Intelligent questioning and answering method and intelligent agent for large-model knowledge graph based on intention driving

Abstract

The invention belongs to the field of artificial intelligence and knowledge graph application, and relates to an intent-driven large model knowledge graph intelligent question and answer method and an intelligent agent. The method comprises the steps of receiving a question input by a user, and identifying a potential intention of the user; according to the identified intention and the information stored in the shared memory, selecting a proper tool from a tool only based on a large language model, a tool based on semantic analysis, a tool based on sub-graph retrieval and a general tool for reasoning; new information generated by the tool is updated into the shared memory; and in each iteration, the large language model evaluates the current reasoning state according to the information stored in the shared memory, and triggers a tool with a corresponding intention to perform supplementary reasoning according to the current reasoning state until enough information is collected to draw a conclusion or a preset upper limit of the number of iterations is reached. The method can improve the reasoning efficiency and accuracy, can be applied to complex question answering in various fields, and provides high-quality question answering service for users.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and knowledge graph applications, and relates to an intelligent question answering technology based on knowledge graphs and large language models. Specifically, it is a method and system for achieving adaptive workflow reasoning through intent-driven knowledge graph intelligent agents to answer complex questions. Background Technology

[0002] Recent advances in Large Language Models (LLMs) have yielded significant performance improvements across numerous Natural Language Processing (NLP) tasks, including text generation, translation, and summarization. However, their performance remains relatively lagging in reasoning tasks, particularly in domains requiring structured knowledge retrieval and logical reasoning.

[0003] Knowledge graphs (KGs) serve as a valuable resource, supporting reasoning based on large language models by providing structured and compact knowledge representations. Traditional KGQA methods primarily employ two paradigms: semantic parsing (SP) and subgraph retrieval (SR). Semantic parsing transforms natural language questions into formal queries that can be executed against the knowledge graph (e.g., SPARQL), ensuring accurate and interpretable answers when entity and relation mappings are accurate. Meanwhile, subgraph retrieval extracts relevant subgraphs from the knowledge graph and utilizes them for reasoning, enabling multi-hop reasoning through relational paths.

[0004] While semantic parsing-based and subgraph retrieval-based methods have proven useful, each has its inherent limitations. Semantic parsing methods are highly fragile because errors in entity, relation, or triple mappings can propagate across multiple stages of inference, significantly impacting overall performance. On the other hand, while subgraph retrieval-based methods are more robust to sparsity, they struggle with long or complex inference paths, where errors or unreliable reasoning frequently occur. Furthermore, many subgraph retrieval-based models heavily rely on information recalled from the graph, failing to fully leverage the parameterized knowledge embedded in Large Language Models (LLMs). This creates a gap in utilizing the rich pre-trained knowledge of LLMs, which has demonstrated strong performance in open-domain question answering tasks.

[0005] Existing methods combining LLMs with KGs, such as retrieval enhancement methods, collaborative enhancement methods, and hybrid methods, while achieving some results, still face challenges such as rigid interaction paradigms, limited generalization capabilities, and difficulties in handling multi-hop or complex reasoning tasks. Although LLM-based agents have made progress in KGQA, most rely on single toolkits and fixed workflows, making it difficult to adapt to complex and dynamic queries and environments. This places excessive demands on LLMs as the primary planner, easily leading to low task execution efficiency or even errors. For example, when encountering unexpected query scenarios, agents with fixed workflows cannot flexibly adjust their strategies, resulting in incorrect or inaccurate answers.

[0006] Therefore, there is an urgent need for a new technical solution that can effectively integrate the advantages of multiple reasoning strategies, overcome the shortcomings of existing methods, and improve performance and adaptability in solving complex problems. Summary of the Invention

[0007] To address the aforementioned problems, this invention provides an intention-driven intelligent question-answering method and agent based on a large model knowledge graph.

[0008] The technical solution adopted in this invention is as follows:

[0009] An intent-driven intelligent question answering method based on large-scale model knowledge graphs includes the following steps:

[0010] Receive user input questions and identify the user's potential intentions;

[0011] Based on the identified intent and the information stored in shared memory, a suitable tool is selected from tools based solely on large language models, tools based on semantic parsing, tools based on subgraph retrieval, and general tools for inference;

[0012] After reasoning using the selected tool, update the shared memory with the new information generated by the tool;

[0013] In each iteration, the large language model evaluates the current reasoning state based on the information stored in shared memory, and triggers tools with corresponding intentions to perform supplementary reasoning based on the current reasoning state, until enough information is collected to draw a conclusion or the preset iteration limit is reached.

[0014] Furthermore, the identification of the user's potential intent involves analyzing the problem using predefined intent templates and a large language model to identify the user's potential intent.

[0015] Furthermore, the information stored in the shared memory includes the current inference progress, identified entities and relationships, and the relevant state of the knowledge graph; the new information generated by the tool includes newly identified entities, relationships, and calculated values.

[0016] Furthermore, among the tools based solely on large language models, one type focuses on generating hypotheses to retrieve relevant entities, while the other type ensures accuracy by standardizing key information to guarantee that the output does not have inconsistent formats.

[0017] Furthermore, the semantic parsing-based tool follows the ChatKBQA method. Natural language problems are first converted into logical forms using the semantic understanding of a large language model. Coreference resolution and entity relation disambiguation are then applied to refine these logical forms. The refined logical forms are then converted into executable SPARQL queries to retrieve relevant answers from the knowledge graph. The subgraph-based retrieval tool generates relational paths from user queries and constructs accurate answers using inference paths. Combined with graph neural network patterns, it enhances reasoning capabilities by retrieving subgraphs while retrieving inference paths, thereby improving performance and scalability in complex tasks.

[0018] Furthermore, the general tools include tools for extracting key information from the knowledge graph, as well as basic logical operations, and also semantic tools; the semantic tools enhance reasoning capabilities by enabling relation retrieval and entity disambiguation, and extend the ability of knowledge graph interaction by utilizing pre-trained models.

[0019] Furthermore, the step of selecting a suitable tool for inference from tools based solely on large language models, semantic parsing-based tools, subgraph retrieval-based tools, and general tools based on the identified intent and information stored in shared memory includes:

[0020] If the intent is identified as requiring precise structured knowledge retrieval, and some relevant entity information already exists in the current shared memory, then a semantic parsing-based tool is selected to generate a logical form, which is then converted into a SPARQL query and the answer is retrieved from the knowledge graph.

[0021] If more potential relationship paths need to be explored during the reasoning process, then choose a tool based on subgraph retrieval to generate relationship paths;

[0022] When the semantic parsing-based tool reaches two or more iterations, the subgraph retrieval-based tool is activated to leverage the advantages of graph neural networks in multi-hop reasoning to improve reasoning performance.

[0023] An intent-driven, large-scale knowledge graph question-answering agent includes:

[0024] The intent recognition module is used to receive user input questions and identify the user's potential intent;

[0025] The tool selection module is used to select appropriate tools for inference from tools based solely on large language models, semantic parsing-based tools, subgraph retrieval-based tools, and general tools, based on the identified intent and information stored in shared memory.

[0026] The shared memory update module is used to update the shared memory with new information generated by the selected tool after inference is performed.

[0027] The iterative reasoning loop module is used to evaluate the current reasoning state of the large language model based on the information stored in the shared memory in each iteration, and to trigger the corresponding intent tool to perform supplementary reasoning based on the current reasoning state, until enough information is collected to draw a conclusion or the preset iteration limit is reached.

[0028] The beneficial effects of this invention are as follows:

[0029] 1. Improved Reasoning Efficiency and Accuracy: The intent-driven approach quickly identifies appropriate reasoning strategies and tools, avoiding blind trial and error and reducing unnecessary computational overhead. Simultaneously, the multi-strategy fusion approach fully leverages the advantages of LLMs, SP-based, and SR-based methods, significantly improving reasoning accuracy when handling multi-hop reasoning and complex problems. For example, when dealing with complex problems involving multiple entities and relationships, combining the precision of SP with the robustness of SR leads to a more accurate solution.

[0030] 2. Enhanced Adaptability and Flexibility: The agent can dynamically adjust its inference workflow based on the complexity of the problem and the state of the knowledge graph, adapting to different types of complex problems and diverse user needs. The shared memory mechanism enables efficient collaboration between workflows with different intentions. When encountering inference difficulties or new information, the agent can promptly switch strategies, maintaining the smoothness and effectiveness of the inference process. For example, when encountering new knowledge graph structures or updated knowledge, the agent can quickly adjust its tool selection and inference methods to adapt to changes.

[0031] 3. Enhanced System Generalization Ability: The method and system of this invention do not rely on specific knowledge graphs or large language models, exhibiting good versatility. Through training and optimization on different datasets, they can be applied to answering complex questions in various fields, such as intelligent customer service, intelligent assistants, and intelligent education, providing users with high-quality question-and-answer services. For example, in intelligent customer service scenarios, they can accurately answer various complex business questions raised by users, improving user experience. Attached Figure Description

[0032] Figure 1 This is a flowchart of the processing method of the present invention.

[0033] Figure 2This is a schematic diagram of the processing procedure of the method of the present invention. Detailed Implementation

[0034] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to specific embodiments and accompanying drawings.

[0035] To address the limitations of existing technologies, this invention proposes an intent-driven large-scale graph question-answering agent, IKG-Agent. It unifies semantic parsing-based, subgraph retrieval-based, and large-scale language model-based reasoning methods into a coherent system, aiming to fully leverage their respective advantages while mitigating their disadvantages. Unlike semantic parsing-based methods, which are prone to error propagation, or subgraph retrieval-based methods, which struggle with long and complex reasoning paths and fail to fully utilize the parameterized knowledge embedded in the large-scale language model, IKG-Agent introduces an innovative intent-driven architecture to dynamically adjust its reasoning strategy.

[0036] The processing flow of the method of the present invention is as follows: Figure 1 As shown, the processing procedure is as follows: Figure 2 As shown, the process first identifies the user's intent and then constructs a customized inference workflow using a large language model. During execution, the agent interacts with a shared memory mechanism to reduce the cascading effect of accumulated errors in semantic parsing-based or subgraph retrieval-based inference. Furthermore, when the workflow encounters extended inference loops or errors from long inference chains, the agent can dynamically switch to a new workflow driven by different inference intents. These workflows exchange and integrate information between different inference paradigms through shared memory, leveraging the accuracy of semantic parsing-based methods, the robustness of subgraph retrieval-based methods, and the rich knowledge encoded in the large language model.

[0037] In summary, the main contributions of this invention are as follows:

[0038] 1) This invention proposes IKG-Agent, a unified intelligent agent framework that integrates multiple reasoning strategies based on large language models, semantic parsing, and subgraph retrieval into a single coherent system to enhance knowledge graph-based question answering (KGQA).

[0039] 2) The method of the present invention utilizes an intent-driven reasoning loop to select the most appropriate tools and strategies based on the complexity of the task and the available knowledge obtained from the knowledge graph.

[0040] 3) This invention introduces a shared memory mechanism, which enables multiple inference workflows to collaborate efficiently and enhances the agent's ability to handle complex multi-hop problems.

[0041] The IKG-Agent of this invention is a framework designed for autonomous multi-step reasoning on a knowledge graph (KG). The IKG-Agent consists of four core components: an intent recognition module, a tool selection module, a shared memory update module, and an iterative reasoning loop module. This invention also provides a toolkit that allows the agent to autonomously call APIs.

[0042] 1. Intent Recognition Module

[0043] The system receives user input questions and analyzes them using predefined intent templates and a large language model to identify the user's potential intent. For example, when a user asks, "Which country is in the ASEAN time zone and has the largest population?", the intent recognition module can determine that the question involves geographical regions, population comparisons, and other intents, guiding the subsequent reasoning process.

[0044] 2. Tool Selection Module

[0045] Based on the identified intent, the module selects an appropriate tool from four categories: LLM-only (based solely on a large language model), SP-based (based on semantic parsing), SR-based (based on subgraph retrieval), and General (general tools). The tool selection process is based on information stored in shared memory (including current inference progress, identified entities and relationships, and the relevant state of the knowledge graph) and the current iteration count. For example, if the intent is identified as requiring precise structured knowledge retrieval, and some relevant entity information is already in the shared memory, the tool selection module might choose an SP-based tool, such as `generate_logic_form`, to generate a logical form, which is then converted into a SPARQL query to retrieve the answer from the knowledge graph. If the inference process requires exploring more potential relationship paths, the SR-based tool `generate_relation_path` might be chosen to generate relationship paths. This module also has the ability to switch modes based on specific conditions. For instance, when the initial intent based on semantic parsing (SP-based) reaches two or more iterations, `gnn_mode` (graph retrieval mode) is activated to leverage the advantages of graph neural networks in multi-hop inference to improve inference performance.

[0046] 3. Shared memory

[0047] After each tool execution, the shared memory update module updates the shared memory with new information generated by the tool (such as newly identified entities, relationships, and calculated values). This information comes not only from the reasoning process of the current intent but may also involve the reasoning results of other intents, ensuring that the shared memory reflects the latest knowledge state throughout the reasoning process. This provides a comprehensive and accurate basis for subsequent tool selection and reasoning decisions. For example, in one reasoning operation, if an SR-based tool retrieves new entity relationships, the shared memory update module will store this relationship information for subsequent SP-based tools to reference when generating logical forms.

[0048] 4. Iterative reasoning loop

[0049] Driven by a large language model, in each iteration, the large language model evaluates the current reasoning state based on information in shared memory. If a knowledge gap is found or more information is needed to advance the reasoning, the large language model triggers a corresponding intent workflow to supplement the reasoning. During this process, the agent does not rely on a single workflow but flexibly integrates workflows of different reasoning paradigms (such as SP-based and SR-based reasoning) according to the actual situation, continuously adjusting the reasoning direction and strategy. The iterative process continues until the agent has collected enough information to draw a conclusion or reaches a preset iteration limit. For example, when solving complex problems, an SP-based tool is initially used for preliminary retrieval. If insufficient information is found, an SR-based tool is triggered through iterative reasoning loops to further explore relationship paths, supplement information, and then the large language model re-evaluates and reasons.

[0050] Regarding the discovery of knowledge gaps, the large language model acts as the judge, using cue word engineering to determine whether a knowledge gap exists based on existing information.

[0051] 5. Toolbox

[0052] The tools provided by this invention for intelligent agents can be categorized into four types: those based solely on large language models, those based on semantic parsing, those based on subgraph retrieval, and general-purpose tools. The following is a detailed description of each type of tool:

[0053] 1) Tools based solely on large language models: This set of tools helps large language models leverage their internal knowledge while addressing common output errors. One type of tool focuses on generating hypotheses to retrieve relevant entities, while another ensures accuracy by standardizing key information (such as dates, numbers, and names) to prevent inconsistent output formats.

[0054] 2) Semantic Parsing-Based Tools: Following the ChatKBQA approach, natural language questions are first transformed into logical forms using semantic understanding from a large language model. Coreference resolution and entity relation disambiguation are then applied to refine these logical forms. Finally, the refined logical forms are converted into executable SPARQL queries to retrieve relevant answers from the knowledge graph.

[0055] 3) Subgraph-based retrieval tools: Inspired by RoG (Reasoning on Graphs), subgraph-based retrieval tools aim to generate relational paths from user queries and construct accurate answers using these inference paths. Based on pre-trained RoG weights, these tools can efficiently generate relational paths and perform inference. In advanced iterations, these tools incorporate a graph neural network mode (GNN-mode) to enhance inference capabilities by retrieving subgraphs simultaneously with inference paths, further improving performance and scalability in complex tasks. This design ensures that knowledge graph-based queries yield reliable and interpretable answers.

[0056] 4) General Tools: General tools provide a flexible foundation for interacting with knowledge graphs (KGs) and supporting reasoning tasks. These include tools for extracting key information from knowledge graphs (such as relations, entities, and specific data types), as well as basic logical operations (such as counting, set operations, and conditional validation). Semantic tools are also included, which enhance reasoning capabilities by enabling tasks such as relation retrieval and entity disambiguation, and extend the ability to interact with knowledge graphs using pre-trained models.

[0057] In one embodiment of the present invention, an intention-driven intelligent question answering method based on a large model knowledge graph is provided, comprising the following steps:

[0058] 1) Data preparation

[0059] Training and evaluation were performed using public datasets containing a large number of complex questions, such as WebQuestionSP and ComplexWebQuestions. The questions in these datasets are associated with the Freebase knowledge graph and cover multi-hop questions of varying complexity. Existing state-of-the-art methods (such as SOTASP-based ChatKBQA, SR-based ROG, and ChatGPT) were used to label the questions in the datasets, determining the appropriate reasoning method (LLM-only, SP-based, or SR-based) for each question. For example, for some simple questions, if all of the above methods can answer them correctly, they are labeled as LLM-only; if the SP-based method and ChatGPT can answer them correctly, they are labeled as SP-based, and so on. Based on the structured function calls generated by ChatGPT-4o, and combined with predefined tools (such as HYDE and format_fix in LLM-only; generate_logic_form, unsupervised_retrieval, and generate_sparql_query in SP-based; generate_relation_path, retrieve_entities, and gnn_mode in SR-based; and update_KG, explore_KG, answer_trying, restart, and launch_parallel_workflow in General), inference instructions are synthesized. These instructions are crucial for training large language models in knowledge graph inference tasks.

[0060] 2) Model Training

[0061] A suitable large language model (such as Llama3-8B) was selected as the base model for the system of this invention. During training, fine-tuning was performed using the training sets of the WebQuestionSP and Complex WebQuestions datasets for the two key tasks of intent recognition and tool selection. Training was conducted on eight A100 GPUs for a total of three epochs. By applying contextual learning in the two tasks, effective knowledge transfer between intent recognition and tool selection was facilitated, improving the model's performance in complex problem reasoning.

[0062] 3) Question and Answer Process

[0063] After a user inputs a question, the intent recognition module first analyzes the question to determine the user's intent. For example, for the question "Please find the major cities of Asian countries with a population exceeding 100 million," the intent recognition module identifies intents involving geographical region (Asia), population size criteria (over 100 million), and city queries. The tool selection module selects an appropriate tool based on the intent and shared memory information. Assuming that the shared memory already contains some information about Asian countries, the tool selection module might first select the SP-based tool `generate_logic_form` to generate a logical form, and then convert it into a SPARQL query to retrieve countries matching the population criteria from the knowledge graph. If, during the retrieval process, it is found that further exploration of the relationship path between countries and cities is needed, the tool selection module selects the SR-based tool `generate_relation_path` in subsequent iterations to generate the relationship path and continue mining relevant information. The shared memory update module continuously updates newly discovered information (such as matching countries, relationships between countries and cities, etc.) into the shared memory during tool execution. For example, when the SR-based tool retrieves a new country-city relationship, the shared memory update module stores this information. The iterative reasoning loop continues, with the large language model constantly evaluating the reasoning state based on information in shared memory until enough information is collected to generate an answer. For example, after multiple iterations, once information on the major cities of all eligible countries has been collected, the large language model combines this information with the original question to generate an accurate and fluent answer, such as "Countries in Asia with a population exceeding 100 million include China and India. Major cities in China include Beijing and Shanghai, while major cities in India include Mumbai and New Delhi."

[0064] 4) Experimental verification

[0065] The system of this invention was evaluated using the WebQuestionSP and Complex WebQuestions datasets, and compared with 20 baseline methods (including embedding-based methods, retrieval-augmented methods, semantic parsing methods, LLMs+KGs methods, and LLM-based agent methods). Evaluation metrics included exact match accuracy (Hits@1) and F1 score. Experimental results show that the system of this invention achieves state-of-the-art (SOTA) or second-best performance on most metrics, significantly outperforming single-policy methods (such as semantic parsing and information retrieval) and many existing agent-based methods, validating the effectiveness of the method and system of this invention.

[0066] The system of this invention has good scalability and can access knowledge graphs from different fields, such as those in the medical field and the financial field. By adjusting the training data and parameters of the large language model and adapting it to the structure and content of different knowledge graphs, it can achieve adaptive solutions to complex questions in multiple fields, providing intelligent question-answering services for users in different fields. For example, in the medical field, it can answer complex questions about disease diagnosis and treatment plans; in the financial field, it can answer complex questions about investment strategies and market analysis.

[0067] Another embodiment of the present invention provides an intent-driven large-scale model knowledge graph question-answering agent, which includes:

[0068] The intent recognition module is used to receive user input questions and identify the user's potential intent;

[0069] The tool selection module is used to select appropriate tools for inference from tools based solely on large language models, semantic parsing-based tools, subgraph retrieval-based tools, and general tools, based on the identified intent and information stored in shared memory.

[0070] The shared memory update module is used to update the shared memory with new information generated by the selected tool after inference is performed.

[0071] The iterative reasoning loop module is used to evaluate the current reasoning state of the large language model based on the information stored in the shared memory in each iteration, and to trigger the corresponding intent tool to perform supplementary reasoning based on the current reasoning state, until enough information is collected to draw a conclusion or the preset iteration limit is reached.

[0072] The above division of modules is merely illustrative. In practical applications, the functions described above can be assigned to different functional modules as needed to complete all or part of the functions described in the aforementioned method. The specific working process of each module can be found in the corresponding process in the aforementioned method embodiments, and will not be repeated here.

[0073] Another embodiment of the present invention provides a computer device (computer, server, smartphone, etc.) including a memory and a processor, the memory storing a computer program configured to be executed by the processor, the computer program including instructions for performing the steps of the method of the present invention.

[0074] Another embodiment of the present invention provides a computer-readable storage medium (such as ROM / RAM, disk, optical disk) storing a computer program that, when executed by a computer, implements the various steps of the method of the present invention.

[0075] The specific embodiments of the present invention disclosed above are intended to help understand the content of the present invention and to implement it accordingly. Those skilled in the art will understand that various substitutions, changes, and modifications are possible without departing from the spirit and scope of the present invention. The present invention should not be limited to the content disclosed in the embodiments of this specification; the scope of protection of the present invention is defined by the claims.

Claims

1. An intention-driven intelligent question answering method based on a large-scale model knowledge graph, characterized in that, Includes the following steps: Receive user input questions and identify the user's potential intentions; Based on the identified intent and the information stored in shared memory, a suitable tool is selected from tools based solely on large language models, tools based on semantic parsing, tools based on subgraph retrieval, and general tools for inference; After reasoning using the selected tool, update the shared memory with the new information generated by the tool; In each iteration, the large language model evaluates the current reasoning state based on the information stored in shared memory, and triggers tools with corresponding intentions to perform supplementary reasoning based on the current reasoning state, until enough information is collected to draw a conclusion or the preset iteration limit is reached.

2. The method according to claim 1, characterized in that, The process of identifying a user's latent intent involves analyzing the problem using predefined intent templates and a large language model to identify the user's latent intent.

3. The method according to claim 1, characterized in that, The information stored in the shared memory includes the current reasoning progress, identified entities and relationships, and the relevant state of the knowledge graph; the new information generated by the tool includes newly identified entities, relationships, and calculated values.

4. The method according to claim 1, characterized in that, Among the tools based solely on large language models, one type focuses on generating hypotheses to retrieve relevant entities, while the other type ensures accuracy by standardizing key information to guarantee that the output does not have inconsistent formats.

5. The method according to claim 1, characterized in that, The semantic parsing-based tool follows the ChatKBQA method. Natural language problems are first converted into logical forms using the semantic understanding of a large language model. Coreference resolution and entity relation disambiguation are then applied to refine these logical forms. The refined logical forms are then converted into executable SPARQL queries to retrieve relevant answers from the knowledge graph. The subgraph-based retrieval tool generates relational paths from user queries and constructs accurate answers using inference paths. It also incorporates graph neural network patterns to enhance reasoning capabilities by retrieving subgraphs while retrieving inference paths, thereby improving performance and scalability in complex tasks.

6. The method according to claim 1, characterized in that, The general tools include tools for extracting key information from knowledge graphs, basic logical operations, and semantic tools; the semantic tools enhance reasoning capabilities by enabling relation retrieval and entity disambiguation, and extend the ability of knowledge graph interactions using pre-trained models.

7. The method according to claim 1, characterized in that, The process of selecting appropriate tools for inference from among tools based solely on large language models, semantic parsing-based tools, subgraph retrieval-based tools, and general-purpose tools, based on the identified intent and information stored in shared memory, includes: If the intent is identified as requiring precise structured knowledge retrieval, and some relevant entity information already exists in the current shared memory, then a semantic parsing-based tool is selected to generate a logical form, which is then converted into a SPARQL query and the answer is retrieved from the knowledge graph. If more potential relationship paths need to be explored during the reasoning process, then choose a tool based on subgraph retrieval to generate relationship paths; When the semantic parsing-based tool reaches two or more iterations, the subgraph retrieval-based tool is activated to leverage the advantages of graph neural networks in multi-hop reasoning to improve reasoning performance.

8. A large-scale knowledge graph question-answering agent based on intent-driven principles, characterized in that... include: The intent recognition module is used to receive user input questions and identify the user's potential intent; The tool selection module is used to select appropriate tools for inference from tools based solely on large language models, semantic parsing-based tools, subgraph retrieval-based tools, and general tools, based on the identified intent and information stored in shared memory. The shared memory update module is used to update the shared memory with new information generated by the selected tool after inference is performed. The iterative reasoning loop module is used to evaluate the current reasoning state of the large language model based on the information stored in the shared memory in each iteration, and to trigger the corresponding intent tool to perform supplementary reasoning based on the current reasoning state, until enough information is collected to draw a conclusion or the preset iteration limit is reached.

9. A computer device, characterized in that, It includes a memory and a processor, the memory storing a computer program configured to be executed by the processor, the computer program including instructions for performing the method of any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a computer, implements the method according to any one of claims 1 to 7.

Images