Knowledge graph-based large model answer accuracy enhancement method and system

Abstract

The invention discloses a knowledge graph-based large model answer accuracy enhancement method and system, and relates to the technical field of electric digital data processing. The large model answer accuracy enhancement method based on the knowledge graph comprises the following steps of answer accuracy judgment, graph updating adjustment and graph embedding optimization. According to the method, the answer accuracy of the to-be-evaluated model embedded with the preset knowledge graph is evaluated and judged to obtain the answer accuracy detection result, then whether the initial updating speed of the preset knowledge graph is adjusted or not is judged, and if the initial updating speed of the preset knowledge graph is adjusted, the preset knowledge graph is updated. If yes, the optimization updating speed of the preset knowledge graph is obtained, otherwise, knowledge graph embedding optimization is carried out on the preset knowledge graph in combination with the answer accuracy detection result, and the effect of enhancing the reasoning answer accuracy of the knowledge graph embedding large model is achieved; the problem that in the prior art, reasoning and answering accuracy is not high when a knowledge graph is embedded into a large model is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of digital data processing, and in particular to a method and system for enhancing the answer accuracy of large models based on knowledge graphs. Background Art

[0002] A knowledge graph (KG), as a graphical structure representing world knowledge through entities, relationships, and attributes, provides efficient knowledge representation and reasoning capabilities. The combination of a knowledge graph and the text generation ability of a large model can make up for the deficiencies of traditional large models in logical reasoning and knowledge application. As time goes by, the information in the knowledge graph will gradually become outdated, especially in fields involving rapid changes (such as technology, etc.), thus affecting the answer accuracy of the large model. Therefore, how to efficiently integrate a dynamically updated knowledge graph into a trained large model to avoid the "cold start" problem or poor timeliness requires new technical support. Entities and relationships in the knowledge graph usually need to be transformed into vector representations to be processed in a neural network. However, there are still great difficulties in representing these structured knowledge graphs as embedding vectors suitable for deep learning models. Existing graph embedding methods (such as TransE, DistMult, etc.) may not be able to fully capture the details and deep relationships in the graph when dealing with complex relationships and high-dimensional knowledge graphs. The knowledge graph provides important knowledge supplementation and reasoning capabilities for the large model, enabling the large model to be significantly improved in terms of knowledge application and generation accuracy. In the future, the integration of the large model and the knowledge graph will be closer, and more accurate and reliable intelligent services can be provided in a wider range of application scenarios.

[0003] Existing large model answers based on knowledge graphs usually rely on the structured knowledge of the knowledge graph, combined with entity recognition, relationship extraction, reasoning techniques, etc., and use knowledge graph embedding to enhance the reasoning ability of the large model. First, the question input by the user is processed, including text cleaning, word segmentation, named entity recognition, entity disambiguation, etc., and entities and relationships related to the question are obtained by querying the knowledge graph. If the question requires specific facts or reasoning, the query module can extract relevant information from the graph. Secondly, based on the input question and the information obtained from the graph, a reasonable answer is generated. The large model can reason based on the structured information in the graph and generate an answer consistent with the facts. Through techniques such as multi-modal data integration, context-aware query, and self-supervised learning, the model can dynamically adjust its reasoning process to provide accurate and in-depth answers. Combining the natural language generation ability of the large model with the structured knowledge of the knowledge graph can significantly improve the accuracy and logical consistency of the large model when answering questions.

[0004] For example, a method and system for enhancing the answer accuracy of a large model based on a knowledge graph disclosed in a patent application with publication number CN119271784A includes: obtaining a text data set, preprocessing the text data set, and extracting entities and relationships in the text data set; establishing an entity connection graph based on the entities and relationships in the text data set, performing fusion processing on the entity nodes of the entity connection graph, and creating a knowledge graph according to the entity nodes and relationships after the fusion processing; obtaining a user's query question, and determining a matching node of the user's query question according to the user's query question; identifying the user's question intention according to the user's query question, determining a prompt template according to the user's question intention and the matching node corresponding to the user's query question, inputting the prompt template into the large model, and obtaining a reliable answer to the user's query question.

[0005] For example, a question-answering method and system based on a knowledge graph and a large language model disclosed in the invention patent announcement with announcement number CN116775847B includes: receiving an intelligence question to be processed, extracting entity features to obtain initial feature entities; performing a vector space search in a vector database to determine extended entities; querying a self-established graph database based on the initial feature entities, constructing query auxiliary information to generate an initial intermediate answer; determining whether a new entity appears in the initial intermediate answer to determine whether to perform the next round of entity feature extraction. Each time a round of entity feature extraction is performed, an intermediate answer is generated until it is determined that no new entity appears in the intermediate answer; performing deduplication and fusion on all intermediate answers after one or more rounds of entity feature extraction to obtain a final answer matching the intelligence question to be processed.

[0006] However, in the process of implementing the technical solution of the invention in the embodiments of the present application, it is found that the above technologies have at least the following technical problems:

[0007] When embedding a knowledge graph into a large model, rich semantic information, context relationships, and complex structures of the graph in the knowledge graph may be simplified into relatively abstract low-dimensional vectors, that is, converting entities and relationships in the knowledge graph into vector representations. However, the graph structure in the knowledge graph is often very sparse, and the types of nodes and edges are also very diverse, which often makes it possible to lose some details when converting graph information into vector representations. Especially in complex queries, this lost information may lead to inaccurate or incomplete answers.

[0008] In addition, the update speed of the knowledge graph is relatively slow, and it may contain outdated or inaccurate information, resulting in inaccurate answers from the large model. There is a problem that the accuracy of reasoning and answering when embedding the knowledge graph into the large model is not high. Summary of the Invention

[0009] Embodiments of this application provide a method and system for enhancing the answer accuracy of a large model based on a knowledge graph, which solves the problem of low accuracy in reasoning and answering when a knowledge graph is embedded in a large model in the prior art, and realizes the enhancement of the accuracy of reasoning and answering when a knowledge graph is embedded in a large model.

[0010] Embodiments of this application provide a method for enhancing the answer accuracy of a large model based on a knowledge graph, including the following steps: Step 1, evaluate and judge the answer accuracy of the model to be evaluated after embedding a preset knowledge graph, and obtain an answer accuracy detection result; Step 2, combine the answer accuracy detection result to judge whether to adjust the initial update speed of the preset knowledge graph. If the initial update speed of the preset knowledge graph is adjusted, then based on the answer accuracy detection result and the initial update speed of the preset knowledge graph, obtain the optimized update speed of the preset knowledge graph. The initial update speed adjustment means optimizing the initial update speed of the preset knowledge graph based on the answer accuracy detection result to improve the answer accuracy of the model to be evaluated; Step 3, if the initial update speed of the preset knowledge graph is not adjusted, then combine the answer accuracy detection result to perform knowledge graph embedding optimization on the preset knowledge graph. The knowledge graph embedding optimization is used to improve the embedding quality of the preset knowledge graph embedded in the model to be evaluated to enhance the answer accuracy of the model to be evaluated.

[0011] Further, the specific steps for evaluating and judging the answer accuracy of the model to be evaluated after embedding the preset knowledge graph are as follows: Obtain the quantization data of the embedding quality after the preset knowledge graph is embedded in the model to be evaluated. The embedding quality quantization data includes embedding vector similarity, embedding global clustering coefficient, embedding space density, and embedding reconstruction error; Perform parameter interaction processing on the difference degree of the embedding global clustering coefficient and the embedding vector similarity, and the difference degree of the embedding space density and the embedding reconstruction error respectively. After weighting the corresponding parameter interaction processing results through the embedding quality influence weight, couple them to obtain an embedding quality influence factor; The embedding quality influence weight includes an embedding tightness influence weight and an embedding accuracy influence weight; The embedding quality influence factor is used to describe the influence degree of the embedding quality of the preset knowledge graph embedded in the model to be evaluated on the answer accuracy determination value; The embedding quality influence factor represents the quantization data of the joint influence degree of the embedding quality quantization data on the embedding quality of the preset knowledge graph embedded in the model to be evaluated on the answer accuracy determination value; Obtain the answer accuracy rate of the model to be evaluated, and perform an embedding quality correlation operation on the answer accuracy rate according to the obtained embedding quality influence factor to obtain an answer accuracy determination value. The answer accuracy determination value is used to quantitatively evaluate the corresponding answer accuracy degree after the preset knowledge graph is embedded in the model to be evaluated; The answer accuracy determination value represents the quantization data of the joint influence of the embedding quality influence factor and the answer accuracy rate on the corresponding answer accuracy after the preset knowledge graph is embedded in the model to be evaluated.

[0012] Further, the specific steps for obtaining the answer accuracy detection result are as follows: Combine the obtained answer accuracy determination value with the minimum answer accuracy threshold obtained from the preset database for comparison and judgment; if the answer accuracy determination value is greater than the minimum answer accuracy threshold, record the answer accuracy detection result as the answer accuracy determination value being qualified and the answer accuracy of the model to be evaluated being to be monitored; if the answer accuracy is not greater than the minimum answer accuracy threshold, record the answer accuracy detection result as the answer accuracy determination value being unqualified and the answer accuracy of the model to be evaluated being to be enhanced.

[0013] Further, the specific process of combining the answer accuracy detection result to determine whether to adjust the initial update speed of the preset knowledge graph is as follows: If the answer accuracy detection result is that the answer accuracy determination value is qualified and the answer accuracy of the model to be evaluated is to be monitored, then adjust the initial update speed of the preset knowledge graph; otherwise, do not adjust the initial update speed of the preset knowledge graph.

[0014] Further, the specific steps for adjusting the initial update speed of the preset knowledge graph are as follows: Map from the preset accurate deviation mapping set in the preset database according to the deviation degree between the answer accuracy determination value in the answer accuracy detection result and the maximum answer accuracy threshold to obtain a graph speed update coefficient, where the preset accurate deviation mapping set represents the mapping relationship between the answer accuracy determination value and the maximum answer accuracy threshold; Obtain the initial update speed of the preset knowledge graph and the graph speed update coefficient, and use the graph speed update coefficient to adjust the initial update speed to obtain an optimized update speed.

[0015] Further, the specific process of optimizing the knowledge graph embedding of the preset knowledge graph by combining the answer accuracy detection result is as follows: S31, when the answer accuracy in the answer accuracy detection result is not greater than the minimum answer accuracy threshold, perform graph community detection and division on the preset knowledge graph to obtain the community density of the graph community; S32, optimize the knowledge graph compression embedding of the preset knowledge graph by combining the community density of the graph community, where the knowledge graph compression embedding optimization refers to obtaining a graph compression rate through the graph compression strength to perform knowledge graph compression; S33, after performing the knowledge graph compression embedding optimization, perform hierarchical embedding optimization on the knowledge graph, where the hierarchical embedding optimization refers to performing hierarchical embedding of the knowledge graph based on the obtained hierarchical embedding priority score, and the hierarchical embedding priority score is used to judge the hierarchical embedding priority during the process of performing hierarchical embedding of the knowledge graph.

[0016] Furthermore, obtaining the graph compression rate based on the graph compression strength to perform knowledge graph compression is specifically as follows: Map based on the deviation degree between the community density of the graph community and the maximum community density threshold in the preset strength allocation mapping set in the preset database to obtain the graph compression strength of the graph community, where the preset strength allocation mapping set represents the mapping relationship between the community density and the maximum community density threshold; Map the deviation degree between the obtained graph compression strength of the graph community and the compression strength reference value from the preset compression strength mapping set in the preset database to obtain the graph compression rate of the graph community, where the preset compression strength mapping set represents the mapping relationship between the graph compression strength and the compression strength reference value; Perform knowledge graph compression on the graph community according to the obtained graph compression rate.

[0017] Furthermore, performing knowledge graph hierarchical embedding based on the obtained hierarchical embedding priority score is specifically as follows: Quantify the information retention degree of the graph community for knowledge graph compression by the graph compression data after knowledge graph compression based on the graph community to obtain a graph compression determination value; Map the deviation degree between the obtained graph compression determination value and the graph compression reference value from the preset database to obtain a compression quality score weight; Weight the graph compression determination value through the obtained compression quality score weight, and couple and weight the graph compression rate through the compression rate score weight obtained from the preset database to obtain a hierarchical embedding priority score; Combine the hierarchical embedding priority scores of the graph communities and sort them in descending order to obtain the hierarchical embedding priority level, and perform knowledge graph hierarchical embedding according to the hierarchical embedding priority level.

[0018] Furthermore, quantifying the information retention degree of the graph community for knowledge graph compression by the graph compression data after knowledge graph compression based on the graph community to obtain a graph compression determination value is specifically as follows: Obtain the graph compression data after knowledge graph compression of the graph community, where the graph compression data includes graph density, node retention degree, edge retention degree, and information loss degree; Perform parameter interaction processing on the coupling result of the node retention degree and the edge retention degree through the result of the graph density difference degree, and couple and weight the zeroing operation result of the information loss degree through the compression evaluation weight to obtain a graph compression determination value; The compression evaluation weight includes a compression retention evaluation weight and a compression loss evaluation weight; The graph compression determination value represents the quantitative data for jointly evaluating the information retention degree of the graph community for knowledge graph compression by the graph compression data.

[0019] The embodiment of the present application provides a system for enhancing the answer accuracy of a large model based on a knowledge graph, including: an answer accuracy judgment module, a graph update adjustment module, and a graph embedding optimization module; the answer accuracy judgment module is used to evaluate and judge the answer accuracy of the model to be evaluated after being embedded in a preset knowledge graph, and obtain an answer accuracy detection result; the graph update adjustment module is used to judge whether to adjust the initial update speed of the preset knowledge graph in combination with the answer accuracy detection result. If the initial update speed of the preset knowledge graph is adjusted, the optimized update speed of the preset knowledge graph is obtained based on the answer accuracy detection result and the initial update speed of the preset knowledge graph. The initial update speed adjustment means optimizing the initial update speed of the preset knowledge graph based on the answer accuracy detection result to improve the answer accuracy of the model to be evaluated; the graph embedding optimization module is used to, if the initial update speed of the preset knowledge graph is not adjusted, perform knowledge graph embedding optimization on the preset knowledge graph in combination with the answer accuracy detection result. The knowledge graph embedding optimization is used to improve the embedding quality of the preset knowledge graph embedded in the model to be evaluated to enhance the answer accuracy of the model to be evaluated.

[0020] One or more technical solutions provided in the embodiments of the present application have at least the following technical effects or advantages: 1. By evaluating and judging the answer accuracy of the model to be evaluated after being embedded in a preset knowledge graph to obtain an answer accuracy detection result, then judging whether to adjust the initial update speed of the preset knowledge graph. If the initial update speed of the preset knowledge graph is adjusted, the optimized update speed of the preset knowledge graph is obtained. Otherwise, the knowledge graph embedding of the preset knowledge graph is optimized in combination with the answer accuracy detection result, thereby realizing the improvement of the embedding quality of the knowledge graph embedded in the large model, and further realizing the enhancement of the inference answer accuracy of the knowledge graph embedded in the large model, effectively solving the problem of low inference answer accuracy of the knowledge graph embedded in the large model in the prior art.

[0021] 2. By mapping the deviation degree between the answer accuracy determination value in the answer accuracy detection result and the maximum answer accuracy threshold from the preset accurate deviation mapping set in the preset database, a graph speed update coefficient is obtained, and then the initial update speed is adjusted with the graph speed update coefficient to obtain the optimized update speed, thereby realizing the optimization of the initial update speed of the preset knowledge graph based on the answer accuracy detection result, and further realizing the enhancement of the inference answer accuracy of the large model through knowledge graph update optimization.

[0022] 3. The deviation degree between the obtained graph compression determination value and the graph compression reference value is mapped from a preset database to obtain a compression quality scoring weight. Then, the graph compression determination value is weighted based on the obtained compression quality scoring weight, and the graph compression rate is weighted by combining the compression rate scoring weight obtained from the preset database. After coupling, a hierarchical embedding priority score is obtained. Finally, knowledge graph hierarchical embedding is performed according to the hierarchical embedding priority, thereby improving the reliability of knowledge graph hierarchical embedding into the large model, and further enhancing the accuracy of the large model's reasoning answer through knowledge graph embedding optimization. Description of the Drawings

[0023] Figure 1 It is a flowchart of the method for enhancing the answer accuracy of a large model based on a knowledge graph provided by an embodiment of the present application; Figure 2 It is a schematic structural diagram of the system for enhancing the answer accuracy of a large model based on a knowledge graph provided by an embodiment of the present application. Detailed Embodiment

[0024] In the embodiments of the present application, by providing a method and system for enhancing the answer accuracy of a large model based on a knowledge graph, the problem that the accuracy of reasoning answers is not high when the knowledge graph is embedded in the large model in the prior art is solved. The answer accuracy of the model to be evaluated after embedding the preset knowledge graph is evaluated and judged to obtain an answer accuracy detection result. Then, whether to adjust the initial update speed of the preset knowledge graph is judged in combination with the answer accuracy detection result. If the initial update speed of the preset knowledge graph is adjusted, the optimized update speed of the preset knowledge graph is obtained based on the answer accuracy detection result and the initial update speed of the preset knowledge graph. If the initial update speed of the preset knowledge graph is not adjusted, the knowledge graph embedding optimization of the preset knowledge graph is performed in combination with the answer accuracy detection result, realizing the enhancement of the accuracy of reasoning answers when the knowledge graph is embedded in the large model.

[0025] The technical solutions in the embodiments of the present application are used to solve the problem that the accuracy of reasoning answers is not high when the knowledge graph is embedded in the large model. The overall idea is as follows: By obtaining the answer accuracy detection result to judge whether to adjust the initial update speed of the preset knowledge graph. If the initial update speed of the preset knowledge graph is adjusted, the optimized update speed of the preset knowledge graph is obtained. Otherwise, the knowledge graph embedding optimization of the preset knowledge graph is performed, achieving the effect of enhancing the accuracy of reasoning answers when the knowledge graph is embedded in the large model.

[0026] To better understand the above technical solutions, the above technical solutions will be described in detail below in combination with the accompanying drawings of the specification and specific embodiments.

[0027] Such asFigure 1 As shown in the figure, it is a flowchart of a method for enhancing the answer accuracy of a large model based on a knowledge graph provided by an embodiment of the present application. The method includes the following steps: Step 1, evaluate and judge the answer accuracy of the model to be evaluated after embedding a preset knowledge graph, and obtain an answer accuracy detection result; Step 2, combine the answer accuracy detection result to judge whether to adjust the initial update speed of the preset knowledge graph. If the initial update speed of the preset knowledge graph is adjusted, then obtain the optimized update speed of the preset knowledge graph based on the answer accuracy detection result and the initial update speed of the preset knowledge graph. The initial update speed adjustment means optimizing the initial update speed of the preset knowledge graph based on the answer accuracy detection result to improve the answer accuracy of the model to be evaluated; Step 3, if the initial update speed of the preset knowledge graph is not adjusted, then perform knowledge graph embedding optimization on the preset knowledge graph in combination with the answer accuracy detection result. The knowledge graph embedding optimization is used to improve the embedding quality of the preset knowledge graph embedded in the model to be evaluated to enhance the answer accuracy of the model to be evaluated.

[0028] In this embodiment, the large model to be evaluated is specifically, for example, the GPT series (such as GPT-4), which can be used for tasks such as text generation, translation, and question answering; a knowledge graph is a way of representing entities (such as people, places, things, etc.) and their relationships as a graphical structure. By constructing a knowledge network through nodes (representing entities) and edges (representing relationships between entities), the large model to be evaluated (such as GPT-4) can understand and process the connections between information; the knowledge graph can provide structured and explicit knowledge about the world. Over time, the information in the knowledge graph may become outdated or inaccurate, especially in rapidly developing fields. Therefore, the update of the knowledge graph is very important. At the same time, the large model to be evaluated has advantages in language generation and reasoning capabilities. The combination of the two can make up for their respective shortcomings and improve reasoning capabilities and answer accuracy. By converting the knowledge graph into a format that can be input into the large model to be evaluated, for example, using graph embedding methods (such as TransE, DistMult, etc.) to map the entities and relationships in the knowledge graph to a low-dimensional space, the information in the knowledge graph can be introduced as background knowledge into the reasoning process of the large model to be evaluated, thereby helping the model make more reasonable judgments when facing some reasoning problems without direct text support.

[0029] Although embedding knowledge graphs into large models to be evaluated to enhance reasoning ability has become a popular direction, it still faces the problem of low quality of knowledge graphs, which directly affects the accuracy of reasoning answers of the large models to be evaluated. During the reasoning process, the large models to be evaluated often need to extract information from the knowledge graph and combine it with natural language text information. When the embedding quality of the knowledge graph is low, the smoothness of information flow may be hindered. The large models to be evaluated often need multi-step reasoning to generate answers. If the embedded information in the graph is missing or inaccurate, the reasoning chain will break, resulting in inaccurate final answers. The algorithm of this application optimizes the knowledge graph embedding and the knowledge graph update speed to further improve the embedding quality of the knowledge graph embedded in the large model to be evaluated (taking GPT-4 as an example), so that the large model to be evaluated can more accurately understand and utilize structured knowledge, improve the coherence and correctness of reasoning, and thus provide more accurate and reliable answers.

[0030] Further, evaluate and judge the answer accuracy of the model to be evaluated after embedding the preset knowledge graph. The specific steps are as follows: Obtain the quantitative data of the embedding quality after embedding the preset knowledge graph into the model to be evaluated; Perform parameter interaction processing on the difference degree of the global clustering coefficient of the embedding, the similarity of the embedding vectors, the difference degree of the embedding space density, and the embedding reconstruction error respectively. After weighting the corresponding parameter interaction results by the embedding quality influence weight, couple them to obtain the embedding quality influence factor.

[0031] The method for obtaining the embedding quality influence factor is as follows: ; In the formula, represents the embedding quality influence factor, represents the embedding tightness influence weight, represents the embedding accuracy influence weight, represents the global clustering coefficient of the embedding, represents the reference minimum clustering coefficient, represents the similarity of the embedding vectors, represents the embedding space density, represents the reference embedding space density, represents the embedding reconstruction error.

[0032] It should be added that the parameter interaction processing is used to describe the interaction relationship between the two parameters involved; among them, the embedded quality quantization data includes the embedded vector similarity, the embedded global clustering coefficient, the embedded space density, and the embedded reconstruction error; specifically, the embedded vector similarity is obtained through the sklearn library in Python; the embedded global clustering coefficient is obtained through NetworkX in Python; the embedded space density is obtained through the scikit-learn library in Python; the embedded reconstruction error is obtained through TensorFlow.

[0033] The minimum value of the collected historical embedded global clustering coefficient represents the reference minimum clustering coefficient, and the result of summing and averaging the collected historical embedded space density represents the reference embedded space density.

[0034] The embedded quality influence weights include the embedded tightness influence weight and the embedded accuracy influence weight; the embedded tightness influence weight and the embedded accuracy influence weight respectively correspond to describing the influence degree of the embedded tightness situation (that is, the result of parameter interaction processing of the difference degree of the embedded global clustering coefficient and the embedded vector similarity) and the embedded accuracy situation (that is, the result of parameter interaction processing of the difference degree of the embedded space density and the embedded reconstruction error) on the embedded quality influence factor, and the value range is [0, 1] and the sum is 1. For example, the real-time embedded tightness situation and the embedded accuracy situation are input into the mapping set of the preset embedded tightness situation, the embedded accuracy situation and their respective corresponding weight factors in the database to obtain the corresponding weights.

[0035] Obtain the answer accuracy rate of the model to be evaluated, and perform an embedded quality correlation operation on the answer accuracy rate according to the obtained embedded quality influence factor to obtain the answer accuracy determination value. Among them, the answer accuracy rate is the ratio of the number of times the model to be evaluated answers correctly to the total number of answers, and the answer accuracy rate is obtained through manual annotation.

[0036] The method for obtaining the answer accuracy determination value is as follows: ; In the formula, represents the answer accuracy determination value, represents the embedded quality influence factor, represents the answer accuracy rate.

[0037] It should be understood that the answer accuracy determination value is used to quantitatively evaluate the corresponding answer accuracy degree after the preset knowledge graph is embedded in the model to be evaluated; among them, as the answer accuracy rate increases, the answer accuracy determination value increases accordingly. In addition, the answer accuracy determination value includes parameters in multiple aspects, representing the quantitative data of the combined influence of the embedded quality influence factor and the answer accuracy rate on the answer accuracy after the preset knowledge graph is embedded in the model to be evaluated.

[0038] It should also be understood that the embedding quality impact factor represents the quantitative data of the combined impact of the embedding quality quantization data on the embedding quality of the preset knowledge graph embedding the model to be evaluated on the answer accuracy determination value, and is used to describe the impact degree of the embedding quality of the preset knowledge graph embedding the model to be evaluated on the answer accuracy determination value; the embedding quality impact factor contains various parameters, and there are connections between the parameters and they do not exist independently. For example, with the increase of the embedding global clustering coefficient, it usually means that the entity and relationship of the knowledge graph are more closely organized in the embedding space, with better clustering, making the embedding vectors of similar entities or relationships in the knowledge graph closer, improving the similarity of the vectors, that is, the embedding vector similarity increases accordingly; if the embedding vector can well capture the relationship between similar entities, the model to be evaluated will better retain the global clustering characteristics of the knowledge graph; at the same time, with the increase of the embedding space density, it indicates that the high-density embedding space can provide more information, thereby reducing the reconstruction error, that is, the embedding reconstruction error decreases accordingly.

[0039] Through the numerical evaluation of the corresponding answer accuracy degree after the preset knowledge graph is embedded into the model to be evaluated, a more accurate evaluation of the answer accuracy degree of the model to be evaluated is realized, and further, the reliability of the evaluation of the corresponding answer accuracy degree after the preset knowledge graph is embedded into the model to be evaluated is improved.

[0040] Further, the specific steps to obtain the answer accuracy detection result are as follows: Combine the obtained answer accuracy determination value with the minimum answer accuracy threshold obtained from the preset database for comparison and judgment; if the answer accuracy determination value is greater than the minimum answer accuracy threshold, record the answer accuracy detection result as the answer accuracy determination value is qualified and the answer accuracy of the model to be evaluated is to be monitored; if the answer accuracy is not greater than the minimum answer accuracy threshold, record the answer accuracy detection result as the answer accuracy determination value is unqualified and the answer accuracy of the model to be evaluated is to be enhanced.

[0041] In this embodiment, the minimum answer accuracy threshold is set by professionals according to the standards in the field. For example, the minimum answer accuracy threshold is set to 0.5; by combining the minimum answer accuracy threshold with the obtained answer accuracy determination value for comparison and judgment to obtain the answer accuracy detection result, a more accurate judgment of the corresponding answer accuracy degree after the preset knowledge graph is embedded into the model to be evaluated is realized.

[0042] Further, combine the answer accuracy detection result to judge whether to adjust the initial update speed of the preset knowledge graph. The specific process is as follows: If the answer accuracy detection result is that the answer accuracy determination value is qualified and the answer accuracy of the model to be evaluated is to be monitored, then adjust the initial update speed of the preset knowledge graph; otherwise, do not adjust the initial update speed of the preset knowledge graph.

[0043] It should be added that the initial update speed of the preset knowledge graph is adjusted, and the specific steps are as follows: according to the deviation degree between the answer accuracy judgment value in the answer accuracy detection result and the maximum answer accuracy threshold, a preset accurate deviation mapping set is mapped from the preset database to obtain a graph speed update coefficient, and the preset accurate deviation mapping set represents the mapping relationship between the answer accuracy judgment value and the maximum answer accuracy threshold; the initial update speed and graph speed update coefficient of the preset knowledge graph are obtained, and the initial update speed is adjusted with the graph speed update coefficient to obtain the optimized update speed.

[0044] In this embodiment, the maximum answer accuracy threshold is set by professionals according to standards in the field. For example, the maximum answer accuracy threshold is set to 1; a preset accurate deviation mapping set of answer accuracy judgment values ​​and maximum answer accuracy thresholds is constructed, and the real-time answer accuracy judgment value is input into the preset accurate deviation mapping set to obtain the corresponding graph speed update coefficient; by adjusting the initial update speed of the preset knowledge graph, the knowledge graph update is embedded in the large model to be evaluated to enhance the accuracy of reasoning answers.

[0045] Furthermore, the specific process of performing knowledge graph embedding optimization on the preset knowledge graph in combination with the answer accuracy detection result is as follows: S31, when the answer accuracy in the answer accuracy detection result is not greater than the minimum answer accuracy threshold, performing graph community detection and division on the preset knowledge graph to obtain the community density of the graph community; S32, performing knowledge graph compression embedding optimization on the preset knowledge graph in combination with the community density of the graph community. The knowledge graph compression embedding optimization refers to obtaining the graph compression rate through the graph compression strength to perform knowledge graph compression; S33. After performing the knowledge graph compression embedding optimization, performing hierarchical embedding optimization on the knowledge graph. The hierarchical embedding optimization refers to performing hierarchical embedding of the knowledge graph based on the obtained hierarchical embedding priority score. The hierarchical embedding priority score is used to determine the hierarchical embedding priority in the process of hierarchical embedding of the knowledge graph.

[0046] In this embodiment, a community detection algorithm (such as Louvain algorithm) is used to perform graph community detection and division on the preset knowledge graph, and the community density of the graph community is obtained through the Density provided by Gephi; the preset knowledge graph is embedded in the knowledge graph and optimized by combining the answer accuracy detection result, so that when the answer accuracy detection result is that the answer accuracy judgment value is unqualified and the answer accuracy of the model to be evaluated needs to be enhanced, the knowledge graph embedding optimization is performed to enhance the answer accuracy of the large model to be evaluated.

[0047] Further, the graph compression rate is obtained through the graph compression intensity to perform knowledge graph compression. The specific steps are as follows: Map based on the deviation degree between the community density of the graph community and the maximum community density threshold in the preset intensity allocation mapping set in the preset database to obtain the graph compression intensity of the graph community. The preset intensity allocation mapping set represents the mapping relationship between the community density and the maximum community density threshold; Map the deviation degree between the obtained graph compression intensity of the graph community and the compression intensity reference value from the preset compression intensity mapping set in the preset database to obtain the graph compression rate of the graph community. The preset compression intensity mapping set represents the mapping relationship between the graph compression intensity and the compression intensity reference value; Perform knowledge graph compression on the graph community according to the obtained graph compression rate.

[0048] In this embodiment, a preset intensity allocation mapping set between the community density and the maximum community density threshold is constructed, and the real-time community density is input into the preset intensity allocation mapping set to obtain the corresponding graph compression intensity; A preset compression intensity mapping set between the graph compression intensity and the compression intensity reference value is constructed, and the real-time graph compression intensity is input into the preset compression intensity mapping set to obtain the corresponding graph compression rate; Perform knowledge graph compression on the graph community through a knowledge graph compression algorithm (such as GraphSAGE, Graph Sample and Aggregation); The graph compression rate is obtained through the graph compression intensity to perform knowledge graph compression, which realizes the enhancement of the inference answer accuracy of the large model to be evaluated by optimizing the knowledge graph compression during the embedding of the knowledge graph into the large model to be evaluated.

[0049] Further, perform hierarchical embedding of the knowledge graph based on the obtained hierarchical embedding priority score. The specific steps are as follows: Quantify the information retention degree of the graph community for knowledge graph compression in the graph compression data after knowledge graph compression based on the graph community to obtain a graph compression determination value.

[0050] Specifically, quantify the information retention degree of the graph community for knowledge graph compression in the graph compression data after knowledge graph compression based on the graph community to obtain a graph compression determination value. The specific steps are as follows: Obtain the graph compression data after knowledge graph compression of the graph community; Perform parameter interaction processing on the coupling result of the node retention degree and the edge retention degree through the result of the graph density difference degree, and perform weighted coupling on the result of zeroing the information loss degree operation in combination with the compression evaluation weight to obtain the graph compression determination value.

[0051] The method for obtaining the graph compression determination value is as follows: ; In the formula, represents the graph compression determination value, represents the compression retention evaluation weight, represents the compression loss evaluation weight, represents the graph density, represents the reference graph density, represents the node retention degree, represents the edge retention degree, represents the information loss degree.

[0052] It should be added that the graph compression data includes the graph density, node retention degree, edge retention degree, and information loss degree; the graph density is obtained through NetworkX in Python; the node retention degree and edge retention degree are obtained through Gephi; the information loss degree is obtained through GraphMatching; the reference graph density is represented by the result of summing and averaging the collected historical graph densities.

[0053] The compression evaluation weights include the compression retention evaluation weight and the compression loss evaluation weight; the compression retention evaluation weight and the compression loss evaluation weight respectively correspond to describing the compression retention situation (i.e., the result of parameter interaction processing of the coupling result of the graph density difference degree to the node retention degree and the edge retention degree) and the influence degree of the information loss degree on the graph compression determination value. For example, the real-time compression retention situation and the information loss degree are input into the mapping set of the preset compression retention situation, information loss degree, and their respective weight factors in the database to obtain the corresponding weights.

[0054] It should be understood that the graph compression determination value represents the quantitative data for evaluating the information retention degree of knowledge graph compression of the graph community by the graph compression data. Among them, the graph compression determination value contains parameters in multiple aspects. Specifically, as the graph density, node retention degree, and edge retention degree increase, the graph compression determination value increases accordingly. As the information loss degree decreases, the graph compression determination value also increases accordingly.

[0055] In addition, the various parameters in the graph compression decision value are related and do not exist independently. This algorithm takes into account the correlation and mutual influence among the parameters and conducts comprehensive analysis through quantification. For example, the graph density is used to reflect the overall distribution of nodes and edges in the knowledge graph. If most nodes and edges are retained during the compression process, it means that the structure of the graph is relatively dense, that is, the graph density increases accordingly. Similarly, if a large number of nodes or edges are deleted during the compression process, it may lead to the sparsification of the graph, that is, the graph density decreases accordingly. At the same time, during the compression process, if a large number of nodes are retained, it is usually necessary to retain the edges between the nodes to maintain the structural integrity of the graph. Therefore, there is a certain mutual influence between the edge retention degree and the node retention degree. When compressing the graph, retaining more nodes may also retain more edges, that is, as the node retention degree increases, the edge retention degree also increases. As the node retention degree and the edge retention degree increase, it indicates that more original nodes and edges are retained during the compression process, usually better retaining the structural information and connectivity of the graph, thereby reducing information loss, that is, the information loss degree decreases accordingly.

[0056] The deviation degree between the obtained graph compression decision value and the graph compression reference value is mapped from the preset database to obtain the compression quality scoring weight; the obtained compression quality scoring weight is used to weight the graph compression decision value, and after weighting the graph compression rate with the compression rate scoring weight obtained from the preset database, they are coupled to obtain the hierarchical embedding priority score.

[0057] The method for obtaining the hierarchical embedding priority score is as follows: ; In the formula, represents the hierarchical embedding priority score, represents the compression quality scoring weight, represents the compression rate scoring weight, represents the graph compression decision value, represents the graph compression rate.

[0058] Combining the hierarchical embedding priority scores of the graph communities and sorting them in descending order to obtain the hierarchical embedding priorities, and performing hierarchical embedding of the knowledge graph according to the hierarchical embedding priorities.

[0059] Among them, the maximum value of the collected historical graph compression decision values represents the graph compression reference value. A mapping set between the graph compression decision value and the graph compression reference value is constructed, and the real-time graph compression decision value is input into the mapping set to obtain the corresponding compression quality scoring weight; the compression rate scoring weight is used to describe the influence degree of the graph compression rate on the hierarchical embedding priority score. For example, the real-time graph compression rate is input into the mapping set of the preset graph compression rate and its corresponding weight factor in the database to obtain the corresponding weight.

[0060] Hierarchical embedding of the knowledge graph is achieved by combining the hierarchical embedding priorities obtained from the hierarchical embedding priority scoring of the graph community, which improves the reliability of hierarchical embedding in the process of embedding the knowledge graph into the large model to be evaluated. Furthermore, the improvement of the embedding quality of the knowledge graph enhances the accuracy of the inference answers of the large model to be evaluated.

[0061] As Figure 2 shown, it is a schematic structural diagram of the large model answer accuracy enhancement system based on the knowledge graph provided by the embodiment of the present application. The large model answer accuracy enhancement system based on the knowledge graph provided by the embodiment of the present application includes: an answer accuracy judgment module, a graph update adjustment module, and a graph embedding optimization module; the answer accuracy judgment module is used to evaluate and judge the answer accuracy of the model to be evaluated after embedding the preset knowledge graph, and obtain the answer accuracy detection result; the graph update adjustment module is used to judge whether to adjust the initial update speed of the preset knowledge graph in combination with the answer accuracy detection result. If the initial update speed of the preset knowledge graph is adjusted, the optimized update speed of the preset knowledge graph is obtained based on the answer accuracy detection result and the initial update speed of the preset knowledge graph. The initial update speed adjustment means optimizing the initial update speed of the preset knowledge graph based on the answer accuracy detection result to improve the answer accuracy of the model to be evaluated; the graph embedding optimization module is used to, if the initial update speed of the preset knowledge graph is not adjusted, perform knowledge graph embedding optimization on the preset knowledge graph in combination with the answer accuracy detection result. The knowledge graph embedding optimization is used to improve the embedding quality of the preset knowledge graph embedded into the model to be evaluated to enhance the answer accuracy of the model to be evaluated.

[0062] In summary, in the embodiment of the present application, the answer accuracy of the model to be evaluated after embedding the preset knowledge graph is evaluated and judged to obtain the answer accuracy detection result, and then it is judged whether to adjust the initial update speed of the preset knowledge graph. If the initial update speed of the preset knowledge graph is adjusted, the optimized update speed of the preset knowledge graph is obtained. Otherwise, the knowledge graph embedding optimization of the preset knowledge graph is performed in combination with the answer accuracy detection result, thereby realizing the improvement of the embedding quality of the knowledge graph embedded in the large model, and further realizing the enhancement of the inference answer accuracy of the knowledge graph embedded in the large model, effectively solving the problem of low inference answer accuracy of the knowledge graph embedded in the large model in the prior art.

[0063] Those skilled in the art will appreciate that the embodiments of the present invention may be provided as a method, a system, or a computer program product. Therefore, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code.

[0064] The present invention is described with reference to the flowcharts and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the present invention. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and combinations of flows and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to produce a machine, such that the instructions executed by the processor of the computer or other programmable data processing devices produce means for implementing the functions specified in Figure 1 one or more of the flows Figure 1 or blocks or combinations of blocks.

[0065] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a particular manner, such that the instructions stored in the computer-readable memory produce a manufacture including instruction means that implement the functions specified in Figure 1 one or more of the flows Figure 1 or blocks or combinations of blocks.

[0066] These computer program instructions can also be loaded onto a computer or other programmable data processing device, such that a series of operation steps are performed on the computer or other programmable device to produce a computer-implemented process, and thus the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in Figure 1 one or more of the flows Figure 1 or blocks or combinations of blocks.

[0067] Although the preferred embodiments of the present invention have been described, those skilled in the art can make additional changes and modifications once they learn the basic creative concepts. Therefore, the appended claims are intended to be construed to include the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0068] Obviously, those skilled in the art can make various modifications and variations to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalent technologies, the present invention is also intended to include these modifications and variations.

Claims

1. A large model answer accuracy enhancement method based on knowledge graph, characterized in that: The following steps are involved: Step 1: Evaluate and judge the answer accuracy of the model to be evaluated after embedding the preset knowledge graph, and obtain the answer accuracy test result; Step 2: Determine whether to adjust the initial update speed of the preset knowledge graph based on the answer accuracy test result. If the initial update speed of the preset knowledge graph is adjusted, the optimized update speed of the preset knowledge graph is obtained based on the answer accuracy test result and the initial update speed of the preset knowledge graph. The initial update speed adjustment means optimizing the initial update speed of the preset knowledge graph based on the answer accuracy test result to improve the answer accuracy of the model to be evaluated. Step three, if the initial update speed of the preset knowledge graph is not adjusted, the preset knowledge graph is optimized for embedding in combination with the answer accuracy detection result. The knowledge graph embedding optimization is used to improve the embedding quality of the preset knowledge graph embedded in the model to be evaluated to enhance the answer accuracy of the model to be evaluated.

2. The method for enhancing the accuracy of large model answers based on knowledge graph as claimed in claim 1, characterized in that: The specific steps of evaluating and judging the answer accuracy of the model to be evaluated after embedding the preset knowledge graph are as follows: Obtaining embedding quality quantification data after a preset knowledge graph is embedded in a model to be evaluated, wherein the embedding quality quantification data includes embedding vector similarity, embedding global clustering coefficient, embedding space density, and embedding reconstruction error; The difference degree of embedding global clustering coefficient and embedding vector similarity, the difference degree of embedding space density and embedding reconstruction error are processed by parameter interaction respectively, and the embedding quality impact factor is obtained by coupling the corresponding parameter interaction processing results with the embedding quality impact weight. The embedding quality impact weight includes an embedding tightness impact weight and an embedding accuracy impact weight; The embedding quality impact factor is used to describe the degree of influence of the embedding quality of the preset knowledge graph embedded in the model to be evaluated on the answer accuracy judgment value; The embedding quality impact factor represents the quantitative data of the embedding quality quantification data on the influence of the embedding quality of the preset knowledge graph embedded in the model to be evaluated on the answer accuracy judgment value; Obtain the answer accuracy of the model to be evaluated, and perform an embedding quality association operation on the answer accuracy according to the obtained embedding quality impact factor to obtain an answer accuracy determination value, wherein the answer accuracy determination value is used to quantitatively evaluate the answer accuracy corresponding to the preset knowledge graph after being embedded in the model to be evaluated; The answer accuracy judgment value represents the quantitative data of the answer accuracy corresponding to the preset knowledge graph after being embedded in the model to be evaluated, which is the result of the embedding quality impact factor and the answer accuracy rate.

3. The method for enhancing the accuracy of large model answers based on knowledge graph as claimed in claim 2, characterized in that: The specific steps of obtaining the answer accuracy test result are as follows: The obtained answer accuracy determination value is combined with the minimum answer accuracy threshold obtained from the preset database for comparison and judgment; If the answer accuracy determination value is greater than the minimum answer accuracy threshold, the answer accuracy test result is recorded as the answer accuracy determination value is qualified and the answer accuracy of the model to be evaluated is to be monitored; If the answer accuracy is not greater than the minimum answer accuracy threshold, the answer accuracy test result will be recorded as the answer accuracy judgment value is unqualified and the answer accuracy of the model to be evaluated needs to be enhanced.

4. The method for enhancing the accuracy of large model answers based on knowledge graph as claimed in claim 3, characterized in that: The specific process of judging whether to adjust the initial update speed of the preset knowledge graph in combination with the answer accuracy test result is as follows: If the answer accuracy test result is that the answer accuracy judgment value is qualified and the answer accuracy of the model to be evaluated is to be monitored, the initial update speed of the preset knowledge graph is adjusted; Otherwise, the initial update speed of the preset knowledge graph is not adjusted.

5. The method for enhancing the accuracy of large model answers based on knowledge graph as claimed in claim 4, characterized in that: The specific steps of adjusting the initial update speed of the preset knowledge graph are as follows: According to the deviation degree between the answer accuracy determination value and the maximum answer accuracy threshold in the answer accuracy detection result, a preset accurate deviation mapping set in a preset database is mapped to obtain a graph speed update coefficient, wherein the preset accurate deviation mapping set represents a mapping relationship between the answer accuracy determination value and the maximum answer accuracy threshold; The initial update speed and graph speed update coefficient of the preset knowledge graph are obtained, and the initial update speed is adjusted with the graph speed update coefficient to obtain the optimized update speed.

6. The method for enhancing the accuracy of large model answers based on knowledge graph as claimed in claim 3, characterized in that: The specific process of optimizing the knowledge graph embedding of the preset knowledge graph in combination with the answer accuracy test result is as follows: S31, when the answer accuracy in the answer accuracy detection result is not greater than the minimum answer accuracy threshold, performing graph community detection and division on the preset knowledge graph to obtain the community density of the graph community; S32, performing knowledge graph compression embedding optimization on the preset knowledge graph in combination with the community density of the graph community, wherein the knowledge graph compression embedding optimization refers to obtaining the graph compression rate through the graph compression strength to perform knowledge graph compression; S33, after performing knowledge graph compression embedding optimization, performing hierarchical embedding optimization on the knowledge graph, the hierarchical embedding optimization refers to performing hierarchical embedding of the knowledge graph based on the obtained hierarchical embedding priority score, and the hierarchical embedding priority score is used to determine the hierarchical embedding priority in the process of performing hierarchical embedding of the knowledge graph.

7. The method for enhancing the accuracy of large model answers based on knowledge graph as claimed in claim 6, characterized in that: The specific steps of obtaining the graph compression rate by the graph compression strength to compress the knowledge graph are as follows: Based on the deviation degree between the community density of the graph community and the maximum community density threshold, a preset strength distribution mapping set in a preset database is mapped to obtain the graph compression strength of the graph community, wherein the preset strength distribution mapping set represents the mapping relationship between the community density and the maximum community density threshold; Mapping the deviation degree between the obtained graph compression strength and the compression strength reference value of the graph community from the preset compression strength mapping set in the preset database to obtain the graph compression rate of the graph community, wherein the preset compression strength mapping set represents the mapping relationship between the graph compression strength and the compression strength reference value; The knowledge graph of the graph community is compressed according to the obtained graph compression rate.

8. The method for enhancing the accuracy of large model answers based on knowledge graph as claimed in claim 6, characterized in that: The hierarchical embedding of the knowledge graph is performed based on the obtained hierarchical embedding priority score, and the specific steps are as follows: Quantify the information retention of the knowledge graph compression performed by the graph community based on the graph compression data after the knowledge graph is compressed by the graph community, and obtain the graph compression judgment value; Mapping the obtained deviation degree between the atlas compression judgment value and the atlas compression reference value from a preset database to obtain a compression quality score weight; The compression quality score weight is used to weight the atlas compression judgment value, and the compression rate score weight obtained from the preset database is used to weight the atlas compression rate, and then coupled to obtain the hierarchical embedding priority score; The hierarchical embedding priority scores of the graph community are sorted in descending order to obtain the hierarchical embedding priority, and the knowledge graph is hierarchically embedded according to the hierarchical embedding priority.

9. The method for enhancing the accuracy of large model answers based on knowledge graph as claimed in claim 8, characterized in that: The graph compression data after the knowledge graph compression based on the graph community is quantified to determine the information retention of the knowledge graph compression by the graph community, and the graph compression judgment value is obtained. The specific steps are as follows: Obtaining graph compression data after the graph community compresses the knowledge graph, wherein the graph compression data includes graph density, node retention, edge retention, and information loss; The coupling results of node retention and edge retention are processed by parameter interaction through the results of the difference in graph density, and the results of the information loss degree zeroing operation are weighted by combining the compression evaluation weight, and then coupled to obtain the graph compression judgment value; The compression assessment weight includes a compression retention assessment weight and a compression loss assessment weight; The graph compression judgment value represents the quantitative data of the graph compression data jointly evaluating the knowledge graph compression information retention degree of the graph community.

10. A large model answer accuracy enhancement system based on knowledge graph, characterized by: include: Answer accurate judgment module, graph update adjustment module and graph embedding optimization module; The answer accuracy judgment module is used to evaluate and judge the answer accuracy of the model to be evaluated after embedding the preset knowledge graph, and obtain the answer accuracy detection result; The graph update adjustment module is used to determine whether to adjust the initial update speed of the preset knowledge graph in combination with the answer accuracy detection result. If the initial update speed of the preset knowledge graph is adjusted, the optimized update speed of the preset knowledge graph is obtained based on the answer accuracy detection result and the initial update speed of the preset knowledge graph. The initial update speed adjustment means optimizing the initial update speed of the preset knowledge graph based on the answer accuracy detection result to improve the answer accuracy of the model to be evaluated; The graph embedding optimization module is used to perform knowledge graph embedding optimization on the preset knowledge graph in combination with the answer accuracy detection result if the initial update speed of the preset knowledge graph is not adjusted. The knowledge graph embedding optimization is used to improve the embedding quality of the preset knowledge graph embedded in the model to be evaluated to enhance the answer accuracy of the model to be evaluated.

Images