariance matrix of the candidate knowledge node semantic samples and the covariance matrix of the question semantic samples, U 1 and U 2 are the orthogonal matrix of the semantic principal component of candidate knowledge nodes and the orthogonal matrix of the semantic principal component of question, Λ 1 and Λ 2 are the semantic diagonal matrix of candidate knowledge nodes and the semantic diagonal matrix of question, respectively, diag(λ 11 ,λ 12 ,…,λ 1m ) is the matrix whose diagonal elements are λ 11 ,λ 12 ,…,λ 1m The semantic diagonal matrix of candidate knowledge nodes, λ 11 ,λ 12 ,…,λ 1mare the weight values ​​of the semantic principal component feature vectors of each candidate knowledge node, diag(λ 21 ,λ 22 ,…,λ 2m ) is the matrix whose diagonal elements are λ 21 ,λ 22 ,…,λ 2m The problem semantic diagonal matrix, λ 21 ,λ 22 ,…,λ 2m are the weight values ​​of the semantic principal component feature vectors of each question, U 1 T and U 2 T U 1 and U 2 The transposed matrix, v 11 ,v 12 ,…,v 1m is the semantic principal component feature vector of each candidate knowledge node in the set of semantic principal component feature vectors of the candidate knowledge nodes, v 21 ,v 22 ,…,v 2m is each question semantic principal component feature vector in the set of question semantic principal component feature vectors, <v 1i ·v 2j >To calculate the v 1i and v 2j The inner product of the vectors between ||·|| is the norm of the calculated vector, To return the j value corresponding to the maximum value, k is the maximum approximate matching value. ⊙ and are position difference, position dot multiplication and position addition respectively, [·,·,·] is cascade processing, conv1D(·) is a one-dimensional convolutional encoding operation, MaxPool(·) is a maximum pooling operation, v pi is the ith candidate knowledge node-question semantics principal component fusion feature vector in the set of candidate knowledge node-question semantics principal component fusion feature vectors, m is the number of feature vectors in the set of candidate knowledge node-question semantics principal component fusion feature vectors, v f It is the problem-alternative knowledge point node semantic joint principal component fusion representation vector.

[0061] In the embodiment of the present application, the semantic matching module 160 is used to obtain a semantic matching degree based on the semantically significant fusion representation features of the question-candidate knowledge point node, and return the candidate knowledge node based on the semantic matching degree. Specifically, Figure 6 FIG. 1 is a block diagram of a semantic matching module in an intelligent question-answering system based on heterogeneous graph RAG technology according to an embodiment of the present application. Figure 6 As shown, the semantic matching module 160 includes: a semantic measurement unit 161, which is used to input the semantic joint principal component fusion representation vector of the question-alternative knowledge point node into a semantic measurement network to obtain the semantic matching degree; an alternative knowledge node feedback unit 162, which is used to return the alternative knowledge node in response to the semantic matching degree being greater than or equal to a preset threshold. More specifically, in an embodiment of the present application, the semantic measurement unit is used to: input the semantic joint principal component fusion representation vector of the question-alternative knowledge point node into a decoder-based semantic measurement network to obtain the semantic matching degree. That is, the semantic joint principal component fusion representation vector of the question-alternative knowledge point node obtained by principal component fusion of the semantic encoding feature vector of the selected knowledge node and the semantic encoding feature vector of the question is decoded to automatically obtain the semantic matching degree, and automatically determine whether to return the alternative knowledge node based on the comparison of the semantic matching degree with the preset threshold. In this way, compared with the rule-based system, this system based on deep learning and heterogeneous graphs has higher flexibility and stronger generalization ability. At the same time, it can capture the complex structure and meaning in natural language, thereby improving the accuracy of problem parsing.

[0062] In particular, considering that the question semantic encoding feature vector and the alternative knowledge node semantic encoding feature vector respectively represent the question input by the user and the encoding semantic features of the alternative knowledge node extracted from the heterogeneous graph model, when they are subjected to alternative knowledge-question semantic significance query matching fusion, the semantic difference of the encoding features based on the source semantic difference will cause the local matching degree difference of the feature principal component query, so that the question-alternative knowledge point node semantic joint principal component fusion representation vector will also have a fusion significance difference based on the local matching degree difference. Therefore, it is expected to improve the detail semantic aggregation expression effect based on the fusion significance difference of the question-alternative knowledge point node semantic joint principal component fusion representation vector.

[0063] Preferably, inputting the semantic joint principal component fusion representation vector of the question-alternative knowledge point node into the decoder-based semantic measurement network to obtain the semantic matching degree includes: calculating the sum of the absolute values ​​of each eigenvalue of the semantic joint principal component fusion representation vector of the question-alternative knowledge point node to obtain the first question-alternative knowledge point node semantic joint principal component fusion and modulation value, and calculating the square root of the sum of the squares of each eigenvalue of the semantic joint principal component fusion representation vector of the question-alternative knowledge point node to obtain the second question-alternative knowledge point node semantic joint principal component fusion and modulation value; performing dot-wise subtraction on the semantic joint principal component fusion representation vector of the question-alternative knowledge point node and the second question-alternative knowledge point node semantic joint principal component fusion and modulation value, respectively performing dot-wise multiplication with the number of eigenvalues ​​of the semantic joint principal component fusion representation vector of the question-alternative knowledge point node and the inverse of the semantic joint principal component fusion and modulation value of the first question-alternative knowledge point node, and taking the inverse of each eigenvalue to obtain the first question-alternative knowledge point node semantic joint principal component fusion and modulation value. node semantic joint principal component fusion phase conversion vector; after dot-wise subtraction of the question-candidate knowledge point node semantic joint principal component fusion representation vector and the first question-candidate knowledge point node semantic joint principal component fusion and modulation value, dot-wise multiplication with the square root of the number of eigenvalues ​​of the question-candidate knowledge point node semantic joint principal component fusion representation vector and the inverse of the second question-candidate knowledge point node semantic joint principal component fusion and modulation value, and taking the inverse of each eigenvalue to obtain the second question-candidate knowledge point node semantic joint principal component fusion phase conversion vector; dot-wise subtraction of the first question-candidate knowledge point node semantic joint principal component fusion phase conversion vector from the dot product vector of the second question-candidate knowledge point node semantic joint principal component fusion phase conversion vector and the weighted hyperparameter to obtain the optimized question-candidate knowledge point node semantic joint principal component fusion representation vector; the optimized question-candidate knowledge point node semantic joint principal component fusion representation vector is input into the decoder-based semantic measurement network to obtain the semantic matching degree.

[0064] Here, the optimization expression of the problem-candidate knowledge point node semantic joint principal component fusion representation vector is:

[0065]

[0066]

[0067] Where V is the semantic joint principal component fusion representation vector of the question-candidate knowledge point node, v irepresents the eigenvalue of the i-th position in the semantic joint principal component fusion representation vector of the question-candidate knowledge point node, n represents the number of eigenvalues ​​of the semantic joint principal component fusion representation vector of the question-candidate knowledge point node, α represents the semantic joint principal component fusion and modulation value of the first question-candidate knowledge point node, β represents the semantic joint principal component fusion and modulation value of the second question-candidate knowledge point node, ⊙ represents the point multiplication by position, [·] ⊙-1 To take the inverse of each eigenvalue in the vector, V 1 represents the first question-candidate knowledge point node semantic joint principal component fusion phase transformation vector, V 2 represents the second question-alternative knowledge point node semantic joint principal component fusion phase transformation vector, represents point-by-point subtraction, ω represents the weighted hyperparameter, and V' represents the optimized problem-candidate knowledge point node semantic joint principal component fusion representation vector.

[0068] That is, the present application uses the difference of the feature value of the semantic joint principal component fusion representation vector of the question-alternative knowledge point node relative to the feature set of the vector of the semantic joint principal component fusion representation vector of the question-alternative knowledge point node as the difference of the modulation representation as the semantic change intensity information, and performs the phase-like conversion corresponding to the position-based intensity modulation through different and modulation representation forms, so as to make the aggregation enhancement of the semantic change phase perception be able to improve the axial aggregation receptive field along the feature aggregation direction by performing the spatial translation operation based on alternating stacking under the vector set scale balance of the semantic joint principal component fusion representation vector of the question-alternative knowledge point node, so as to improve the perception effect of the aggregated semantics of the semantic joint principal component fusion representation vector of the question-alternative knowledge point node for the semantic change of the time sequence, so as to improve the expression effect of the semantic joint principal component fusion representation vector of the question-alternative knowledge point node, and improve the accuracy of the semantic matching degree obtained by inputting the semantic metric network based on the decoder. In this way, compared with the rule-based system, this system based on deep learning and heterogeneous graph has higher flexibility and stronger generalization ability. At the same time, it can capture the complex structure and meaning in natural language, thereby improving the accuracy of problem parsing.

[0069] In summary, the intelligent question-answering system 100 based on heterogeneous graph RAG technology according to the embodiment of the present application is explained, which constructs a heterogeneous graph model, extracts alternative knowledge nodes from the heterogeneous graph model, obtains questions input by users, and uses deep learning-based data processing technology to semantically encode the alternative knowledge nodes and the questions input by users, thereby automatically obtaining the semantic matching degree based on the principal component significant fusion representation between the semantic features of the alternative knowledge nodes and the semantic features of the input questions, and automatically determines whether to return the alternative knowledge nodes based on the comparison of the semantic matching degree with a preset threshold. This intelligent question-answering system based on deep learning and heterogeneous graphs has higher flexibility and stronger generalization ability, and can capture the complex structure and meaning in natural language, thereby improving the accuracy of question parsing.

[0070] As described above, the intelligent question and answer system 100 based on heterogeneous graph RAG technology according to the embodiment of the present application can be implemented in various terminal devices, such as a server for intelligent question and answer based on heterogeneous graph RAG technology. In one example, the intelligent question and answer system 100 based on heterogeneous graph RAG technology according to the embodiment of the present application can be integrated into the terminal device as a software module and / or hardware module. For example, the intelligent question and answer system 100 based on heterogeneous graph RAG technology can be a software module in the operating system of the terminal device, or it can be an application developed for the terminal device; of course, the intelligent question and answer system 100 based on heterogeneous graph RAG technology can also be one of the many hardware modules of the terminal device.

[0071] Alternatively, in another example, the intelligent question and answer system 100 based on heterogeneous graph RAG technology and the terminal device may also be separate devices, and the intelligent question and answer system 100 based on heterogeneous graph RAG technology may be connected to the terminal device via a wired and / or wireless network, and transmit interactive information in accordance with an agreed data format.

[0072] In the system embodiments provided in this application, it should be understood that the disclosed system can be implemented in other ways. For example, the system embodiments described above are only illustrative, for example, the division of the modules is only a logical function division, and there may be other division methods in actual implementation.

[0073] The modules described as separate components may or may not be physically separated, and the components shown as modules may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0074] In addition, each functional module in each embodiment of the present application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The above integrated unit can be implemented in the form of hardware or in the form of hardware plus software functional modules.

[0075] It is obvious to those skilled in the art that the present application is not limited to the details of the above exemplary embodiments, and that the present application can be implemented in other specific forms without departing from the spirit or essential features of the present application.

Claims

1. An intelligent question-answering system based on heterogeneous graph RAG technology, characterized in that: include: A heterogeneous graph model construction module, used to construct a heterogeneous graph model, wherein the heterogeneous graph model includes knowledge nodes and relationship edges between knowledge nodes; A question data collection module is used to obtain questions input by users; A question semantic encoding module, used for semantically encoding the question input by the user to obtain a question semantic encoding feature vector; A candidate knowledge node semantic encoding module, used for extracting candidate knowledge nodes from the heterogeneous graph model, and semantically encoding the candidate knowledge nodes to obtain a candidate knowledge node semantic encoding feature vector; A question-alternative knowledge point node semantic joint principal component fusion module is used to perform alternative knowledge-question semantic saliency query matching and fusion on the alternative knowledge node semantic encoding feature vector and the question semantic encoding feature vector to obtain a question-alternative knowledge point node semantic saliency fusion representation feature; A semantic matching module, used to obtain a semantic matching degree based on the semantically significant fusion representation features of the question-candidate knowledge point node, and return the candidate knowledge node based on the semantic matching degree; The question-candidate knowledge point node semantic joint principal component fusion module includes: A candidate knowledge node question semantic feature sample covariance calculation unit is used to calculate the sample covariance matrix of the candidate knowledge node semantic encoding feature vector and the question semantic encoding feature vector to obtain a candidate knowledge node semantic sample covariance matrix and a question semantic sample covariance matrix; A principal component feature extraction unit for the covariance matrix of candidate knowledge node question semantic samples, used for extracting principal component feature vectors based on matrix decomposition from the covariance matrix of candidate knowledge node semantic samples and the covariance matrix of question semantic samples to obtain a set of principal component feature vectors of candidate knowledge node semantics and a set of principal component feature vectors of question semantics; The candidate knowledge node question semantic principal component fusion unit is used to perform approximate query matching semantic association analysis on the set of candidate knowledge node semantic principal component feature vectors and the set of question semantic principal component feature vectors to obtain a question-alternative knowledge point node semantic joint principal component fusion representation vector as the question-alternative knowledge point node semantic significant fusion representation feature.

2. The intelligent question-answering system based on heterogeneous graph RAG technology according to claim 1 is characterized in that: The candidate knowledge node question semantic feature sample covariance calculation unit includes: A candidate knowledge node question semantics standardization subunit is used to standardize the candidate knowledge node semantic coding feature vector and the question semantic coding feature vector to obtain a standardized candidate knowledge node semantic coding feature vector and a standardized question semantic coding feature vector; The standardized candidate knowledge node question semantic sample covariance matrix calculation subunit is used to calculate the sample covariance matrix of the standardized candidate knowledge node semantic encoding feature vector and the standardized question semantic encoding feature vector to obtain the candidate knowledge node semantic sample covariance matrix and the question semantic sample covariance matrix.

3. The intelligent question-answering system based on heterogeneous graph RAG technology according to claim 2 is characterized in that: The candidate knowledge node question semantic standardization subunit is used to: Respectively calculating the mean and standard deviation of the semantic coding feature vector of the candidate knowledge node to obtain the mean of the semantic features of the candidate knowledge node and the standard deviation of the semantic features of the candidate knowledge node; after subtracting the semantic coding feature vector of the candidate knowledge node from the mean of the semantic features of the candidate knowledge node by position, performing positional division of the calculated semantic difference vector of the candidate knowledge node and the standard deviation of the semantic features of the candidate knowledge node to obtain the standardized semantic coding feature vector of the candidate knowledge node; The mean and standard deviation of the question semantic encoding feature vector are calculated respectively to obtain the question semantic feature mean and the question semantic feature standard deviation; after subtracting the question semantic encoding feature vector from the question semantic feature mean by position, the calculated question semantic difference vector is divided by the question semantic feature standard deviation by position to obtain the standardized question semantic encoding feature vector.

4. The intelligent question-answering system based on heterogeneous graph RAG technology according to claim 3 is characterized in that: The standardized candidate knowledge node question semantic sample covariance matrix calculation subunit is used to: After multiplying the transposed vector of the standardized candidate knowledge node semantic coding feature vector by the standardized candidate knowledge node semantic coding feature vector, the obtained standardized candidate knowledge node semantic association matrix is ​​divided by the value obtained by subtracting one from the length of the standardized candidate knowledge node semantic coding feature vector to obtain the candidate knowledge node semantic sample covariance matrix; After multiplying the transposed vector of the standardized question semantic encoding feature vector by the standardized question semantic encoding feature vector, the obtained standardized question semantic association matrix is ​​divided by the value obtained by subtracting one from the length of the standardized question semantic encoding feature vector by position to obtain the question semantic sample covariance matrix.

5. The intelligent question-answering system based on heterogeneous graph RAG technology according to claim 4 is characterized in that: The candidate knowledge node question semantic principal component fusion unit comprises: The candidate knowledge node-question semantic maximum approximation query subunit is used to input the set of candidate knowledge node semantic principal component feature vectors and the set of question semantic principal component feature vectors into the maximum approximation query matching network to obtain the set of best matching pairs of candidate knowledge node semantic principal component feature vectors and question semantic principal component feature vectors; The candidate knowledge node-question semantic principal component fusion subunit is used to input the best matching pairs of the candidate knowledge node semantic principal component feature vector and the question semantic principal component feature vector in the set of the best matching pairs of the candidate knowledge node semantic principal component feature vector and the question semantic principal component feature vector into the semantic fine-grained gating joint module to obtain a set of candidate knowledge node-question semantic principal component fusion feature vectors; The question-alternative knowledge point node semantic joint principal component fusion representation generation subunit is used to cascade the set of the alternative knowledge node-question semantic principal component fusion feature vectors to obtain the question-alternative knowledge point node semantic joint principal component fusion representation vector.

6. The intelligent question-answering system based on heterogeneous graph RAG technology according to claim 5 is characterized in that: The candidate knowledge node-problem semantic maximum approximation query subunit is used to: Extracting predetermined candidate knowledge node semantic principal component feature vectors from the set of candidate knowledge node semantic principal component feature vectors; Calculating the cosine similarity between the predetermined candidate knowledge node semantic principal component feature vector and each question semantic principal component feature vector in the set of question semantic principal component feature vectors to obtain a set of matching query similarities; The question semantic principal component feature vector corresponding to the largest matching query similarity in the set of matching query similarities and the predetermined candidate knowledge node semantic principal component feature vector are taken as the best matching pair of the predetermined candidate knowledge node semantic principal component feature vector and the question semantic principal component feature vector.

7. The intelligent question-answering system based on heterogeneous graph RAG technology according to claim 6 is characterized in that: The candidate knowledge node-question semantic principal component fusion subunit is used to: Respectively calculating the positional difference, positional dot product and positional addition between the best matching pair of the candidate knowledge node semantic principal component feature vector and the question semantic principal component feature vector to obtain a candidate knowledge node-question semantic principal component pattern difference vector, a candidate knowledge node-question semantic principal component pattern dot product vector and a candidate knowledge node-question semantic principal component pattern sum vector; The candidate knowledge node-question semantics principal component pattern difference vector, the candidate knowledge node-question semantics principal component pattern dot product vector and the candidate knowledge node-question semantics principal component pattern sum vector are cascaded and then one-dimensional convolutional coding is performed to obtain a candidate knowledge node-question semantics principal component pattern multi-dimensional fusion vector; The candidate knowledge node-question semantics principal component pattern multi-dimensional fusion vector is subjected to local window-based maximum pooling processing to obtain the candidate knowledge node-question semantics principal component fusion feature vector.

8. The intelligent question-answering system based on heterogeneous graph RAG technology according to claim 7 is characterized in that: The semantic matching module comprises: A semantic measurement unit, used for inputting the semantic joint principal component fusion representation vector of the question-candidate knowledge point node into a semantic measurement network to obtain the semantic matching degree; The candidate knowledge node feedback unit is used to return the candidate knowledge node in response to the semantic matching degree being greater than or equal to a preset threshold.

9. The intelligent question-answering system based on heterogeneous graph RAG technology according to claim 8, characterized in that: The semantic measurement unit is used to: input the question-candidate knowledge point node semantic joint principal component fusion representation vector into a decoder-based semantic measurement network to obtain the semantic matching degree.

Images