Graph recall method based on large language model, computer readable medium and equipment
By generating semantic semantic sequences based on random walking algorithms based on large language models and embedding for vector embedding, the existing diagram recovery algorithm in data structure understanding, node attribute ignoring, model update, and border relationships have achieved more accurate and diverse diverse diverse and diverse The effectiveness of the map recall.
Patent Information
- Application Number
- CN202411421944.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-12
- Publication Date
- 2025-01-28
AI Technical Summary
The existing graph recall algorithms have problems such as writing graph query languages that need to understand the graph data structure, random walk algorithms ignore node attributes and edges, algorithms based on graph embedding need to retrain the model to update data, and content-based algorithms ignore edge relationships.
Using a random walking algorithm based on a large language model, by generating random walking language sequences, converting it into sequence vectors embedded, and using vector database for retrieval, so as to extract the nodes and edges in the data of the data Essence
A more accurate node similarity calculation and link prediction have been implemented, the diversity and depth of recall are improved, the path of related nodes can be generated according to the user's requirements, enhanced the global or local excellence ability of the algorithm, and improved the transparency of decision -making.
Smart Images
Figure CN119377451A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of data processing technology, and in particular relates to a graph recall method based on a large language model, a computer-readable medium and a device. Background Art
[0002] A graph is a data structure that stores nodes and edges. Graph recall refers to the process of retrieving specific information from a graph. Existing graph recall algorithms include recall algorithms written by graph query languages, random walk algorithms, recall algorithms based on graph embedding, and recall algorithms based on content. Among them, recall algorithms written by graph query languages write specific query statements, such as Cypher, Gremlin, nGQL, GSQL, PGQL, etc., and use specific graph engines in specific modes to execute graph query languages to retrieve information from graph structures. The random walk algorithm generates random paths in the graph, which can be used to generate embedded representations of nodes and then used for recall. For example, the traditional random walk algorithm is based on random sampling, calculates the transition probability of the nodes that have been walked and the unknown nodes, and finally generates a sequence of nodes in the graph; the Node2vec algorithm learns the embedded representation of nodes in the graph through random walks. The recall algorithm based on graph embedding converts the text information of graph nodes into vector embeddings and learns the embedded representations of the nodes in the graph for recall. These embedded representations can capture the local and global neighborhood information of the nodes, and recall the nodes in the graph based on vector similarity during query. For example, DeepWalk and Node2Vec are two popular graph embedding algorithms based on random walks. The content-based recall algorithm calculates the similarity with the description information in the query statement by analyzing the content features or attributes of the nodes, and then finds the nodes based on these similarity analyses. The above graph recall algorithms can be used alone or in combination.
[0003] However, the existing graph recall algorithms have the following problems:
[0004] The method of writing a graph query language requires understanding the graph data structure in advance, such as the type and attribute information of known nodes, the type and attribute information of edges, etc., and building the corresponding graph query language;
[0005] The random walk algorithm ignores node attributes and edges and simply records the connection relationship between nodes. The generated sequence is likely to have no actual semantic meaning as a whole.
[0006] For the recall algorithm based on graph embedding, when the data in the graph is updated, for example, when new nodes or edges are added, the model needs to be retrained to obtain a new graph embedding.
[0007] Content-based recall methods can only recall nodes with high text similarity, ignoring the relationship between edges. Summary of the invention
[0008] In view of this, on the one hand, some embodiments disclose a graph recall method based on a large language model, comprising the steps of:
[0009] S1. Use the random walk algorithm based on the large language model to construct a random walk semantic sequence;
[0010] S2. embed the constructed random walk semantic sequence into a sequence vector and store it in a vector database;
[0011] S3. Convert the natural language input by the user into a retrieval vector embedding;
[0012] S4. Input the search vector embedded in the vector database to retrieve the pre-stored semantic sequence;
[0013] S5. Extract the nodes and edges of the retrieved semantic sequence from the graph data, and the retrieved nodes and edges constitute a subgraph as the query result.
[0014] Furthermore, some embodiments disclose a graph recall method based on a large language model, characterized in that step S1 comprises:
[0015] Based on several topics and graph data structure information of the graph model, the prompt method of the large language model is used to walk through the graph and generate prompt description sentences;
[0016] The suggestive description sentences are input into the large language model to generate candidate walking paths;
[0017] Using the generated candidate walk paths, data sampling is performed from the graph data to obtain a random walk semantic sequence.
[0018] In some embodiments, the graph recall method based on a large language model disclosed in step S2 includes:
[0019] Train a graph embedding model;
[0020] The graph node data of the random walk semantic sequence is passed into the graph vector embedding model in sequence order, converted into vector embedding and stored in the vector database, and the vector embedding is used as the embedding vector of the graph data.
[0021] Furthermore, in some embodiments of the graph recall method based on a large language model, in step S3, the natural language input by the user is converted into a retrieval vector embedding using the same graph vector embedding model as in step S2.
[0022] In some embodiments, the graph recall method based on a large language model disclosed in step S2 includes:
[0023] Train a text vector embedding model;
[0024] Organize the graph node data of the random walk semantic sequence into appropriate suggestive description sentences in sequence order, and pass the suggestive description sentences into the large language model to generate summary description text for the random walk semantic sequence;
[0025] The summary description text is passed into the text vector embedding model to be converted into a vector embedding and stored in the vector database, and the vector embedding is used as the embedding vector of the graph data.
[0026] Furthermore, some embodiments disclose a graph recall method based on a large language model. In step S3, the natural language input by the user is converted into a retrieval vector embedding using the same text vector embedding model as in step S2.
[0027] In some embodiments of the graph recall method based on a large language model, in step S4, a similarity calculation method is used for retrieval.
[0028] In some embodiments of the graph recall method based on a large language model, the generated candidate wandering paths are used as wandering paths after manual review.
[0029] On the other hand, some embodiments disclose a computer-readable medium storing a computer program, which, when executed, implements a graph recall method based on a large language model.
[0030] On the other hand, some embodiments disclose a device including a memory and an information processor, wherein the memory stores a computer program, and when the information processor reads the computer program, it executes a graph recall method based on a large language model.
[0031] The graph recall method based on a large language model disclosed in the embodiment of the present invention, combined with the random walk algorithm of the large language model, can better understand the semantic information of nodes in the graph, and realize more accurate node similarity calculation and link prediction in graph recall; realize multiple information retrieval, expand the search scope through methods such as entity linking and synonym expansion, and improve the diversity and depth of recall; can generate relevant node paths according to the specific requirements of users, and realize personalized graph search; enhance the global or local optimization ability of the algorithm, improve the convergence speed of the algorithm and the quality of the solution; when generating text, it usually has good interpretability, and improves the decision transparency in the graph recall process.
[0032] The graph recall method based on a large language model disclosed in the embodiment of the present invention can directly query corresponding information from a graph using natural language. The semantic sequence generated by the walk algorithm based on a large language model is more interpretable than the random walk algorithm. The walk algorithm based on a large model can generate walk sequences of different topics. The query efficiency based on vector similarity is higher. Users can query multiple data to form a subgraph without writing a graph query statement. BRIEF DESCRIPTION OF THE DRAWINGS
[0033] Figure 1 This is a flow chart of the graph recall method based on a large language model disclosed in Example 1;
[0034] Figure 2 This is a flow chart of the graph recall method based on a large language model disclosed in Example 2;
[0035] Figure 3 This is a flow chart of the graph recall method based on a large language model disclosed in Example 3. DETAILED DESCRIPTION
[0036] The word "embodiment" used herein as an "exemplary" does not necessarily mean that any embodiment described is superior or better than other embodiments. Performance index tests in the embodiments of the present invention are performed using conventional test methods in the art unless otherwise specified. It should be understood that the terms described in the embodiments of the present invention are only used to describe specific implementation methods and are not intended to limit the contents disclosed in the embodiments of the present invention.
[0037] Unless otherwise specified, the technical and scientific terms used herein have the same meanings as commonly understood by ordinary technicians in the technical field to which the embodiments of the present invention belong; other experimental methods and technical means not specifically specified in the embodiments of the present invention refer to experimental methods and technical means commonly used by ordinary technicians in the field.
[0038] The terms "substantially" and "approximately" used herein are used to describe small fluctuations. For example, they can refer to less than or equal to ±5%, such as less than or equal to ±2%, such as less than or equal to ±1%, such as less than or equal to ±0.5%, such as less than or equal to ±0.2%, such as less than or equal to ±0.1%, such as less than or equal to ±0.05%. The numerical data represented or presented in the range format herein are used only for convenience and brevity, and should therefore be flexibly interpreted as including not only the values clearly listed as the limits of the range, but also all independent values or sub-ranges contained in the range. For example, the numerical range of "1-5%" should be interpreted as including not only the clearly listed values of 1% to 5%, but also the independent values and sub-ranges within the range shown. Therefore, independent values such as 2%, 3.5% and 4% and sub-ranges such as 1%-3%, 2%-4% and 3%-5% are included in this numerical range. This principle also applies to the range of only one numerical value. In addition, such an interpretation applies regardless of the width of the range or the characteristics described.
[0039] In this document, including in the claims, transitional words such as "comprises," "includes," "with," "having," "containing," "involving," "accommodating," etc. are understood to be open-ended, i.e., meaning "including but not limited to." Only the transitional words "consisting of" and "composed of" are closed transitional words.
[0040] In order to better illustrate the present invention, numerous specific details are provided in the specific examples below. It should be understood by those skilled in the art that the present invention can be implemented without certain specific details. In the embodiments, some methods, means, instruments, equipment, etc. well known to those skilled in the art are not described in detail in order to highlight the gist of the present invention.
[0041] Under the premise of no conflict, the technical features disclosed in the embodiments of the present invention may be arbitrarily combined, and the resulting technical solutions belong to the contents disclosed in the embodiments of the present invention.
[0042] In some implementations, the large language model-based graph recall method includes the steps of:
[0043] S1. Use a random walk algorithm based on a large language model to construct a random walk semantic sequence. Usually, based on several topics of the graph model pre-entered by the user, combined with the graph data structure, such as the type information of the edges and the type information of the nodes in the graph, the large language model can be used to sample from the graph data to generate possible path structure information, and the random walk path can be extracted from the graph data according to the path structure information to form a random walk semantic sequence.
[0044] S2. embed the constructed random walk semantic sequence into a sequence vector and store it in a vector database;
[0045] S3. converting the natural language input by the user into a search vector embedding; the search vector embedding generally refers to a vector embedding for search obtained by converting the natural language input by the user to meet the search needs of the user;
[0046] S4. Input the search vector embedded in the vector database to retrieve the pre-stored semantic sequence;
[0047] S5. Extract the nodes and edges of the retrieved semantic sequence from the graph data, and the retrieved nodes and edges constitute a subgraph as the query result.
[0048] Furthermore, some embodiments disclose a graph recall method based on a large language model, characterized in that step S1 comprises:
[0049] Based on several topics of the graph model and the graph data structure information, the prompt method of the large language model is used to walk through the graph and generate prompt description sentences; usually the graph structure information includes the type information of the node, such as person-type node, mobile phone number-type node, WeChat account-type node, and the type information of the edge, for example, friends, colleagues, relatives, etc. represent the edges between person-type nodes and person-type nodes, hold, held, etc. represent the edges between person-type nodes and mobile phone model nodes, hold, held, etc. represent the edges between person-type nodes and WeChat account-type nodes, register, etc. represent the edges between mobile phone number-type nodes and WeChat account-type nodes, call, etc. represent the edges between mobile phone number-type nodes and mobile phone number-type nodes, follow, etc. represent the edges between WeChat account-type nodes and WeChat account-type nodes, etc.;
[0050] Usually, several topics of the graph model are graph model topics input by the user, including strong relationships such as communication and chatting, and weak relationships such as possible acquaintance;
[0051] Usually, we can select a topic, start from any node, extract its adjacent nodes, and combine the topic, two points and edge attributes into a prompt. The large language model determines the direction of the next hop and generates a prompt description language for the current random walk semantic sequence.
[0052] The suggestive description sentence is input into the large language model to generate candidate walking paths; for example, the candidate walking paths are:
[0053] Strong relationship - communication chat: person->holder->WeChat->follow->WeChat->holder->person;
[0054] Weak relationship - possible acquaintance: person->holder->phone number->dial->phone number->holder->person; person->holder->phone number->dial->phone number->register->WeChat->holder->person;
[0055] Using the generated candidate walk paths, we sample data from the graph data to get a random walk semantic sequence. For example, if the walk path is: person->hold->WeChat->follow->WeChat->hold->person, we generate a graph query cypher, sample data from the graph data, and get a random walk semantic sequence:
[0056] Xiaohong->Hold->WeChat-Honghong->Follow->WeChat-Huahua->Hold->Xiaohua;
[0057] Huazai->holds->WeChat-In the Mood for Love->follow->WeChat-Get Rich->held->Zhou Zhou.
[0058] In some embodiments, the graph recall method based on a large language model disclosed in step S2 includes:
[0059] Training graph vector embedding model; usually, the image embedding model can use the commonly used graph data structure model, such as DeepWalk algorithm, Node2Vec algorithm, etc.; among them, DeepWalk algorithm uses the random walk of the graph to generate node sequence, and then uses Word2Vec to learn the embedded representation of the node. It does not rely on the structural information of the graph, but explores the graph through random walk. Its basic training method is: generate node sequence through random walk, and then use Word2Vec's Skip-Gram model for training; Node2Vec algorithm is an algorithm based on random walk. When learning the embedded representation of nodes by random walk in the graph, it combines the strategies of depth-first search (DFS) and breadth-first search (BFS) to explore the local and global structures in the graph. Its basic training method is: generate node sequence by walking in the graph through a certain search strategy, and then use Word2Vec's Skip-Gram model to learn the embedding of nodes;
[0060] The graph node data of the random walk semantic sequence is passed into the graph vector embedding model in sequence order, converted into vector embeddings and stored in the vector database, and the vector embeddings are used as the embedding vectors of the graph data;
[0061] Step S3 includes:
[0062] The natural language input by the user is converted into retrieval vector embedding using the same graph vector embedding model as in step S2.
[0063] In some embodiments, the graph recall method based on a large language model disclosed in step S2 includes:
[0064] Train a text vector embedding model; usually, a text vector embedding model can use a commonly used text data model, such as the Word2Vec algorithm, the GloVe algorithm, the FastText algorithm, etc.; the Word2Vec algorithm is a model that predicts the context of a word, which can learn the distributed representation of a word; the Word2Vec algorithm has two architectures: continuous bag of words (CBOW) and Skip-Gram; its basic training method is to use a neural network to train by maximizing the probability of predicting the occurrence of a word in the context; during the training process, the model learns the embedding representation of the word;
[0065] The GloVe (Global Vectors for Word Representation) algorithm is a matrix factorization-based method that converts word co-occurrence statistics into word vectors; its basic training method is to optimize word vectors by minimizing the difference between the co-occurrence probability of word pairs and the model's predicted probability; the FastText algorithm is a model developed by Facebook that learns vector representations of words by considering the subword (character n-gram) information of the word; its basic training method is to use the Skip-Gram model similar to Word2Vec, but consider both words and their subwords during the training process;
[0066] Organize the graph node data of the random walk semantic sequence into appropriate suggestive description sentences in sequence order, and pass the suggestive description sentences into the large language model to generate summary description text for the random walk semantic sequence;
[0067] The summary description text is passed into the text vector embedding model to be converted into a vector embedding and stored in the vector database, and the vector embedding is used as the embedding vector of the graph data;
[0068] Step S3 includes: using the same text vector embedding model as in step S2, converting the natural language input by the user into a retrieval vector embedding.
[0069] In some embodiments of the graph recall method based on a large language model, in step S4, a similarity calculation method is used for retrieval, which generally includes cosine similarity, Euclidean similarity, Pearson correlation coefficient, Manhattan distance and other methods.
[0070] In some embodiments, the graph recall method based on the large language model disclosed in the embodiments generates candidate wandering paths, which are used as wandering paths after manual review. Generally, manual review can be used to check whether the node paths generated with the assistance of the large model have actual semantic meaning, whether they conform to the predetermined theme, etc., and retain the expected results.
[0071] Some embodiments disclose a computer-readable medium storing a computer program. When the computer program is executed, a graph recall method based on a large language model is implemented.
[0072] Some embodiments disclose a device including a memory and an information processor, wherein the memory stores a computer program, and when the information processor reads the computer program, a graph recall method based on a large language model is executed.
[0073] The graph recall method based on a large language model disclosed in the embodiment of the present invention, combined with the random walk algorithm of the large language model, can better understand the semantic information of nodes in the graph, and realize more accurate node similarity calculation and link prediction in graph recall; realize multiple information retrieval, expand the search scope through methods such as entity linking and synonym expansion, and improve the diversity and depth of recall; can generate relevant node paths according to the specific requirements of users, and realize personalized graph search; enhance the global or local optimization ability of the algorithm, improve the convergence speed of the algorithm and the quality of the solution; when generating text, it usually has good interpretability, and improves the decision transparency in the graph recall process.
[0074] The graph recall method based on a large language model disclosed in the embodiment of the present invention can directly query corresponding information from a graph using natural language. The semantic sequence generated by the walk algorithm based on a large language model is more interpretable than the random walk algorithm. The walk algorithm based on a large model can generate walk sequences of different topics. The query efficiency based on vector similarity is higher. Users can query multiple data to form a subgraph without writing a graph query statement.
[0075] The technical details are further illustrated below in conjunction with embodiments.
[0076] Example 1
[0077] Figure 1 This is a flow chart of the graph recall method based on a large language model disclosed in Example 1.
[0078] In Example 1, the graph recall method based on a large language model includes the following steps:
[0079] S1. Use the random walk algorithm based on the large language model to construct a random walk semantic sequence;
[0080] S2, embed the constructed random walk semantic sequence into a sequence vector and store it in a vector database;
[0081] S3, convert the natural language input by the user into retrieval vector embedding;
[0082] S4, input the search vector to embed in the vector database to search for the pre-stored semantic sequence;
[0083] S5. Extract the nodes and edges of the retrieved semantic sequence from the graph data, and the retrieved nodes and edges constitute a subgraph as the query result.
[0084] Example 2
[0085] Figure 2 This is a flow chart of the graph recall method based on a large language model disclosed in Example 2.
[0086] In Example 2, the graph recall method based on a large language model includes the following steps:
[0087] S11. Based on several topics and graph data structure information of the graph model, use the prompt method of the large language model to walk in the graph and generate prompt description sentences;
[0088] S12, inputting the suggestive description sentence into the large language model to generate candidate walking paths;
[0089] S13, using the generated candidate walk paths, sampling data from the graph data to obtain a random walk semantic sequence;
[0090] S21, training graph vector embedding model;
[0091] S22, passing the graph node data of the random walk semantic sequence into the graph vector embedding model in sequence order, converting them into vector embeddings and storing them in the vector database, and using the vector embeddings as the embedding vectors of the graph data;
[0092] S3, using the same graph vector embedding model as in step S22, converting the natural language input by the user into a retrieval vector embedding;
[0093] S4, input the search vector embedding, and use the similarity calculation method to search the pre-stored semantic sequence in the vector database;
[0094] S5. Extract the nodes and edges of the retrieved semantic sequence from the graph data, and the retrieved nodes and edges constitute a subgraph as the query result.
[0095] Example 3
[0096] Figure 3 This is a flow chart of the graph recall method based on a large language model disclosed in Example 3.
[0097] In Example 3, the graph recall method based on a large language model includes the following steps:
[0098] S11. Based on several topics and graph data structure information of the graph model, use the prompt method of the large language model to walk in the graph and generate prompt description sentences;
[0099] S12, inputting the suggestive description sentence into the large language model to generate candidate walking paths;
[0100] S13, using the generated candidate walk paths, sampling data from the graph data to obtain a random walk semantic sequence;
[0101] S21, training text vector embedding model;
[0102] S22, organizing the graph node data of the random walk semantic sequence into appropriate suggestive description sentences in sequence order, and passing the suggestive description sentences into the large language model to generate a summary description text for the random walk semantic sequence;
[0103] S23, converting the summary description text into a text vector embedding model and storing it in a vector database, using the vector embedding as an embedding vector of the graph data;
[0104] S3, using the same text vector embedding model as in step S23, converting the natural language input by the user into a retrieval vector embedding;
[0105] S4, input the search vector embedding, and use the similarity calculation method to search the pre-stored semantic sequence in the vector database;
[0106] S5. Extract the nodes and edges of the retrieved semantic sequence from the graph data, and the retrieved nodes and edges constitute a subgraph as the query result.
[0107] The graph recall method based on a large language model disclosed in the embodiment of the present invention can directly query corresponding information from a graph using natural language. The semantic sequence generated by the walk algorithm based on a large language model is more interpretable than the random walk algorithm. The walk algorithm based on a large model can generate walk sequences of different topics. The query efficiency based on vector similarity is higher. Users can query multiple data to form a subgraph without writing a graph query statement.
[0108] The technical solutions disclosed in the embodiments of the present invention and the technical details disclosed in the embodiments are merely illustrative of the inventive concept of the present invention and do not constitute a limitation on the technical solutions of the embodiments of the present invention. Any conventional changes, replacements or combinations of the technical details disclosed in the embodiments of the present invention have the same inventive concept as the present invention and are within the protection scope of the claims of the present invention.
Claims
1. A graph recall method based on a large language model, characterized in that: Includes steps: S1. Use the random walk algorithm based on the large language model to construct a random walk semantic sequence; S2. embed the constructed random walk semantic sequence into a sequence vector and store it in a vector database; S3. Convert the natural language input by the user into a retrieval vector embedding; S4. Input the search vector embedded in the vector database to retrieve the pre-stored semantic sequence; S5. Extract the nodes and edges of the retrieved semantic sequence from the graph data, and the retrieved nodes and edges constitute a subgraph as the query result.
2. The graph recall method based on a large language model according to claim 1, characterized in that: Step S1 includes: Based on several topics and graph data structure information of the graph model, the prompt method of the large language model is used to walk through the graph and generate prompt description sentences; The suggestive description sentence is input into the large language model to generate candidate walking paths; Using the generated candidate walk paths, data sampling is performed from the graph data to obtain a random walk semantic sequence.
3. The graph recall method based on a large language model according to claim 1, characterized in that: Step S2 includes: Train a graph embedding model; The graph node data of the random walk semantic sequence is passed into the graph vector embedding model in sequence order, converted into vector embedding and stored in the vector database, and the vector embedding is used as the embedding vector of the graph data.
4. The graph recall method based on a large language model according to claim 3, characterized in that: In step S3, the natural language input by the user is converted into a retrieval vector embedding using the same graph vector embedding model as in step S2.
5. The graph recall method based on a large language model according to claim 1, characterized in that: Step S2 includes: Train a text vector embedding model; Organizing the graph node data of the random walk semantic sequence into appropriate suggestive description sentences in sequence order, and passing the suggestive description sentences into a large language model to generate a summary description text for the random walk semantic sequence; The summary description text is passed into the text vector embedding model to be converted into a vector embedding and stored in the vector database, and the vector embedding is used as the embedding vector of the graph data.
6. The graph recall method based on a large language model according to claim 5, characterized in that: In step S3, the natural language input by the user is converted into a retrieval vector embedding using the same text vector embedding model as in step S2.
7. The graph recall method based on a large language model according to claim 1, characterized in that: In step S4, a similarity calculation method is used for retrieval.
8. The graph recall method based on a large language model according to claim 2, characterized in that: The generated candidate wandering paths are used as wandering paths after manual review.
9. A computer readable medium, characterized in that The computer readable medium stores a computer program, and when the computer program is executed, the method according to any one of claims 1 to 8 is implemented.
10. A device comprising a memory and an information processor, characterized in that The memory stores a computer program, and when the information processor reads the computer program, the method according to any one of claims 1 to 8 is executed.
Citation Information
Cited By
Structured knowledge base mapping retrieval method based on visual semantic high-dimensional vectorization
CN122527179A