Intelligent go playing method and system based on multi-modal retrieval enhancement

Abstract

The application belongs to the field of multi-modal large model application, and provides an intelligent go game analysis method and system based on multi-modal retrieval enhancement, which acquires historical games and corresponding comments, extracts feature embeddings of the historical games and the corresponding comments, trains a game feature extraction network based on the similarity of the historical game feature embeddings and the corresponding comment feature embeddings, obtains the trained game feature extraction network, extracts feature embeddings of the current game and historical game features based on the trained game feature extraction network, calculates the similarity of the current game feature embedding and each historical game feature embedding, obtains similar games according to the similarity calculation result, finds corresponding comments according to the index of the similar games, combines the current game AI prediction result and the index of the similar games to find corresponding comments to form prompt words, and obtains game analysis results of the current game according to the prompt words and a large language model. The credibility of the pre-trained large language model in generating comments is improved.

Description

Technical Field

[0001] This invention belongs to the field of multimodal large model applications, and in particular relates to an intelligent Go game analysis method and system based on multimodal retrieval enhancement. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Due to the extreme complexity of Go, machines cannot play by listing all possible moves; traditional Go programs can only reach a high-level beginner level. In recent years, with the development of deep learning, reinforcement learning-based Go engines, particularly AlphaZero and KataGo, have developed playing strength surpassing that of human players. However, these Go engines can only provide win rates and predictions; they cannot directly enable users to deeply understand the decisions made in the game. Users still need expert commentary to understand the current state and variations of the game.

[0004] To address this issue, large language models based on retrieval-enhanced generative techniques have made intelligent Go game commentary possible. Large language models (LLMs) are trained on massive amounts of data, enabling them to understand fundamental knowledge and terminology in the Go domain. They utilize hundreds of billions of parameters to generate raw outputs for tasks such as answering questions and completing sentences. Currently, large language models, such as GPT-4o, can play against humans and generate analytical commentary on the current move and the overall game situation based on clues such as win rate and territory allocation.

[0005] While pre-trained large language models can understand Go terminology and game rules, they face a significant challenge: AI hallucinations. This means that the model, lacking any supporting facts, can generate completely fabricated information. This often leads to the risk of generating incorrect or false information, which can have serious consequences for specialized applications, particularly Go game analysis.

[0006] One effective approach to addressing this challenge is to employ multimodal retrieval-augmented generation (RAG) techniques. RAG extends the already powerful natural language generation capabilities of large language models by granting them access to internal knowledge bases within specific domains or organizations. This allows them to retrieve relevant information from authoritative, pre-defined knowledge sources, enhancing the relevance and accuracy of generated text within vertical domains. Go, in essence, can be viewed as a pattern recognition game: from eyes and liberties to joseki and territory control, professional players rely on a rich set of concepts to represent the structure on the board. Based on existing Go game datasets containing numerous human annotations, retrieving both the board and natural language annotations as two modalities can provide large language models with more accurate information sources, enhancing the credibility of generated text. However, Go game datasets are typically sparse; not every move can necessarily be found in the existing data, and the generated comments often lack credibility. Summary of the Invention

[0007] To address at least one of the technical problems mentioned above, this invention provides an intelligent Go commentary system for beginners based on multimodal retrieval enhancement. This system, for the current game, retrieves information from a database within the professional Go field and constructs prompts based on similar games and comments found, thereby enhancing the credibility of comments generated by a pre-trained large language model.

[0008] To achieve the above objectives, the present invention adopts the following technical solution:

[0009] The first aspect of this invention provides an intelligent Go game analysis method based on multimodal retrieval enhancement, comprising the following steps:

[0010] Obtain historical game positions and corresponding comments, and extract feature embeddings from historical game positions and corresponding comments;

[0011] The game feature extraction network is trained based on the similarity between historical game feature embeddings and corresponding comment feature embeddings, resulting in the trained game feature extraction network.

[0012] The trained chess game feature extraction network extracts feature embeddings of the current chess game and features of historical chess games.

[0013] Calculate the similarity between the current game feature embedding and the feature embedding of each historical game. Based on the similarity calculation results, obtain similar games and find the corresponding comments based on the index of the similar games.

[0014] The AI ​​prediction of the current game and the index of similar games are used to find corresponding comments to form prompt words;

[0015] The game analysis results for the current chess position are obtained based on the prompt words and the large language model.

[0016] Furthermore, the game feature extraction network is trained based on the similarity between historical game feature embeddings and corresponding comment feature embeddings to obtain the trained game feature extraction network, including:

[0017] The first game feature embedding is extracted based on the pre-trained KataGo residual neural network;

[0018] The first game feature embedding is output to a trainable neural network Φ to obtain the second game feature embedding;

[0019] The loss function of the global model is constructed based on the similarity between the second game feature embedding and the corresponding comment feature embedding. When the global model loss function converges to a predetermined accuracy or reaches a predetermined number of training rounds, the trained neural network Φ is obtained.

[0020] Furthermore, the loss function of the global model is:

[0021]

[0022]

[0023] Where τ is the temperature parameter of the neural network, and the key-value pair of historical game-comment is (B i C i ), Embedding features for the second game. Embedding of comment features; C j This refers to any comment in the comment collection.

[0024] Furthermore, after obtaining the historical and current game positions, the historical and current game positions are converted into the SGF standard format.

[0025] Furthermore, the large language model adopts GPT-4o or Tongyiqianwen 2.5.

[0026] Furthermore, the current game AI prediction result is obtained through KataGo analysis, including win rate analysis and next move prediction.

[0027] A second aspect of the present invention provides an intelligent Go game analysis system based on multimodal retrieval enhancement, comprising:

[0028] The chess game feature extraction network training module is configured to acquire historical chess games and corresponding comments, and extract feature embeddings of historical chess games and corresponding comments; the chess game feature extraction network is trained based on the similarity between the historical chess game feature embeddings and the corresponding comment feature embeddings to obtain the trained chess game feature extraction network;

[0029] The feature extraction module is configured to extract feature embeddings of the current game and historical game based on the trained game feature extraction network;

[0030] The prompt word generation module is configured to calculate the similarity between the feature embedding of the current game and the feature embedding of each historical game, obtain similar games based on the similarity calculation results, and find the corresponding comments based on the index of similar games; the prompt word is composed of the AI ​​prediction result of the current game and the corresponding comments found by the index of similar games.

[0031] The game commentary generation module is configured to generate game analysis results for the current game based on prompt words and a large language model.

[0032] A third aspect of the present invention provides a computer-readable storage medium.

[0033] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the intelligent Go game analysis method based on multimodal retrieval enhancement as described above.

[0034] A fourth aspect of the present invention provides a computer device.

[0035] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the intelligent Go game analysis method based on multimodal retrieval enhancement as described above.

[0036] A fifth aspect of the present invention provides a program product.

[0037] A program product, which is a computer program product, includes a computer program that, when executed by a processor, implements the steps in the intelligent Go game analysis method based on multimodal retrieval enhancement as described above.

[0038] Compared with the prior art, the beneficial effects of the present invention are:

[0039] This invention, based on the data transmission of Go game data between the front-end and back-end of a server, retrieves similar Go games and existing comments from a vector database. Combining this with Go AI predictions, a large language model generates a commentary on the current game. Simultaneously, based on a designed retrieval mechanism, a game embedding model is trained to guide the retrieval process of similar games. Ultimately, this achieves an intelligent Go commentary system based on multimodal retrieval enhancement, improving retrieval accuracy and enhancing the credibility of the generated comments.

[0040] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0041] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0042] Figure 1 This is a framework diagram of an intelligent Go game analysis system based on multimodal retrieval enhancement provided by an embodiment of the present invention;

[0043] Figure 2 This is a schematic diagram of the retrieval training and inference structure provided in an embodiment of the present invention. Detailed Implementation

[0044] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0045] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0046] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0047] As mentioned in the background section of this application, Go game datasets are typically sparse, and each move may not necessarily yield identical game results in existing data. Furthermore, the generated commentaries often suffer from low credibility. This invention addresses this issue by retrieving similar games and existing commentaries from a vector database, based on front-end and back-end Go game data transmission on a server. Combining this with Go AI predictions, a large language model generates a commentary on the current game. Simultaneously, based on a designed retrieval mechanism, a game embedding model is trained to guide the retrieval process for similar games. Ultimately, this results in an intelligent Go commentary system based on multimodal retrieval enhancement, improving retrieval accuracy and enhancing the credibility of the generated commentaries.

[0048] Example 1

[0049] like Figure 1 and Figure 2 As shown, this embodiment provides an intelligent Go game analysis method based on multimodal retrieval enhancement, including the following steps:

[0050] Step 1: Obtain historical game positions and corresponding comments, and extract feature embeddings from historical game positions and corresponding comments;

[0051] In this embodiment, historical chess games and corresponding comments are represented as: existing chess games B = {B1, B2, ..., B...} N} and the corresponding comments C = {C1, C2, ..., C} N};

[0052] For a given key-value pair of game state and comment (B i C i ), assuming It is a pre-trained KataGo residual neural network The C-dimensional vector space of the encoded H×W feature plane is processed by a trainable neural network Φ to obtain the game feature embedding.

[0053] Assuming the corresponding comment is C i Embedding of features obtained from a pre-trained text encoder and and Same dimensions.

[0054] Step 2: Train the game feature extraction network based on the similarity between the historical game feature embedding and the corresponding comment feature embedding to obtain the trained game feature extraction network;

[0055] The similarity between the chess game and the commentary can be represented as:

[0056]

[0057] Among them, data pairs (B) i C i (This is a chess game - commentary)

[0058] The loss function during training can be defined as:

[0059]

[0060] Where τ is the temperature parameter of the neural network, C j It is any comment that belongs to comment set C;

[0061] During training, sim(B) i C i ) is B i and all comments C in the batch j Calculate the similarity score separately.

[0062] Repeat the above training steps until the global model loss function converges to the predetermined accuracy or reaches the predetermined number of training rounds, and obtain the trained neural network Φ.

[0063] Step 3: Obtain the current game state and preprocess it;

[0064] In this embodiment, after obtaining the current game state, the user's current game state is converted into the standard format SGF (SmartGameFormat).

[0065] Step 4: Extract feature embeddings of the current game and historical game features based on the trained game feature extraction network;

[0066] Let there be a chess position B = {B1, B2, ..., B} N The feature vector embeddings are obtained sequentially through a pre-trained KataGo deep neural network and a fully trained neural network Φ. The game is stored on the server using the vector database FAISS. MySQL is used to store the game index and corresponding comments C = {C1, C2, ..., C}. N}

[0067] In the reasoning process based on a multimodal retrieval system, for the new chess position B input from the front end... q Vector feature embeddings are obtained through a pre-trained KataGo residual neural network and a trained multilayer neural network Φ.

[0068] Step 5: Calculate the similarity between the current game feature embedding and the feature embedding of each historical game. Based on the similarity calculation results, obtain similar games and find the corresponding comments based on the index of the similar games.

[0069] In this embodiment, the current game feature embedding and vector storage E are calculated. B The similarity formula for each feature vector embedding is:

[0070]

[0071] Where i∈{1,2,...,N}, the top k maximum similarity scores are selected from TopK(sim(B)). q B i )), resulting in a similar chess position R B ={B r}, r∈{1,2,...,k}. Find the corresponding comment R based on the index of similar chess positions. c ={C r}, r∈{1,2,...,k} is an important component of the prompt words.

[0072] Step 6: Combine the prediction results of the current game state through the KataGo network with the index of similar games to find the corresponding comments and form prompt words;

[0073] Regarding the new chess position B passed in from the front end q Assume that the result predicted by the KataGo analysis is P. q The prompt word can be represented as Prompt(P) q R C ).

[0074] Step 7: Obtain the game analysis results of the current game based on the prompt words and the large language model.

[0075] In this embodiment, the pre-trained large language model G is used as the generator. Therefore, the game analysis output for the current chess position is:

[0076] Analysis = G(Prompt(P) q R C )),

[0077] The pre-trained large language model G can be any large language model, such as GPT-4o, Tongyi 1000 Questions 2.5, etc.

[0078] The above scheme implements an intelligent Go game analysis function using a multimodal retrieval-enhanced generative model. It retrieves similar games and existing comments from a vector database, combines Go AI predictions, and generates a game analysis of the current game using a large language model. Furthermore, this invention designs a retrieval tool that trains a game embedding model to guide the retrieval process of similar games. Ultimately, this achieves an intelligent Go game analysis system based on multimodal retrieval-enhanced generative model, improving retrieval accuracy and enhancing the credibility of generated comments.

[0079] Example 2

[0080] This embodiment provides an intelligent Go game explanation system based on multimodal retrieval enhancement for beginners, including a server and a user terminal;

[0081] The user terminal is configured to process the current game situation of the user and send it to the server.

[0082] In this embodiment, the user terminal is configured to send the current game position of the user to the server via API in the standard format SGF (SmartGameFormat);

[0083] The server includes a KataGo analysis module, a game feature extraction network training module, a feature extraction module, a prompt word generation module, and a game commentary generation module;

[0084] The KataGo analysis module is configured to receive the current game situation from the user, analyze the current game situation based on KataGo, and obtain the KataGo game situation analysis result.

[0085] After receiving the game board in SGF format via API, the KataGo analysis module processes the current game board and sends it to the KataGo server deployed in the cloud for win rate analysis and next move prediction, and obtains the return results from KataGo.

[0086] The chess game feature extraction network training module is configured to acquire historical chess games and corresponding comments, extract feature embeddings of historical chess games and corresponding comments, and train the chess game feature extraction network based on the similarity between the historical chess game feature embeddings and the corresponding comment feature embeddings to obtain the trained chess game feature extraction network.

[0087] The feature extraction module is configured to extract the feature embeddings of the current game and the feature embeddings of the historical game based on the trained game feature extraction network.

[0088] The prompt word generation module is configured to calculate the similarity between the feature embedding of the current game and the feature embedding of each historical game, obtain similar games based on the similarity calculation results, find the corresponding comments based on the index of the similar games, and combine the AI ​​prediction result of the current game and the corresponding comments found by the index of the similar games to form prompt words;

[0089] This invention performs a similarity search between the current game state and the trained multimodal retrieval system and pre-stored game states in a vector database. By querying the game state with the highest similarity through an index, the game state comment corresponding to the game state with the highest similarity is combined with the KataGo game state analysis results to form prompt words, which helps to improve the credibility of comments generated by the pre-trained large language model.

[0090] The game commentary generation module is configured to obtain the game analysis results of the current game based on prompt words and a large language model, and return them to the user's Go game webpage interface for display.

[0091] The implementation process of each module in this embodiment is the same as that in Embodiment 1.

[0092] Example 3

[0093] This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the intelligent Go game analysis method based on multimodal retrieval enhancement described above.

[0094] Example 4

[0095] This embodiment provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps in the intelligent Go game analysis method based on multimodal retrieval enhancement described above.

[0096] Example 5

[0097] This embodiment provides a program product, which is a computer program product including a computer program. When the computer program is executed by a processor, it implements the steps in the intelligent Go game analysis method based on multimodal retrieval enhancement described above.

[0098] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of hardware embodiments, software embodiments, or embodiments combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage and optical storage) containing computer-usable program code.

[0099] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0100] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0101] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0102] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0103] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for intelligent Go game analysis based on multimodal retrieval enhancement, characterized in that, Includes the following steps: Obtain historical game positions and corresponding comments, and extract feature embeddings from historical game positions and corresponding comments; The game feature extraction network is trained based on the similarity between historical game feature embeddings and corresponding comment feature embeddings to obtain the trained game feature extraction network; wherein, the training of the game feature extraction network based on the similarity between historical game feature embeddings and corresponding comment feature embeddings to obtain the trained game feature extraction network includes: The first game feature embedding is extracted based on the pre-trained KataGo residual neural network; The features of the first game are embedded into the output of a trainable neural network. Φ Obtain the feature embedding of the second chess game; A loss function for the global model is constructed based on the similarity between the feature embeddings of the second game and the corresponding comment feature embeddings. When the global model loss function converges to a predetermined accuracy or reaches a predetermined number of training rounds, the trained neural network is obtained. Φ Based on the trained game feature extraction network, feature embeddings of the current game and historical game features are extracted. These historical game features are then sequentially passed through a pre-trained KataGo residual deep neural network and a fully trained neural network. Φ The feature embeddings are obtained, and the game index and corresponding comments are stored; for the current game, a pre-trained KataGo residual deep neural network and a fully trained neural network are used. Φ Obtain the feature embedding; Calculate the similarity between the current game feature embedding and the feature embedding of each historical game. Based on the similarity calculation results, obtain similar games and find the corresponding comments based on the index of the similar games. The AI ​​prediction of the current game and the index of similar games are used to find corresponding comments to form prompt words; The game analysis results for the current chess position are obtained based on the prompt words and the large language model.

2. The intelligent Go game analysis method based on multimodal retrieval enhancement as described in claim 1, characterized in that, The loss function for the global model is: ， ， in, This is the temperature parameter of the neural network, and the key-value pairs of historical chess games and comments are ( , ), Embedding features for the second game. Embedding for comment features; For any comment in the comment collection, The number of historical chess games.

3. The intelligent Go game analysis method based on multimodal retrieval enhancement as described in claim 1, characterized in that, After obtaining the historical and current game positions, convert them into the SGF standard format.

4. The intelligent Go game analysis method based on multimodal retrieval enhancement as described in claim 1, characterized in that, The large language model used is GPT-4o or Tongyi 1000 Questions 2.

5.

5. The intelligent Go game analysis method based on multimodal retrieval enhancement as described in claim 1, characterized in that, The current game AI prediction results are obtained through KataGo analysis, including win rate analysis and next move prediction.

6. An intelligent Go game analysis system based on multimodal retrieval enhancement, characterized in that, include: The chess game feature extraction network training module is configured to acquire historical chess games and corresponding comments, and extract feature embeddings from historical chess games and corresponding comments. The game feature extraction network is trained based on the similarity between historical game feature embeddings and corresponding comment feature embeddings to obtain the trained game feature extraction network; wherein, the training of the game feature extraction network based on the similarity between historical game feature embeddings and corresponding comment feature embeddings to obtain the trained game feature extraction network includes: The first game feature embedding is extracted based on the pre-trained KataGo residual neural network; The features of the first game are embedded into the output of a trainable neural network. Φ Obtain the feature embedding of the second chess game; A loss function for the global model is constructed based on the similarity between the feature embeddings of the second game and the corresponding comment feature embeddings. When the global model loss function converges to a predetermined accuracy or reaches a predetermined number of training rounds, the trained neural network is obtained. Φ ; The feature extraction module is configured to extract feature embeddings of the current game and historical games based on the trained game feature extraction network. Historical game features are then sequentially processed through a pre-trained KataGo residual deep neural network and a fully trained neural network. Φ The feature embeddings are obtained, and the game index and corresponding comments are stored; for the current game, a pre-trained KataGo residual deep neural network and a fully trained neural network are used. Φ Obtain the feature embedding; The prompt word generation module is configured to calculate the similarity between the feature embedding of the current game and the feature embedding of each historical game, obtain similar games based on the similarity calculation results, and find the corresponding comments based on the index of similar games; the prompt word is composed of the AI ​​prediction result of the current game and the corresponding comments found by the index of similar games. The game commentary generation module is configured to generate game analysis results for the current game based on prompt words and a large language model.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps in the intelligent Go game analysis method based on multimodal retrieval enhancement as described in any one of claims 1-5.

8. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the intelligent Go game analysis method based on multimodal retrieval enhancement as described in any one of claims 1-5.

9. A program product, said program product being a computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps in the intelligent Go game analysis method based on multimodal retrieval enhancement as described in any one of claims 1-5.

Images