Large model thinking map prompt learning system for common sense questions and answers

Abstract

The invention belongs to the technical field of large model thinking maps, and discloses a large model thinking map prompt learning system for common sense questions and answers. The big model thinking map prompt learning system for common sense questions and answers provided by the invention comprises a knowledge map generation module, a thinking map reasoning module, a prompt learning module, a domain adaptation module and a reasoning interpretability module. The invention provides a large model thinking map prompt learning system for common knowledge questions and answers, and the system improves the answering ability of a common knowledge question and answer model through multi-level knowledge structure reasoning in a mode of combining the thinking map and prompt learning. The system not only enhances the inference ability for complex common sense questions, but also improves the accuracy of domain common sense questions and answers by constructing a domain-specific thinking map, and provides the interpretability of the inference process.

Description

Technical Field

[0001] The present invention belongs to the technical field of large model mind maps, and in particular relates to a large model mind map prompt learning system for common sense question and answer. Background Art

[0002] Prior art: Common sense question answering system based on pre-trained language model

[0003] Current common sense question answering systems are mainly based on large-scale pre-trained language models (such as GPT, BERT, etc.). These models can capture a large amount of language patterns and semantic information through unsupervised learning on large-scale text corpora, thus having strong question-answering capabilities. However, existing common sense question answering systems still have some significant technical bottlenecks and challenges:

[0004] Lack of multi-level reasoning ability for complex questions: Although current pre-trained language models have the ability to answer common sense questions, they rely more on superficial semantic associations and lack multi-level reasoning ability. When faced with complex common sense questions, the model tends to directly generate answers based only on the context, rather than gradually deriving results through logical reasoning.

[0005] Knowledge fragmentation and lack of systematicity: During the reasoning process, the existing common sense question-answering model cannot effectively combine relevant information from different fields and different knowledge points to answer cross-domain common sense questions. This is because the knowledge learned during pre-training is fragmented and lacks an understanding of the systematic relationship between various knowledge points.

[0006] Lack of sensitivity to domain-specific common sense: Common sense question answering problems cover a wide range of fields, and common sense in different fields often has different reasoning logic and thinking patterns. However, existing models are not specially trained or enhanced for certain specific fields, resulting in weak performance when facing domain-specific common sense questions.

[0007] The reasoning process lacks explainability: Most existing pre-trained models are "black box" operations, and it is difficult for users to understand how the model derives the answer from the input question. Especially in common sense question-answering scenarios, the reasoning process needs to have a certain degree of explainability to enhance user trust.

[0008] Through the above analysis, the problems and defects of the prior art are as follows:

[0009] Existing common-sense question-answering systems lack the ability to reason at multiple levels about complex common-sense questions, lack knowledge fragmentation and systematic connections, lack domain-specific common-sense reasoning, and the reasoning process is unexplainable. Summary of the invention

[0010] In view of the problems existing in the prior art, the present invention provides a large-model mind map prompt learning system for common sense question and answer.

[0011] The present invention is implemented in this way: a large model mind map prompt learning system for common sense question answering includes:

[0012] A knowledge graph generation module is connected to the mind map reasoning module and the reasoning explainability module. The knowledge graph generation module is used to automatically extract knowledge from a large-scale corpus through a natural language processing (NLP) algorithm to construct a common sense knowledge graph in multiple fields. The knowledge graph contains multiple nodes, each node represents an entity, and the edges between the nodes represent the common sense reasoning relationship between the entities. The extracted entities and relationships are represented and learned using a graph neural network (GNN) or other graph structure learning algorithms to form a vectorized knowledge graph that can be used for reasoning.

[0013] The mind map reasoning module is connected with the knowledge map generation module, the prompt learning module, the domain adaptation module, and the reasoning interpretability module. The mind map reasoning module is used to automatically generate a dynamically updated mind map based on the questions raised by the user by using a reinforcement learning algorithm or a reasoning model based on deep learning (such as a neural symbolic reasoning model). The mind map displays multi-level information related to the question in a hierarchical and logical manner. The reasoning path is optimized by combining a Bayesian network or a path planning algorithm to ensure the rationality and accuracy of the reasoning path;

[0014] A prompt learning module is connected to the mind map reasoning module and the domain adaptation module. The prompt learning module interacts with a pre-trained language model (such as the GPT series and the BERT model) and combines the prompt learning technology to input each node information and relationship chain information into the pre-trained language model in the form of a prompt based on the generated mind map. The model performs reasoning and answering based on the prompt. The prompt can be adaptively adjusted through training to improve the accuracy of the answer. The specific prompt learning strategies include soft prompts and reinforcement prompts.

[0015] The domain adaptation module is connected with the mind map reasoning module, the prompt learning module, and the reasoning explainability module. The domain adaptation module implements adaptive training of the system for a specific domain through knowledge distillation or domain transfer learning technology. When answering questions, it generates a mind map for a specific domain through a deep domain adaptation algorithm based on a pre-stored domain knowledge base or domain expert data, and optimizes prompts in combination with a domain-specific model during the reasoning process.

[0016] The reasoning explainability module is connected with the knowledge graph generation module, the mind map reasoning module, and the domain adaptation module. The reasoning explainability module records and tracks the node reasoning process in the mind map through heuristic rules or logistic regression models to form a reasoning chain that can be reviewed by users. At the same time, it combines visualization algorithms (such as T-SNE or UMAP) to generate a visualized mind map to display the reasoning path, thereby enhancing the transparency and explainability of the system.

[0017] Data module: The system relies on large-scale multi-domain datasets for training and fine-tuning various models involved in the system. The data includes public common sense corpora, domain expert data, and data generated by user interactions. The system optimizes model performance by combining large-scale unsupervised pre-training and supervised fine-tuning, and improves the generalization ability of the inference model through data augmentation technology.

[0018] Algorithm module, the algorithms used by the system include but are not limited to graph neural network (GNN) for the generation and update of knowledge graphs, deep reinforcement learning (DRL) for the optimization of mind map reasoning paths, prompt learning algorithm for guiding the pre-trained language model to generate reasoning answers, transfer learning algorithm for domain adaptation, and logistic regression, neural symbolic reasoning and visualization algorithms for the interpretability of the reasoning process.

[0019] Furthermore, the knowledge graph generation module method is as follows:

[0020] First, common sense knowledge is extracted from a large-scale corpus and a multi-domain knowledge graph is constructed; this knowledge graph contains not only individual entities and facts, but also common sense reasoning relationships between different entities;

[0021] During the graph generation process, the system uses natural language processing technology to parse the text, identify knowledge points, and associate them through relationship extraction technology to build a complete knowledge network;

[0022] Entities represent concepts or objects in the system, such as "animals", "plants", "solar system", while relationships define how entities are connected, such as causal relationships, attribution relationships, and time order;

[0023] Extract key information from the corpus and generate nodes and edges of the knowledge graph;

[0024] Nodes represent entities or facts, and edges represent reasoning relationships or causal relationships;

[0025] This module outputs a domain-wide common-sense knowledge graph for reasoning in subsequent modules.

[0026] Further, the mind map reasoning module method is as follows:

[0027] When a user asks a question, the system first analyzes the question and uses the entities and relationships in the knowledge graph to generate a mind map related to the question;

[0028] The mind map is a hierarchical structure that logically displays multi-level information related to the problem. When the user asks "Why do humans need sleep?", the system will generate a mind map related to "sleep", covering different levels of information such as "physiological functions of sleep", "impacts of insufficient sleep", and "brain operation".

[0029] The system uses reasoning algorithms to find nodes related to the question in the knowledge graph, and then displays this information in a hierarchical manner; this mind map not only helps the system reason, but also provides rich context for the final prompt learning;

[0030] Parse user questions and locate relevant entities and relationships;

[0031] Generate a multi-level mind map based on the information in the knowledge map to show the logical relationship behind the problem;

[0032] The output contains a mind map structure with multiple levels of nodes, which is used as the input for subsequent prompt learning modules.

[0033] Further, the prompt learning module method is as follows:

[0034] Based on the pre-trained language model, combined with the mind map, prompts are generated and the model is guided to generate answers with logical reasoning capabilities;

[0035] The system extracts key information from the mind map, including the knowledge points of the nodes and the reasoning relationship between the nodes, and then inputs this information into the pre-trained model in the form of prompts; the prompts can include the context of the question, related knowledge points, and reasoning chains;

[0036] Extract key information from the mind map, including nodes and relationship chains;

[0037] Convert the extracted key information into prompt sentences and input them into the language model;

[0038] The language model generates more reasoned and logical answers based on the prompts.

[0039] Further, the domain adaptation module method is as follows:

[0040] Used to handle problems in specific fields, such as medicine, law, and finance. These fields have unique common sense and reasoning logic, so the system will generate customized mind maps for these problems and provide optimized prompts through the domain knowledge base;

[0041] The system calls the corresponding domain knowledge base according to the domain attributes of the problem to ensure that domain-specific knowledge nodes and reasoning relationships are included when generating the mind map. For example, when the system processes a question like "Why does the heart beat?", it will first extract information related to the heart and physiological operations from the knowledge base in the medical field to generate a mind map specifically for the medical field.

[0042] When generating prompts, the system also adjusts the prompt strategy to better suit the reasoning needs of the field, such as guiding the language model to use specific medical terms or logical reasoning patterns;

[0043] Detect the domain to which the problem belongs and call the domain knowledge base to generate a customized mind map;

[0044] Include domain-related nodes and relationship chains in the mind map to ensure that learning prompts focus on domain common sense;

[0045] Generate optimization hints that enable pre-trained models to provide accurate answers in specific domains.

[0046] Further, the reasoning explainability module method is as follows:

[0047] By recording and visualizing the reasoning process of the mind map, an explanation of the answers generated by the system is provided; this module not only shows how the model starts from the question and gets the answer through multi-level reasoning, but also shows the logical relationship behind each step of reasoning;

[0048] The reasoning path for each question is recorded, and users can view the reasoning chain the system uses to generate answers through a visual interface. When users ask, "Why do plants need sunlight?", the system will display a related mind map to explain the principles of plant photosynthesis and how to derive "photosynthesis" and "energy acquisition" information from "sunlight";

[0049] Record the key nodes and reasoning chains in each model reasoning process;

[0050] Visualize the reasoning chain in the mind map and show it to the user;

[0051] Provide explanations of the reasoning process to help users understand the model's answer logic.

[0052] Another object of the present invention is to provide a large model mind map prompt learning method for common sense question answering, comprising:

[0053] Step 1: Extract knowledge from large-scale corpus through the knowledge graph generation module to build a common sense knowledge graph in multiple fields; the graph contains entities and facts, and common sense reasoning relationships between entities;

[0054] Step 2: Based on the question raised by the user, the mind map reasoning module generates a mind map using the knowledge map. The map displays multi-level information and reasoning paths related to the question in a hierarchical and logical way.

[0055] Step 3: Based on the pre-trained language model, the prompt learning module combines the prompt learning technology to input the node information and relationship chain information in the mind map into the language model through prompts to guide the model to give a more reasoning answer;

[0056] Step 4: For questions in a specific field, the system generates a customized mind map from the corresponding domain knowledge base through the domain adaptation module, and optimizes the prompts through the domain adaptation module to ensure that the model can accurately answer domain-specific common sense questions;

[0057] Step 5, record and visualize the reasoning process of each node in the mind map through the reasoning explainability module, and provide the reasoning chain associated with the answer generated by the model for users to review and reference, thereby enhancing the transparency and explainability of the system.

[0058] Furthermore, the knowledge graph generation module further includes a named entity recognition (NER) algorithm and a relationship extraction (RE) algorithm based on deep learning, which are used to automatically extract entities and relationships between entities from unstructured text. The entities and relationships are mapped to a vector space, and a high-dimensional vector representation that can be used for reasoning is generated through a graph embedding algorithm (Graph Embedding), thereby enhancing the accuracy and generalization ability of reasoning.

[0059] Furthermore, the prompt learning technology adopted by the prompt learning module includes a self-supervised learning mechanism, which automatically adjusts the structure and content of the prompt input by generating pseudo-label data for the model, and updates the prompt strategy using the reward signal of reinforcement learning, thereby improving the performance of the pre-trained model in complex common sense reasoning tasks.

[0060] Furthermore, the reasoning explainability module further combines a visualized heat map to display the contribution of different reasoning nodes, which is calculated by Shapley value or attention mechanism, so that users can intuitively understand the key decision nodes and information sources in the reasoning process.

[0061] Another object of the present invention is to provide a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the large-model mind map prompt learning method for common sense questioning and answering.

[0062] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, enables the processor to execute the steps of the large-model mind map prompt learning method for common sense question and answer.

[0063] Another object of the present invention is to provide an information data processing terminal, which is used to implement the large-model mind map prompt learning system for common sense question and answer.

[0064] In combination with the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solutions to be protected by the present invention are as follows:

[0065] The present invention provides a large-model mind map prompt learning system for common sense question answering. The system uses a combination of mind map and prompt learning to improve the answering ability of the common sense question answering model through multi-level knowledge structure reasoning. The system not only enhances the reasoning ability of complex common sense questions, but also improves the accuracy of domain common sense question answering by constructing domain-specific mind maps, and provides the explainability of the reasoning process.

[0066] 1. Improvement of multi-level reasoning ability:

[0067] The present invention significantly enhances the multi-level reasoning ability of the model by combining mind maps and prompt learning. Unlike traditional language models that generate answers based only on semantic information, this system can generate a logical reasoning chain based on the question, ensuring that complex questions can be answered step by step through multiple reasoning steps, thus improving the system's ability to handle complex common sense questions.

[0068] 2. Enhanced knowledge systematization:

[0069] By introducing the knowledge graph, the system can organize common sense knowledge in a structured manner and generate a systematic mind map. This improvement solves the problem of knowledge fragmentation in existing technologies, allowing the system to efficiently integrate knowledge points from different fields when answering cross-domain questions and provide more systematic and comprehensive answers.

[0070] 3. Improved sensitivity to domain common sense:

[0071] The system uses a domain adaptation module to generate customized mind maps for problems in different fields and adjust the prompt learning strategy. This domain-specific optimization improves the model's sensitivity to common sense questions in specific fields, ensuring that more professional answers can be provided for problems in professional fields such as medicine and law.

[0072] 4. Explainability of the reasoning process:

[0073] Most existing pre-trained models are "black box operations" and it is difficult for users to understand their reasoning process. This system makes the reasoning process highly explainable through the visualization of the mind map reasoning chain. Users can see how the model reasoned and got the answer step by step, which enhances users' trust in the system and facilitates debugging and improvement in error situations.

[0074] 5. Automation and intelligent optimization:

[0075] The system introduces a feedback mechanism that can self-optimize the prompt learning and mind map generation strategies based on user feedback. This automated learning and optimization process not only improves the system's question-answering quality, but also continuously improves its reasoning ability and accuracy in the long run.

[0076] Through this mind map prompt learning system, the common sense question answering model has stronger reasoning ability, domain common sense sensitivity, and systematic knowledge integration ability. In addition, the visualization of the reasoning chain provides users with explainability, significantly improving users' trust in the system and usage experience. The system has made significant technical progress in handling complex common sense question answering problems and provides strong support for question answering applications in various fields. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] Figure 1 It is a structural block diagram of a large-model mind map prompt learning system for common sense question and answering provided by an embodiment of the present invention.

[0078] Figure 2 It is a flow chart of the knowledge graph generation module method provided by an embodiment of the present invention.

[0079] Figure 3 It is a flow chart of a large-model mind map prompt learning method for common sense question and answering provided by an embodiment of the present invention.

[0080] Figure 1 In: 1. Knowledge graph generation module; 2. Mind map reasoning module; 3. Prompt learning module; 4. Domain adaptation module; 5. Reasoning explainability module. DETAILED DESCRIPTION

[0081] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0082] The large-model mind map prompt learning system for common sense question answering provided by the present invention comprises:

[0083] In the knowledge graph generation module, advanced natural language processing (NLP) algorithms, such as BERT variants or Transformer models, are used to efficiently extract entities and relationships from large-scale multi-domain corpora and build graphs containing rich common sense knowledge.

[0084] Using variants of graph neural networks (GNNs), such as Graph Convolutional Networks (GCN) or Graph Attention Networks (GAT), representation learning is performed on the constructed knowledge graph to convert entities and relations into vector form for use in subsequent reasoning;

[0085] Implement a dynamic knowledge graph update mechanism to maintain the timeliness and comprehensiveness of the knowledge graph by continuously learning new common sense information, ensuring that the system can handle the latest common sense question and answer needs.

[0086] The large-model mind map prompt learning system for common sense question answering provided by the present invention combines deep reinforcement learning (DRL) algorithms, such as Deep Q-Networks (DQN) or Policy Gradient method, in the mind map reasoning module to optimize the mind map generation process and ensure that the reasoning path is both efficient and accurate.

[0087] The prompt learning module uses advanced soft prompt technology, such as continuous prompt vectors and reinforcement prompting strategies, to iteratively optimize prompt content and improve the pre-trained language model's ability to understand and utilize mind map information.

[0088] Realize the dynamic adjustment function of the mind map, automatically adjust the structure and content of the mind map according to user feedback and reasoning results, so as to improve the accuracy and adaptability of the system in answering future problems.

[0089] The large-model mind map prompt learning system for common sense question answering provided by the present invention adopts knowledge distillation technology in the domain adaptation module, and uses a small model to imitate the behavior of the large model to achieve efficient model deployment in a specific domain. At the same time, it uses a domain transfer learning algorithm, such as Transfer Learning with Adapters, to quickly adapt the system to a new domain.

[0090] The reasoning explainability module uses a fine-grained heuristic rule set and a logistic regression model to record the reasoning basis and path selection of each node in the mind map in detail and generate a detailed reasoning chain report;

[0091] Combined with advanced visualization technologies, such as D3.js or Plotly, a user-friendly interface is developed to intuitively display the reasoning path and node relationships of the mind map, enhancing users' trust and understanding of the system reasoning process.

[0092] like Figure 1 As shown, a large model mind map prompt learning system for common sense question answering provided by an embodiment of the present invention includes:

[0093] Knowledge graph generation module 1, mind map reasoning module 2, prompt learning module 3, domain adaptation module 4, reasoning explainability module 5.

[0094] The knowledge graph generation module 1 is connected with the mind map reasoning module 2 and the reasoning explainability module 5 to extract knowledge from large-scale corpus and construct a common sense knowledge graph in multiple fields; the graph contains entities and facts, and the common sense reasoning relationship between entities;

[0095] The mind map reasoning module 2 is connected with the knowledge map generation module 1, the prompt learning module 3, the domain adaptation module 4, and the reasoning explainability module 5, and is used to generate a mind map using the knowledge map according to the question raised by the user. The map displays the multi-level information and reasoning path related to the question in a hierarchical and logical way;

[0096] The prompt learning module 3 is connected with the mind map reasoning module 2 and the domain adaptation module 4, and is used to input the node information and relationship chain information in the mind map into the language model through prompts based on the pre-trained language model and combined with the prompt learning technology, so as to guide the model to give a more reasoning answer;

[0097] The domain adaptation module 4 is connected with the mind map reasoning module 2, the prompt learning module 3, and the reasoning explainability module 5. For problems in a specific domain, the system generates a customized mind map from the corresponding domain knowledge base and optimizes prompts through the domain adaptation module to ensure that the model can accurately answer domain-specific common sense questions;

[0098] The reasoning explainability module 5 is connected with the knowledge graph generation module 1, the mind map reasoning module 2 and the domain adaptation module 4, and is used to record and visualize the reasoning process of each node in the mind map, and provide the reasoning chain associated with the answer generated by the model for users to review and refer to, thereby enhancing the transparency and explainability of the system.

[0099] like Figure 2 As shown, the knowledge graph generation module method provided by the embodiment of the present invention is as follows:

[0100] S101 first extracts common sense knowledge from a large-scale corpus and constructs a multi-domain knowledge graph; this knowledge graph contains not only individual entities and facts, but also common sense reasoning relationships between different entities;

[0101] S102, during the graph generation process, the system will use natural language processing technology to parse the text, identify knowledge points, and associate them through relationship extraction technology to build a complete knowledge network;

[0102] S103, entities represent concepts or objects in the system, such as "animals", "plants", and "solar system", while relationships define how entities are connected, such as causal relationships, attributive relationships, and temporal order;

[0103] S104, extract key information from the corpus and generate nodes and edges of the knowledge graph; nodes represent entities or facts, and edges represent reasoning relationships or causal relationships; this module outputs a domain-wide common-sense knowledge graph for subsequent module reasoning.

[0104] The mind map reasoning module method provided by the embodiment of the present invention is as follows:

[0105] When a user asks a question, the system first analyzes the question and uses the entities and relationships in the knowledge graph to generate a mind map related to the question;

[0106] The mind map is a hierarchical structure that logically displays multi-level information related to the problem. When the user asks "Why do humans need sleep?", the system will generate a mind map related to "sleep", covering different levels of information such as "physiological functions of sleep", "impacts of insufficient sleep", and "brain operation".

[0107] The system uses reasoning algorithms to find nodes related to the question in the knowledge graph, and then displays this information in a hierarchical manner; this mind map not only helps the system reason, but also provides rich context for the final prompt learning;

[0108] Parse user questions and locate relevant entities and relationships;

[0109] Generate a multi-level mind map based on the information in the knowledge map to show the logical relationship behind the problem;

[0110] The output contains a mind map structure with multiple levels of nodes, which is used as the input for subsequent prompt learning modules.

[0111] The prompt learning module method provided by the embodiment of the present invention is as follows:

[0112] Based on the pre-trained language model, combined with the mind map, prompts are generated and the model is guided to generate answers with logical reasoning capabilities;

[0113] The system extracts key information from the mind map, including the knowledge points of the nodes and the reasoning relationship between the nodes, and then inputs this information into the pre-trained model in the form of prompts; the prompts can include the context of the question, related knowledge points, and reasoning chains;

[0114] Extract key information from the mind map, including nodes and relationship chains;

[0115] Convert the extracted key information into prompt sentences and input them into the language model;

[0116] The language model generates more reasoned and logical answers based on the prompts.

[0117] The domain adaptation module method provided by the embodiment of the present invention is as follows:

[0118] Used to handle problems in specific fields, such as medicine, law, and finance. These fields have unique common sense and reasoning logic, so the system will generate customized mind maps for these problems and provide optimized prompts through the domain knowledge base;

[0119] The system calls the corresponding domain knowledge base according to the domain attributes of the problem to ensure that domain-specific knowledge nodes and reasoning relationships are included when generating the mind map. For example, when the system processes a question like "Why does the heart beat?", it will first extract information related to the heart and physiological operations from the knowledge base in the medical field to generate a mind map specifically for the medical field.

[0120] When generating prompts, the system also adjusts the prompt strategy to better suit the reasoning needs of the field, such as guiding the language model to use specific medical terms or logical reasoning patterns;

[0121] Detect the domain to which the problem belongs and call the domain knowledge base to generate a customized mind map;

[0122] Include domain-related nodes and relationship chains in the mind map to ensure that learning prompts focus on domain common sense;

[0123] Generate optimization hints that enable pre-trained models to provide accurate answers in specific domains.

[0124] The reasoning explainability module method provided by the embodiment of the present invention is as follows:

[0125] By recording and visualizing the reasoning process of the mind map, an explanation of the answers generated by the system is provided; this module not only shows how the model starts from the question and gets the answer through multi-level reasoning, but also shows the logical relationship behind each step of reasoning;

[0126] The reasoning path for each question is recorded, and users can view the reasoning chain the system uses to generate answers through a visual interface. When users ask, "Why do plants need sunlight?", the system will display a related mind map to explain the principles of plant photosynthesis and how to derive "photosynthesis" and "energy acquisition" information from "sunlight";

[0127] Record the key nodes and reasoning chains in each model reasoning process;

[0128] Visualize the reasoning chain in the mind map and show it to the user;

[0129] Provide explanations of the reasoning process to help users understand the model's answer logic.

[0130] like Figure 3 As shown, a large model mind map prompt learning method for common sense question answering provided by an embodiment of the present invention includes:

[0131] S201, extracting knowledge from large-scale corpus through the knowledge graph generation module to build a common sense knowledge graph in multiple fields; the graph contains entities and facts, and common sense reasoning relationships between entities;

[0132] S202, using the mind map reasoning module to generate a mind map based on the question raised by the user using the knowledge map, and the map displays multi-level information and reasoning paths related to the question in a hierarchical and logical manner;

[0133] S203, using a prompt learning module based on a pre-trained language model and combined with a prompt learning technology, the node information and relationship chain information in the mind map are input into the language model through prompts to guide the model to give a more reasoning answer;

[0134] S204, for questions in a specific domain, the system generates a customized mind map from the corresponding domain knowledge base through the domain adaptation module, and performs prompt optimization through the domain adaptation module to ensure that the model can accurately answer domain-specific common sense questions;

[0135] S205, records and visualizes the reasoning process of each node in the mind map through the reasoning explainability module, and provides the reasoning chain associated with the answer generated by the model for user review and reference, thereby enhancing the transparency and explainability of the system.

[0136] Another object of the present invention is to provide a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the large-model mind map prompt learning method for common sense questioning and answering.

[0137] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, enables the processor to execute the steps of the large-model mind map prompt learning method for common sense question and answer.

[0138] Another object of the present invention is to provide an information data processing terminal, which is used to implement the large-model mind map prompt learning system for common sense question and answer.

[0139] The present invention provides a large-model mind map prompt learning system for common sense question answering. The system uses a combination of mind map and prompt learning to improve the answering ability of the common sense question answering model through multi-level knowledge structure reasoning. The system not only enhances the reasoning ability of complex common sense questions, but also improves the accuracy of domain common sense question answering by constructing domain-specific mind maps, and provides the explainability of the reasoning process.

[0140] The system consists of the following modules:

[0141] 1. Knowledge graph generation module:

[0142] Extract knowledge from large-scale corpora and build common-sense knowledge graphs in multiple fields. The graphs contain not only entities and facts, but also common-sense reasoning relationships between entities.

[0143] 2. Mind map reasoning module:

[0144] Based on the questions raised by the user, a mind map (MindMap) is generated using the knowledge graph, which displays the multi-level information and reasoning paths related to the question in a hierarchical and logical manner.

[0145] 3. Prompt learning module:

[0146] Based on pre-trained language models (such as GPT-4, T5, etc.), combined with prompt learning technology, the node information, relationship chain information, etc. in the mind map are input into the language model through prompts to guide the model to give more reasoning answers.

[0147] 4. Domain Adaptation Module:

[0148] For questions in specific fields, the system generates customized mind maps from the corresponding domain knowledge base, and optimizes prompts through the domain adaptation module to ensure that the model can accurately answer domain-specific common sense questions.

[0149] 5. Reasoning Explainability Module:

[0150] This module records and visualizes the reasoning process of each node in the mind map, and provides the reasoning chain associated with the answers generated by the model for users to review and reference, enhancing the transparency and explainability of the system.

[0151] The specific working principle of the large-scale mind map prompt learning system for common sense question answering can be understood from the functions and interaction processes of its various modules, covering knowledge extraction, mind map generation, prompt learning, and the generation and interpretation of reasoning results. The following is a detailed description of the system.

[0152] Working principle:

[0153] 1. Knowledge graph generation module

[0154] Working principle:

[0155] The system first extracts common sense knowledge from a large-scale corpus and constructs a multi-domain knowledge graph. This knowledge graph contains not only individual entities and facts (such as "a dog is an animal"), but also common sense reasoning relationships between different entities (such as "a dog needs food to survive").

[0156] During the graph generation process, the system uses natural language processing technology to parse the text, identify knowledge points (such as entities, attributes, actions, etc.), and associate them through relationship extraction technology to build a complete knowledge network.

[0157] Entities represent concepts or objects in the system, such as "animal", "plant", "solar system", etc., while relationships define how entities are connected, such as causal relationships, attribution relationships, time order, etc.

[0158] step:

[0159] Extract key information from the corpus and generate nodes and edges of the knowledge graph.

[0160] Nodes represent entities or facts, and edges represent inference relationships or causal relationships.

[0161] This module outputs a domain-wide common-sense knowledge graph for reasoning in subsequent modules.

[0162] 2. Mind map reasoning module

[0163] Working principle:

[0164] When a user asks a question, the system will first parse the question and use the entities and relationships in the knowledge graph to generate a mind map related to the question.

[0165] A mind map is a hierarchical structure that logically displays multiple levels of information related to a question. For example, when a user asks "Why do humans need sleep?", the system will generate a mind map related to "sleep", covering different levels of information such as "physiological functions of sleep", "impacts of insufficient sleep", and "brain operation".

[0166] The system uses reasoning algorithms to find nodes related to the question in the knowledge graph and then displays this information in a hierarchical manner. This mind map not only helps the system reason, but also provides rich context for the final prompt learning.

[0167] step:

[0168] Parse user questions and locate relevant entities and relationships.

[0169] Generate a multi-level mind map based on the information in the knowledge graph to show the logical relationship behind the problem.

[0170] The output contains a mind map structure with multiple levels of nodes, which is used as the input for subsequent prompt learning modules.

[0171] 3. Prompt learning module

[0172] Working principle:

[0173] The prompt learning module is the core of the entire system. It generates prompts based on a pre-trained language model (such as GPT-4 or T5) combined with a mind map, and guides the model to generate answers with logical reasoning capabilities.

[0174] The system extracts key information from the mind map, including the knowledge points of the nodes, the reasoning relationship between the nodes (such as causal chain, time sequence, etc.), and then inputs this information into the pre-trained model in the form of prompts. The prompt content can include the context of the question, relevant knowledge points, and reasoning chains.

[0175] By prompting the model, the pre-trained model can not only generate direct answers based on the questions, but also consider the reasoning chain and background knowledge in the process of generating answers to ensure the logic and rationality of the answers.

[0176] step:

[0177] Extract key information from mind maps, including nodes, relationship chains, etc.

[0178] The extracted key information is converted into prompt sentences and input into the language model.

[0179] The language model generates more reasoned and logical answers based on the prompts.

[0180] 4. Domain Adaptation Module

[0181] Working principle:

[0182] The domain adaptation module is used to handle problems in specific fields, such as medicine, law, finance, etc. These fields have unique common sense and reasoning logic, so the system will generate customized mind maps for these problems and optimize prompts through the domain knowledge base.

[0183] The system calls the corresponding domain knowledge base according to the domain attributes of the problem to ensure that domain-specific knowledge nodes and reasoning relationships are included when generating the mind map. For example, when the system processes a question like "Why does the heart beat?", it will first extract information related to the heart and physiological operations from the knowledge base in the medical field and generate a mind map specifically for the medical field.

[0184] The system also adjusts the prompt strategy when generating prompts to better suit the reasoning needs of the domain, such as guiding the language model to use specific medical terms or logical reasoning patterns.

[0185] step:

[0186] Detect the domain to which the problem belongs and call the domain knowledge base to generate a customized mind map.

[0187] Include domain-related nodes and relationship chains in the mind map to ensure that learning prompts focus on domain common sense.

[0188] Generate optimization hints that enable pre-trained models to provide accurate answers in specific domains.

[0189] 5. Reasoning Explainability Module

[0190] Working principle:

[0191] The reasoning explainability module provides an explanation of the answers generated by the system by recording and visualizing the reasoning process of the mind map. This module not only shows how the model starts from the question and gets the answer through multi-level reasoning, but also shows the logical relationship behind each step of reasoning.

[0192] The reasoning path for each question is recorded, and users can view the reasoning chain that the system relies on when generating answers through a visual interface. For example, when a user asks "Why do plants need sunlight?", the system will display a related mind map to explain the principle of plant photosynthesis and how to derive "photosynthesis" and "energy acquisition" from "sunlight".

[0193] This explainability greatly enhances the transparency of the system, helps users understand the model's reasoning process, and also facilitates debugging and improvement of the system when it gives incorrect answers.

[0194] step:

[0195] Record the key nodes and reasoning chains in each model reasoning process.

[0196] Visualize the reasoning chain in the mind map and show it to the user.

[0197] Provide explanations of the reasoning process to help users understand the model's answer logic.

[0198] The overall workflow of the system

[0199] 1. Question analysis and knowledge extraction: After receiving the question raised by the user, the system first extracts relevant knowledge points and reasoning relationships from the multi-domain knowledge base through the knowledge graph generation module. Based on this information, a preliminary mind map related to the question is generated.

[0200] 2. Mind map generation and reasoning chain construction: The mind map reasoning module displays these knowledge points in a hierarchical and logical manner and constructs a reasoning chain related to the problem. The mind map displays information at different levels, such as the core knowledge points of the problem and the secondary conclusions and reasoning paths derived from them.

[0201] 3. Prompt learning and answer generation: The prompt learning module generates prompt sentences based on the nodes and relationship chains in the mind map and inputs them into the pre-trained language model. Under the guidance of the prompts, the model generates answers with multi-level reasoning logic.

[0202] 4. Domain adaptation and optimization: If the question belongs to a specific domain, the domain adaptation module calls the corresponding domain knowledge base, generates a domain-specific mind map and optimizes the prompts to ensure that the model can output answers that are consistent with domain common sense.

[0203] 5. Reasoning chain explainability display: The reasoning explainability module records and visualizes the reasoning chain. Users can view each reasoning step, understand how the system arrives at the answer, and review the reasoning process.

[0204] Significant technological advancement brought by the system

[0205] 1. Multi-level reasoning ability: Through the hierarchical display of mind maps, the system can perform multi-level reasoning when answering common sense questions, solving the problem of lack of reasoning depth in traditional systems.

[0206] 2. Flexible adaptation of domain knowledge: The domain adaptation module enables the system to cope with problems in different fields, ensuring that it is more accurate when answering common sense questions in specific fields.

[0207] 3. Prompt learning improves reasoning accuracy: By combining prompt learning, the system effectively guides the pre-trained language model to generate more logical and reasoning answers, avoiding simple surface-level association answers.

[0208] 4. Explainability enhances user trust: The reasoning explainability module provides a transparent display of the reasoning path, enabling users to understand the system's reasoning process, greatly improving users' trust in the system's answers.

[0209] 5. Reasoning chain explainability display: The reasoning explainability module records and visualizes the reasoning chain. Users can view each reasoning step, understand how the system arrives at the answer, and review the reasoning process.

[0210] Through the synergistic effect of the above working principles, the large-model mind map prompt learning system for common sense question answering has greatly improved the accuracy, efficiency and explainability of common sense question answering.

[0211] Significant technological advancement brought by the system

[0212] 1. Multi-level reasoning ability: Through the hierarchical display of mind maps, the system can perform multi-level reasoning when answering common sense questions, solving the problem of lack of reasoning depth in traditional systems.

[0213] 2. Flexible adaptation of domain knowledge: The domain adaptation module enables the system to cope with problems in different fields, ensuring that it is more accurate when answering common sense questions in specific fields.

[0214] 3. Prompt learning improves reasoning accuracy: By combining prompt learning, the system effectively guides the pre-trained language model to generate more logical and reasoning answers, avoiding simple surface-level association answers.

[0215] 4. Explainability enhances user trust: The reasoning explainability module provides a transparent display of the reasoning path, enabling users to understand the system's reasoning process, greatly improving users' trust in the system's answers.

[0216] Through the synergistic effect of the above working principles, the large-model mind map prompt learning system for common sense question answering has greatly improved the accuracy, efficiency and interpretability of common sense question answering.

[0217] Method steps

[0218] 1. Knowledge extraction and mind map generation:

[0219] First, after the system receives the user's question, the knowledge graph generation module extracts relevant common sense facts and knowledge points from the knowledge base. These knowledge points are organized in the form of a graph to form a mind map related to the question. The mind map includes multiple nodes, each node represents a fact or knowledge point, and the nodes are connected by different types of reasoning relationships, such as causal relationships, category relationships, time sequence, etc.

[0220] 2. Mind map prompt generation:

[0221] After the mind map is generated, the prompt learning module will organize the information of each node and edge into natural language prompts according to the structure of the mind map, and input these prompts into the pre-trained language model. The prompt content includes the context of the question, the reasoning relationship chain in the mind map, and the important knowledge nodes related to the answer. This step enhances the model's logical reasoning ability when answering questions by inputting prompt information.

[0222] 3. Field adaptation and adjustment:

[0223] If the problem involves a specific field, the domain adaptation module will call the knowledge base of the field, generate a domain-specific mind map, and optimize the prompts for the domain-specific reasoning logic. For example, in the medical field, the system will generate a pathology-related reasoning chain to enhance the model's reasoning accuracy in this field.

[0224] 4. Model reasoning and answer generation:

[0225] The pre-trained language model combines the prompt information and generates answers through prompt learning. Since the prompt learning module has provided a complete mind map reasoning path, the model's answer is not only based on the context of the question itself, but also combines the multi-level reasoning in the mind map to ensure that the answer is more accurate and logical.

[0226] 5. Reasoning chain verification and explanation:

[0227] The reasoning explainability module records the entire mind map reasoning process of the model generating answers. Users can view the reasoning chain in the mind map through a visual interface to understand each step of the reasoning logic of the model generating answers. The module can also verify the rationality of the reasoning chain to ensure that the answer output by the system is based on the correct reasoning path.

[0228] 6. Feedback and self-optimization:

[0229] The system continuously improves mind map generation and prompt learning strategies through user feedback mechanisms. Users can mark the accuracy of the system's answers, and the system will adjust the mind map structure and prompt generation method based on the feedback data, gradually optimizing the model's performance in specific common sense question-answering scenarios.

[0230] A large model mind map for common sense question answering prompts the detailed signal processing process in the learning system

[0231] The signal processing process of the system starts from the user inputting the question, and then goes through a complete process of multi-level knowledge extraction, reasoning, answer generation, and result feedback. The following are the specific signal processing mechanisms and their operation details in each step:

[0232] 1. Signal processing for knowledge extraction and mind map generation

[0233] User question input: First, the system receives the question input by the user through the keyboard or voice. This question is used as an input signal and enters the language processing module of the system.

[0234] Natural language parsing: The input text signal is parsed by the natural language processing (NLP) module and converted into structured data. This process includes:

[0235] Lexical analysis: Break down the vocabulary of the input question and identify important lexical signals, such as entities (nouns), actions (verbs), etc.

[0236] Syntactic analysis: Analyze the sentence structure and clarify the grammatical relationships such as subject, predicate and object to determine the core intention of the question.

[0237] Semantic understanding: Through the knowledge graph and pre-trained models, the knowledge points in the question and the target areas of required reasoning are identified. For example, when a user asks "Why do plants need sunlight?", the system identifies "plants", "sunlight" and "need" as key nodes.

[0238] Signal mapping to knowledge graph: The system maps the structured signals obtained from the analysis to the knowledge graph module to find knowledge points related to the problem. After signal processing, these knowledge points are converted into nodes (facts) and reasoning links (reasoning links) to generate a preliminary mind map.

[0239] 2. Signal processing of mind map prompt generation

[0240] Node signal conversion to prompt signal: The generated mind map contains multiple nodes and edges, each node represents a knowledge point, and the edge represents the reasoning relationship. The system converts these structural signals into natural language prompts through the following steps:

[0241] Node information processing: The system processes each node in the mind map and converts them into language signals. These signals are facts (such as "photosynthesis requires sunlight") or inference relationships (such as "sunlight provides energy for plants").

[0242] Relationship chain extraction: The system extracts the reasoning relationship chain in the mind map and converts the associations between different nodes into linear prompt language signals. For example, the system will convert "sunlight provides energy to plants" and "photosynthesis produces glucose" as part of the reasoning chain into prompt information.

[0243] Prompt signals are input into the language model: The sorted prompt signals are input into the pre-trained language model (such as GPT-4) through the prompt learning module. These prompt signals have clear reasoning chains and can guide the model to generate logical answers.

[0244] 3. Signal Processing for Domain Adaptation and Adjustment

[0245] Domain Identification: When a question involves a specific field (such as medicine, law, technology, etc.), the system identifies the specific field through key signals in the question. For example, when the system receives a question related to "pathology", it automatically identifies that the question belongs to the medical field.

[0246] Signal adaptation domain knowledge base: The system calls the domain-specific knowledge base signal through the domain adaptation module to regenerate the domain-specific mind map. In this process, signal processing will become more sophisticated, taking into account the unique knowledge and reasoning logic in the domain.

[0247] Domain-specific reasoning chain optimization: When generating prompt signals, the system will optimize the reasoning logic. For example, in the medical field, the system will give priority to the reasoning relationship chain related to the disease and pathological process, and adjust the generation method of prompt signals based on the domain knowledge base.

[0248] 4. Signal Processing for Model Reasoning and Answer Generation

[0249] Language signal generation driven by prompt learning: After the pre-trained language model (such as GPT-4) receives the prompt signal, it starts reasoning processing.

[0250] Multi-level reasoning signal processing: Driven by the prompt information, the signal processing inside the model gradually generates a reasonable answer signal through a multi-level reasoning path. The model first generates a preliminary answer signal, and then continuously adjusts and optimizes it through the internal feedback mechanism until the final answer is generated.

[0251] Context-based signal response: The answer generated by the pre-trained model combined with the prompt signal is not only based on the context of the user input, but also combines the reasoning relationship chain in the prompt signal. For example, for the question "Why do plants need sunlight?", the signal chain generated by the system includes the reasons for photosynthesis, energy sources, etc.

[0252] 5. Signal Processing for Validation and Interpretation of Reasoning Chains

[0253] Reasoning chain recording and signal visualization: The reasoning interpretability module records the entire reasoning signal chain of the model to generate the answer. The system tracks each signal node and generates a signal path based on time sequence. Users can view these signal chains through a visualization interface to understand how the system gradually infers and generates answers from the input question.

[0254] Signal verification and adjustment: The inference chain verification module compares the rationality of the signal chain to ensure that each inference node and edge (signal) is derived based on the correct inference path. If a signal error is found in the inference chain (such as unreasonable reasoning or signal disconnection), the system will send an adjustment signal to optimize the inference path.

[0255] 6. Feedback and self-optimizing signal processing

[0256] Feedback signal capture: The system receives user feedback signals through the user interface. Feedback signals can be scores for answers or markings as correct or incorrect. The system will adjust the strategies for subsequent reasoning graph generation and prompt signal generation based on these feedback signals.

[0257] Signal Self-Optimization:

[0258] The system learns by comparing user feedback signals with the reasoning signals used to generate answers. For example, when a user reports that an answer is incomplete, the system learns from the feedback signal and optimizes itself by adding more reasoning nodes or more complex reasoning chains.

[0259] These feedback signals are stored in the system’s optimization module and used to train the model to improve the accuracy of signal processing. For example, when the system recognizes that the user has given negative feedback many times, it will retrain the prompt generation model to improve future answer signal generation.

[0260] System advantages brought by signal processing

[0261] 1. Enhanced logical reasoning: Through the generation of mind map prompts, the signal processing process can significantly enhance the system's logical reasoning ability. Prompt signals help the pre-trained language model generate more logical answers through multi-level reasoning chains.

[0262] 2. Domain-specific problem optimization: Adaptation to specific domains during signal processing enables the system to better respond to questions in different domains. Through domain-specific knowledge signals, the system can generate more professional answers, improving the wide adaptability of common sense question answering.

[0263] 3. Improved explainability and transparency: By recording and visualizing the reasoning signal chain, the system significantly improves the transparency of the reasoning process. Users can intuitively see how the system generates answers step by step, and this explainability enhances users' trust in the system.

[0264] 4. Signal self-optimization of feedback mechanism: Feedback signals enable the system to optimize its own performance in continuous application. By capturing feedback signals, the system automatically adjusts the reasoning graph generation and prompt signal strategies, achieving self-learning and continuous optimization, and further improving the accuracy and logic of the answers.

[0265] Through detailed signal processing, the large-model mind map prompt learning system for common sense question answering can achieve efficient and accurate logical reasoning and domain adaptation from inputting questions to generating complex reasoning answers. The multi-level architecture of signal processing not only enhances the reasoning ability of the system, but also achieves self-optimization and interpretability through feedback signals, ensuring that the system can cope with complex common sense question answering challenges.

[0266] Implementation case: Answering a common sense question based on "Why do plants need sunlight?"

[0267] Problem background:

[0268] The user asked a common sense question: "Why do plants need sunlight?" This question involves the fields of botany and basic biology, and requires the system to answer common sense related to plant photosynthesis. Existing common sense question-answering systems directly give superficial answers, such as "Plants produce energy through photosynthesis", but lack the support of detailed reasoning process and domain knowledge background.

[0269] System Implementation Process

[0270] 1. Knowledge graph generation module

[0271] Workflow:

[0272] The system extracts knowledge points related to plants, sunlight, and photosynthesis from the knowledge base. For example:

[0273] “Plants photosynthesize”

[0274] “Sunlight is the source of energy for photosynthesis”

[0275] “Plants need energy to carry out their metabolic activities”

[0276] The extraction of knowledge points is not limited to plants and sunlight, but also includes the details of photosynthesis, such as how photosynthesis decomposes carbon dioxide and water to produce oxygen and glucose.

[0277] The extracted knowledge is organized in the form of a graph. The nodes in the graph represent concepts or entities, such as "sunlight", "plants", "energy", and "photosynthesis", and the edges represent the associations between these concepts, such as "sunlight is the source of energy for photosynthesis".

[0278] Generated knowledge graph (simplified):

[0279] Nodes: Plants, Sunlight, Photosynthesis, Energy, Carbon Dioxide, Water, Oxygen, Glucose

[0280] Side: Sunlight -> Photosynthesis (energy source), Photosynthesis -> Plants (energy supply), Carbon dioxide + water -> Glucose + Oxygen (products of photosynthesis)

[0281] 2. Mind map reasoning module

[0282] Workflow:

[0283] The system analyzes the question raised by the user, “Why do plants need sunlight?” and extracts relevant reasoning paths from the generated knowledge graph.

[0284] Based on the information in the knowledge graph, the system generates a multi-level mind map that shows why plants need sunlight and the logical chain behind it.

[0285] For example, the first layer of the mind map shows: "Plants need sunlight for photosynthesis"; the second layer shows: "Photosynthesis produces glucose and oxygen by breaking down carbon dioxide and water"; the third layer shows: "Plants use glucose as an energy source for metabolic activities."

[0286] Generated mind map:

[0287] Layer 1: Sunlight -> Photosynthesis (Sunlight is the energy source for photosynthesis)

[0288] Layer 2: Photosynthesis -> Breaks down carbon dioxide and water (producing glucose and oxygen)

[0289] Layer 3: Glucose -> Plant Energy Source (provides energy for plant growth and metabolism)

[0290] 3. Prompt learning module

[0291] Workflow:

[0292] The system converts the generated mind map into prompt sentences and inputs them into a pre-trained language model (such as GPT-4) to guide the model to answer.

[0293] Example prompt: The question is, "Why do plants need sunlight?" Related prompts include:

[0294] 1. "Plants use sunlight as a source of energy for growth and metabolic activities through photosynthesis."

[0295] 2. "During photosynthesis, plants break down carbon dioxide and water into glucose and oxygen."

[0296] 3. “Glucose provides energy for plants and supports their basic metabolic activities.”

[0297] These hints ensure that when answering questions, the model does not just rely on superficial associations, but can instead gradually arrive at accurate answers through chains of reasoning.

[0298] 4. Domain Adaptation Module

[0299] Workflow:

[0300] This question involves the field of botany, and the domain adaptation module calls a special botanical knowledge base to further optimize the mind map. At this time, the system not only provides prompts from the perspective of basic biology, but also supplements knowledge related to plant physiology.

[0301] The domain adaptation module adds specific steps about photosynthesis to the mind map, such as: the two stages of photosynthesis (light reaction and dark reaction), how to decompose water molecules in the light reaction and produce ATP and NADPH, etc.

[0302] Optimized tips:

[0303] 1. "Photosynthesis is divided into light reaction and dark reaction stages. Sunlight provides energy to decompose water molecules, release oxygen, and produce ATP and NADPH for use in the dark reaction."

[0304] 2. “The dark reaction uses ATP and NADPH to convert carbon dioxide into glucose, which is the plant’s energy source.”

[0305] 5. Reasoning Explainability Module

[0306] Workflow:

[0307] The system records the entire reasoning chain and generates a visual mind map, allowing users to view the system's reasoning process when generating answers.

[0308] The Reasoning Explainability module shows the complete reasoning path from "sunlight" to "photosynthesis" to "plants obtain energy."

[0309] Through the visual interface, users can step by step view the detailed logic of how the system deduces the answer "sunlight provides energy for plants through photosynthesis" from the question "plants need sunlight".

[0310] Inference chain visualization:

[0311] Users can click on each node (such as "photosynthesis", "energy source", etc.) to view the relevant reasoning path and specific knowledge explanation.

[0312] This not only allows users to understand how the system reaches its conclusions, but also enhances the transparency and user trust of the system.

[0313] 6. Feedback and self-optimization

[0314] Workflow:

[0315] Users evaluate the generated answers. If the user believes that the system's answer is accurate, the system will record the feedback and mark it as a success case; if the user believes that the answer is not accurate or complete, the system will adjust the prompt strategy or update the knowledge base.

[0316] The system continuously optimizes mind map generation and prompts learning strategies through a feedback mechanism, gradually improving the quality of answers to such questions.

[0317] For example, after users reported that the system lacked a description of the specific processes of "light reaction" and "dark reaction", the system would add relevant steps in the next answer and improve the detailed description of plant physiological processes.

[0318] Implementation results and significant technological progress

[0319] 1. Multi-level reasoning and logic enhancement:

[0320] Guided by the mind map, the system not only gave the basic answer "Plants need sunlight for photosynthesis", but also explained the specific process of photosynthesis and its relationship with energy metabolism through multi-level reasoning. This kind of logical reasoning chain is difficult to achieve in traditional question-answering systems, but this system shows a clear reasoning path through the mind map, which greatly enhances the logic of the answer.

[0321] 2. Enhanced field adaptability:

[0322] The domain adaptation module enables the system to generate customized mind maps and prompts for questions in the field of botany, and optimizes photosynthesis details, plant physiology, etc. This significantly improves the system's performance in dealing with professional field problems, making the generated answers more professional and accurate.

[0323] 3. Enhanced interpretability:

[0324] The reasoning explainability module shows the complete reasoning chain from question parsing to answer generation, and users can intuitively view the basis of each step of the system's reasoning. This not only enhances users' trust in the system, but also makes the system's answering process transparent, making it easier to debug and optimize.

[0325] 4. Feedback mechanism and self-optimization:

[0326] The system can continuously optimize itself based on user feedback, gradually improving its ability to answer complex common sense questions. For example, through the feedback mechanism, the system will gradually enhance its description of the complex process of photosynthesis, so that it can be more accurate and detailed in subsequent answers.

[0327] Summarize:

[0328] This implementation case demonstrates the working principle of the large-scale mind map prompt learning system for common sense question answering in practical applications. Through the collaborative work of modules such as knowledge graph, mind map reasoning, prompt learning and domain adaptation, the system can accurately answer complex common sense questions and provide explainability of multi-level reasoning processes, greatly improving the reasoning ability, domain adaptability and user trust of the question answering system.

[0329] Implementation plan: A large-model mind map prompt learning system for common sense question answering

[0330] The system combines large-scale pre-trained language models (such as GPT), knowledge graphs, prompt learning technology, and domain adaptation modules, aiming to implement the reasoning process of common sense question answering through algorithms and programs. The implementation of the system includes multiple core modules, such as knowledge graph generation, mind map reasoning, prompt learning, domain adaptation, and reasoning explainability. The specific implementation method of each module is described in detail below, including the algorithm used and the program implementation process.

[0331] 1. Knowledge graph generation module

[0332] Extract common-sense entities and reasoning relationships between entities from large-scale unstructured text to generate a domain-wide knowledge graph.

[0333] #Core algorithm:

[0334] Named Entity Recognition (NER): used to identify key entities in text (such as names, places, events, etc.). A pre-trained NER model based on BERT is used.

[0335] Relation Extraction (RE): used to identify the relationship between entities. A relation extraction model based on BERT or RoBERTa is used.

[0336] Graph Embedding: Map the nodes and relationships in the knowledge graph into a high-dimensional vector space using a graph convolutional network (GCN) or TransE model.

[0337] #Implementation steps:

[0338] 1. Data preprocessing: Convert large-scale corpus data into a standard text format and segment it into sentences. Use Python's spaCy library for preprocessing, including word segmentation and stop word removal.

[0339] 2. Entity extraction: Use the pre-trained BERT NER model to call the `transformers` library to annotate entities in unstructured text. Example code:

[0340] ```Python

[0341] from transformers import BertTokenizer,BertForTokenClassification

[0342] tokenizer=BertTokenizer.from_pretrained("bert-base-uncased")

[0343] model=BertForTokenClassification.from_pretrained("bert-base-uncased",num_labels=9)

[0344] inputs=tokenizer(text,return_tensors="pt")

[0345] outputs = model(inputs)

[0346] predictions=outputs.logits.argmax(dim=2)

[0347] ```

[0348] 3. Relation extraction: Detect semantic relationships between entities through the RoBERTa model. Use the `transformers` library to implement:

[0349] ```Python

[0350] from transformers import RobertaTokenizer,RobertaForSequenceClassification

[0351] tokenizer=RobertaTokenizer.from_pretrained('roberta-base')

[0352] model=RobertaForSequenceClassification.from_pretrained('roberta-base',num_labels=2)

[0353] inputs=tokenizer(text,return_tensors='pt')

[0354] outputs = model(inputs)

[0355] relations=outputs.logits.argmax(dim=1)

[0356] ```

[0357] 4. Knowledge graph construction: Build entities and their relationships into a knowledge graph, and use the `networkx` library to create the graph structure.

[0358] ```Python

[0359] import networkx as nx

[0360] G = nx.Graph()

[0361] G.add_node(entity_1)

[0362] G.add_node(entity_2)

[0363] G.add_edge(entity_1,entity_2,relation='is_related_to')

[0364] 2. Mind map reasoning module

[0365] Generate problem-related mind maps based on the knowledge graph and display the reasoning path in a hierarchical manner.

[0366] #Core algorithm:

[0367] Graph search algorithm: Use depth-first search (DFS) or breadth-first search (BFS) to find reasoning paths related to the question in the knowledge graph.

[0368] Hierarchical display: Use visualization tools such as `graphviz` to display the reasoning path in the form of a hierarchical mind map.

[0369] #Implementation steps:

[0370] 1. Graph search: Based on the questions input by the user, use the BFS or DFS algorithm to search for nodes and edges in the knowledge graph to find relevant reasoning paths.

[0371] ```Python

[0372] defbfs_search(graph,start_node,target_node):

[0373] visited=set()

[0374] queue=[(start_node,[start_node])]

[0375] while queue:

[0376] (node,path)=queue.pop(0)

[0377] ifnode not in visited:

[0378] ifnode == target_node:

[0379] returnpath

[0380] visited.add(node)

[0381] queue.extend((neighbor,path+[neighbor])for neighbor in graph[node])

[0382] 2. Mind map construction: Generate a hierarchical mind map from the searched path nodes and display it through `graphviz`.

[0383] ```Python

[0384] import graphviz

[0385] defvisualize_path(path):

[0386] dot=graphviz.Digraph(comment="Thinking Path")

[0387] for i,node in enumerate(path):

[0388] dot.node(str(i),node)

[0389] ifi>0:

[0390] dot.edge(str(i-1),str(i))

[0391] dot.render("output / path.gv",view=True)

[0392] 3. Prompt learning module

[0393] Combined with knowledge graph information, prompt tuning is used to guide the pre-trained language model (such as GPT) to generate inferential answers.

[0394] #Core algorithm:

[0395] Prompt Tuning: Convert node information in the mind map into prompt input and guide the pre-trained model.

[0396] Language model reasoning: Use the GPT model to generate answers after prompts through the `transformers` library.

[0397] #Implementation steps:

[0398] 1. Prompt design: According to the reasoning path of the mind map, construct the prompt text and input it into the pre-trained model. For example:

[0399] ```Python

[0400] prompt="Given the following information:{thinking_path},answerthequestion:{question}"

[0401] input_prompt=prompt.format(thinking_path="->".join(path), question=user_question)

[0402] ```

[0403] 2. Model generates answers:

[0404] ```Python

[0405] from transformers import GPT2Tokenizer,GPT2LMHeadModel

[0406] tokenizer=GPT2Tokenizer.from_pretrained('gpt2')

[0407] model=GPT2LMHeadModel.from_pretrained('gpt2')

[0408] inputs=tokenizer(input_prompt,return_tensors="pt")

[0409] outputs=model.generate(inputs["input_ids"],max_length=100)

[0410] answer=tokenizer.decode(outputs[0],skip_special_tokens=True)

[0411] 4. Domain Adaptation Module

[0412] Generate customized mind maps for specific fields and optimize the prompt guidance model.

[0413] #Core algorithm:

[0414] Domain adaptation: Use domain-specific pre-trained models (such as medicine, law, etc.) combined with knowledge graphs to generate customized mind maps for domain problems.

[0415] Prompt optimization: Adjust prompt templates by domain relevance to ensure the model understands the domain context.

[0416] #Implementation steps:

[0417] 1. Domain-specific models: Use pre-trained models for different fields. For example, in the medical field, you can use `BioBERT`:

[0418] ```Python

[0419] from transformers import BertTokenizer,BertForSequenceClassification

[0420] tokenizer=BertTokenizer.from_pretrained("biobert")

[0421] model=BertForSequenceClassification.from_pretrained("biobert")

[0422] ```

[0423] 5. Reasoning Explainability Module

[0424] Record and display the whole process of mind map reasoning, and provide the reasoning chain associated with the final answer.

[0425] #Core algorithm:

[0426] Shapley value calculation: used to show the contribution of the inference node.

[0427] Visualization tools: Use tools such as `matplotlib` or `seaborn` to visualize the inference path.

[0428] #Implementation steps:

[0429] 1. Reasoning chain record: Each reasoning step is recorded in a chain structure:

[0430] ```Python

[0431] reasoning_chain = []

[0432] for step in path:

[0433] reasoning_chain.append(step)

[0434] ```

[0435] 2. Contribution calculation and visualization:

[0436] ```Python

[0437] import shap

[0438] shap_values=shap.TreeExplainer(model).shap_values(inputs)

[0439] shap.summary_plot(shap_values,inputs)

[0440] Through knowledge graph generation, prompt learning, mind map reasoning and reasoning explainability, combined with deep learning algorithms such as LSTM and BERT, a reasoning system for common sense question answering was designed and implemented.

[0441] It should be noted that the above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by any technician familiar with the technical field within the technical scope disclosed by the present invention and within the spirit and principles of the present invention should be covered within the protection scope of the present invention.

Claims

1. A large-model mind map prompt learning system for common sense question answering, characterized in that: include: The knowledge graph generation module, mind map reasoning module, prompt learning module, domain adaptation module, reasoning explainability module, and algorithm components involving data processing and model training are as follows: A knowledge graph generation module is connected to the mind map reasoning module and the reasoning explainability module. The knowledge graph generation module is used to automatically extract knowledge from a large-scale corpus through a natural language processing (NLP) algorithm to construct a common sense knowledge graph in multiple fields. The knowledge graph contains multiple nodes, each node represents an entity, and the edges between the nodes represent the common sense reasoning relationship between the entities. The extracted entities and relationships are represented and learned using a graph neural network (GNN) or other graph structure learning algorithms to form a vectorized knowledge graph that can be used for reasoning. The mind map reasoning module is connected with the knowledge map generation module, the prompt learning module, the domain adaptation module, and the reasoning interpretability module. The mind map reasoning module is used to automatically generate a dynamically updated mind map based on the questions raised by the user by using a reinforcement learning algorithm or a reasoning model based on deep learning (such as a neural symbolic reasoning model). The mind map displays multi-level information related to the question in a hierarchical and logical manner. The reasoning path is optimized by combining a Bayesian network or a path planning algorithm to ensure the rationality and accuracy of the reasoning path; A prompt learning module is connected to the mind map reasoning module and the domain adaptation module. The prompt learning module interacts with the pre-trained language model and combines the prompt learning technology to input each node information and relationship chain information into the pre-trained language model in the form of prompts based on the generated mind map. The model performs reasoning and answering according to the prompts. The prompts can be adaptively adjusted through training to improve the accuracy of the answers. The specific prompt learning strategies include soft prompts and reinforcement prompts. The domain adaptation module is connected with the mind map reasoning module, the prompt learning module, and the reasoning interpretability module. The domain adaptation module realizes the adaptive training of the system for a specific domain through knowledge distillation or domain transfer learning technology. When answering questions, it generates a mind map for a specific domain through a deep domain adaptation algorithm based on a pre-stored domain knowledge base or domain expert data, and optimizes prompts in combination with a domain-specific model during the reasoning process. The reasoning explainability module is connected with the knowledge graph generation module, the mind map reasoning module, and the domain adaptation module. The reasoning explainability module records and tracks the node reasoning process in the mind map through heuristic rules or logistic regression models to form a reasoning chain that can be reviewed by users. At the same time, it combines visualization algorithms to generate a visualized mind map to display the reasoning path, thereby enhancing the transparency and explainability of the system. Data module: The system relies on large-scale multi-domain datasets for training and fine-tuning various models involved in the system. The data includes public common sense corpora, domain expert data, and data generated by user interactions. The system optimizes model performance by combining large-scale unsupervised pre-training and supervised fine-tuning, and improves the generalization ability of the inference model through data augmentation technology. Algorithm module, the algorithms used by the system include but are not limited to graph neural network (GNN) for the generation and update of knowledge graphs, deep reinforcement learning (DRL) for the optimization of mind map reasoning paths, prompt learning algorithm for guiding the pre-trained language model to generate reasoning answers, transfer learning algorithm for domain adaptation, and logistic regression, neural symbolic reasoning and visualization algorithms for the interpretability of the reasoning process.

2. The large-model mind map prompt learning system for common sense question answering according to claim 1 is characterized in that: The knowledge graph generation module further includes a named entity recognition (NER) algorithm and a relationship extraction (RE) algorithm based on deep learning, which are used to automatically extract entities and relationships between entities from unstructured text. The entities and relationships are mapped to a vector space, and a high-dimensional vector representation that can be used for reasoning is generated through a graph embedding algorithm (Graph Embedding), thereby enhancing the accuracy and generalization ability of reasoning.

3. The large-model mind map prompt learning system for common sense question answering according to claim 1 is characterized in that: The prompt learning technology adopted by the prompt learning module includes a self-supervised learning mechanism. The self-supervised mechanism automatically adjusts the structure and content of the prompt input by generating pseudo-label data for the model, and updates the prompt strategy using the reward signal of reinforcement learning, thereby improving the performance of the pre-trained model in complex common sense reasoning tasks.

4. The large-model mind map prompt learning system for common sense question answering according to claim 1 is characterized in that: The reasoning explainability module further combines a visualized heat map to display the contribution of different reasoning nodes, which is calculated through Shapley value or attention mechanism, so that users can intuitively understand the key decision nodes and information sources in the reasoning process.

5. The large-model mind map prompt learning system for common sense question answering as claimed in claim 1, characterized in that: The domain adaptation module method is as follows: Used to handle problems in specific fields, such as medicine, law, and finance. These fields have unique common sense and reasoning logic, so the system will generate customized mind maps for these problems and provide optimized prompts through the domain knowledge base; The system calls the corresponding domain knowledge base according to the domain attributes of the problem to ensure that domain-specific knowledge nodes and reasoning relationships are included when generating the mind map. For example, when the system processes a question like "Why does the heart beat?", it will first extract information related to the heart and physiological operations from the knowledge base in the medical field to generate a mind map specifically for the medical field. When generating prompts, the system also adjusts the prompt strategy to better suit the reasoning needs of the field, such as guiding the language model to use specific medical terms or logical reasoning patterns; Detect the domain to which the problem belongs and call the domain knowledge base to generate a customized mind map; Include domain-related nodes and relationship chains in the mind map to ensure that learning prompts focus on domain common sense; Generate optimization hints that enable pre-trained models to provide accurate answers in specific domains.

6. The large-model mind map prompt learning system for common sense question answering as claimed in claim 1, characterized in that: The reasoning interpretability module method is as follows: By recording and visualizing the reasoning process of the mind map, an explanation of the answers generated by the system is provided; this module not only shows how the model starts from the question and gets the answer through multi-level reasoning, but also shows the logical relationship behind each step of reasoning; The reasoning path of each question will be recorded, and users can view the reasoning chain based on which the system generates answers through a visual interface; when users ask "Why do plants need sunlight?", the system will display a related mind map to explain the principle of plant photosynthesis and how to derive "photosynthesis" and "energy acquisition" information from "sunlight"; Record the key nodes and reasoning chains in each model reasoning process; Visualize the reasoning chain in the mind map and show it to the user; Provide explanations of the reasoning process to help users understand the model's answer logic.

7. The large-model mind map prompt learning system for common sense question answering as claimed in claim 1, characterized in that: Further including: In the knowledge graph generation module, advanced natural language processing (NLP) algorithms, such as BERT variants or Transformer models, are used to efficiently extract entities and relationships from large-scale multi-domain corpora and build graphs containing rich common sense knowledge. Using variants of graph neural networks (GNNs), such as Graph Convolutional Networks (GCN) or Graph Attention Networks (GAT), representation learning is performed on the constructed knowledge graph to convert entities and relations into vector form for use in subsequent reasoning; Implement a dynamic knowledge graph update mechanism to maintain the timeliness and comprehensiveness of the knowledge graph by continuously learning new common sense information, ensuring that the system can handle the latest common sense question and answer needs.

8. The large-model mind map prompt learning system for common sense question answering as claimed in claim 1, characterized in that: Further including: In the mind map reasoning module, the deep reinforcement learning (DRL) algorithm is combined to optimize the mind map generation process to ensure that the reasoning path is both efficient and accurate; The prompt learning module uses advanced soft prompt technology, such as continuous prompt vectors and reinforcement prompting strategies, to iteratively optimize prompt content and improve the pre-trained language model's ability to understand and utilize mind map information. Realize the dynamic adjustment function of the mind map, automatically adjust the structure and content of the mind map according to user feedback and reasoning results, so as to improve the accuracy and adaptability of the system in answering future problems.

9. The large-model mind map prompt learning system for common sense question answering as claimed in claim 1, characterized in that: Further including: In the domain adaptation module, knowledge distillation technology is used to achieve efficient model deployment in specific domains by having small models imitate the behavior of large models. At the same time, domain transfer learning algorithms, such as Transfer Learning with Adapters, are used to quickly adapt the system to new domains. The reasoning explainability module uses a fine-grained heuristic rule set and a logistic regression model to record the reasoning basis and path selection of each node in the mind map in detail and generate a detailed reasoning chain report; Combined with advanced visualization technologies, such as D3.js or Plotly, a user-friendly interface is developed to intuitively display the reasoning path and node relationships of the mind map, enhancing users' trust and understanding of the system reasoning process.

10. A method for learning common sense question and answering with a large model mind map prompt, which implements the large model mind map prompt learning system for common sense question and answering as claimed in any one of claims 1 to 6, characterized in that: The large model mind map prompt learning method for common sense question answering includes: Step 1: Extract knowledge from large-scale corpus through the knowledge graph generation module to build a common sense knowledge graph in multiple fields; the graph contains entities and facts, and common sense reasoning relationships between entities; Step 2: Based on the question raised by the user, the mind map reasoning module generates a mind map using the knowledge map. The map displays multi-level information and reasoning paths related to the question in a hierarchical and logical way. Step 3: Based on the pre-trained language model, the prompt learning module combines the prompt learning technology to input the node information and relationship chain information in the mind map into the language model through prompts to guide the model to give a more reasoning answer; Step 4: For questions in a specific field, the system generates a customized mind map from the corresponding domain knowledge base through the domain adaptation module, and optimizes the prompts through the domain adaptation module to ensure that the model can accurately answer domain-specific common sense questions; Step 5, record and visualize the reasoning process of each node in the mind map through the reasoning explainability module, and provide the reasoning chain associated with the answer generated by the model for users to review and reference, thereby enhancing the transparency and explainability of the system.

Images