Multi-hop reasoning method based on dynamic reasoning guidance and multistage self-feedback retrieval

Abstract

A multi-hop reasoning method based on dynamic reasoning guidance and multistage self-feedback retrieval belongs to the field of natural language processing, and comprises the following steps: deconstructing a multi-hop reasoning process into a target-oriented sequence decision problem, carrying out dynamic reasoning guidance by using a large language model, generating a sub-problem sequence matched with a reasoning progress in real time, and carrying out multi-level self-feedback retrieval on the sub-problem sequence; target document retrieval is guided, and sub-questions are dynamically generated; according to the generated sub-questions, obtaining associated documents by adopting a three-level collaborative retrieval mechanism; and performing information refining on the associated document through a large language model, fusing the refined information into an inference chain, and performing inference to generate an answer. The invention further discloses a multi-hop reasoning system, a storage medium and a computer program product. The method aims at solving the complex multi-hop problem that multiple dispersed knowledge fragments need to be integrated, high-accuracy and high-efficiency reasoning is achieved, the retrieval requirement is dynamically generated through an explicit thinking chain guiding mechanism, and evidence obtaining is optimized and redundant information is filtered in combination with a three-level self-feedback retrieval mechanism.

Description

Technical Field

[0001] This invention belongs to the field of natural language processing technology, specifically relating to a multi-hop reasoning method based on dynamic reasoning guidance and multi-level self-feedback retrieval. Background Technology

[0002] Multi-hop reasoning (MHR) is a challenging fundamental task in Natural Language Processing (NLP), particularly crucial in applications such as question answering (QA) and information extraction. Unlike single-hop reasoning, which extracts answers from a single text fragment, MHR requires models to integrate information from multiple disparate pieces of evidence across different documents or sentences, ultimately arriving at a conclusion through a series of logical deduction steps (i.e., a reasoning chain). In MHR, the model must first identify one or more "bridge entities" or intermediate facts connecting the question and the answer, then execute subsequent deduction steps based on these intermediate links to form a complete reasoning path. This ability to structurally integrate long-distance, multi-source information is a key indicator of complex language understanding. Existing MHR methods have been widely applied to document-based question answering, aiming to handle more complex reasoning problems presented in natural language.

[0003] Despite significant progress in multi-hop reasoning, existing technologies still suffer from the following major limitations and shortcomings when dealing with complex real-world documents and reasoning patterns: (1) The accuracy and robustness of the reasoning path construction are insufficient. The core challenge of multi-hop reasoning lies in constructing the reasoning chain. Existing methods are susceptible to textual noise or interfering information at any stage of the reasoning process, especially in the initial stage. Once the bridging entity or key fact in an intermediate step is misidentified, the reasoning chain breaks, causing all subsequent steps to deviate from the correct path, resulting in poor model robustness. Furthermore, for complex queries that require multiple conditions to be met (such as "intersection type" problems), existing models often lack the ability to integrate and judge multi-source evidence.

[0004] (2) Lack of an effective dynamic reasoning guidance mechanism Most existing multi-hop reasoning models employ relatively fixed or static processes to retrieve and combine evidence. They lack dynamic mechanisms for real-time evaluation of the current reasoning state, guiding adjustments to the focus and direction of information retrieval based on the progress of reasoning. This static reasoning model struggles to adapt to the diversity of problems and the complexity of documents. When documents contain a large amount of noise and redundant information, the model cannot effectively focus on the key entities and relationships required for the current reasoning step, leading to inefficient reasoning and increased error rates.

[0005] (3) The retrieval and evidence feedback mechanisms are simple and weak. In complex multi-hop reasoning scenarios, evidence is often scattered across different levels of a document (such as sentences, paragraphs, or the entire document). Existing retrieval mechanisms typically acquire evidence at a single level and lack a multi-level, fine-grained self-feedback mechanism. This means: Limited retrieval levels: The model struggles to work collaboratively across multiple levels (such as from coarse-grained document retrieval to fine-grained sentence or argument retrieval), making it unable to accurately locate and extract arguments.

[0006] Insufficient feedback correction: The model lacks self-feedback and optimization capabilities for the retrieved evidence. Incorrect intermediate inferences are easily carried over to subsequent steps and cannot be effectively corrected, making it difficult to guarantee the accuracy of the final answer.

[0007] (4) Long-distance dependence and information interference are serious problems. When dealing with long or multi-document tasks, the key information required may be geographically distant, resulting in a serious long-distance dependency problem. At the same time, documents often contain background information unrelated to the target event or problem, as well as other interfering information. Existing models struggle to effectively isolate this noise, which severely interferes with the construction of the core reasoning path and the extraction of key facts. Summary of the Invention

[0008] The purpose of this invention is to address the problems in the prior art by providing a multi-hop reasoning method based on dynamic reasoning guidance and multi-level self-feedback retrieval. This method aims to handle complex multi-hop problems that require the integration of multiple scattered knowledge fragments, achieving high accuracy and high efficiency in reasoning. It dynamically generates retrieval requirements through an explicit chain-of-thought (CoT) guidance mechanism and combines a three-level self-feedback retrieval mechanism to optimize evidence acquisition and filter redundant information.

[0009] To achieve the above objectives, the present invention provides the following technical solution: Firstly, a multi-hop reasoning method based on dynamic reasoning guidance and multi-level self-feedback retrieval is provided, including: The multi-hop reasoning process is deconstructed into a goal-oriented sequential decision problem, and a large language model is used for dynamic reasoning guidance to generate a sequence of sub-problems that match the reasoning progress in real time, guiding the retrieval of target documents and dynamically generating sub-problems. Based on the generated sub-questions, a three-level collaborative retrieval mechanism is used to obtain related documents; By refining information from related documents using a large language model, the refined information is integrated into the reasoning chain to generate the answer.

[0010] As a preferred embodiment, the process of deconstructing the multi-hop reasoning process into a goal-oriented sequential decision problem, and using a large language model for dynamic reasoning guidance, generates a sequence of sub-problems in real time that matches the reasoning progress to guide target document retrieval. The steps of dynamically generating sub-problems include: during reasoning... i Step, receive the current original question. and accumulated reasoning information As input; Before generating the final answer, the large language model is guided by an explicit thought chain mechanism, which forces the generation of intermediate reasoning steps. Based on the currently generated thought chain content, the large language model analyzes the core information or missing bridge entities required for the next reasoning step, and generates a sub-problem accordingly. Through sub-problems Define the target of the next search step, as shown in the following expression:

[0011] in, Representing the i The reasoning state of the step, This is a sub-question that guides the next step of the retrieval.

[0012] As a preferred embodiment, the accumulated reasoning information For the reasoning i -1. Evidence retrieved and thought chain generated in step 1.

[0013] As a preferred embodiment, the step of obtaining related documents using a three-level collaborative retrieval mechanism based on the generated sub-questions includes: The first level uses a dense search engine to retrieve relevant documents for sub-questions from the vector knowledge base, and a relevance judge filters out irrelevant noise and retains documents with the required relevance. If the first-level retrieval fails, the second level calls the large language model to generate relevant knowledge, and uses the "knowledge boundary awareness" instruction to avoid illusions and supplement common sense or long-tail information not covered by the knowledge base. If the first two levels of retrieval fail, the third level extracts the key entities in the sub-questions and re-retrieves based on the key entities to expand information coverage and avoid retrieval failures caused by deviations in the expression of factor questions.

[0014] As a preferred approach, in the step of refining information from related documents using a large language model, integrating the refined information into the reasoning chain, and generating an answer, the historical reasoning chain is combined to retain only the key facts directly related to the current reasoning step, and to unify terminology and logical expressions, so as to ensure that the refined information can be directly integrated into the reasoning chain and avoid irrelevant information from interfering with subsequent reasoning.

[0015] As a preferred approach, the steps for generating the answer through reasoning involve iteratively generating sub-questions, using a three-level collaborative retrieval mechanism to obtain related documents, refining the information in the related documents using a large language model, and integrating the refined information into the reasoning chain to achieve multi-hop reasoning until the final answer is generated, with the number of iterations not exceeding a preset limit.

[0016] Secondly, this paper provides an application of the multi-hop reasoning method based on dynamic reasoning guidance and multi-level self-feedback retrieval in multi-hop question answering tasks on datasets such as HotpotQA, 2WikiMultihopQA, or StrategyQA.

[0017] Thirdly, a multi-hop reasoning system based on dynamic reasoning guidance and multi-level self-feedback retrieval is provided, including: The sub-problem generation module is used to deconstruct the multi-hop reasoning process into a goal-oriented sequence decision problem, and uses a large language model for dynamic reasoning guidance to generate a sequence of sub-problems that matches the reasoning progress in real time, guiding the retrieval of target documents and dynamically generating sub-problems. The three-level collaborative retrieval module is used to retrieve related documents based on the generated sub-questions using a three-level collaborative retrieval mechanism. The Information Refinement and Reasoning module is used to refine information from related documents using a large language model, integrate the refined information into the reasoning chain, and generate answers through reasoning.

[0018] Fourthly, a computer-readable storage medium is provided, wherein at least one instruction is stored in the computer-readable storage medium, the at least one instruction being executed by a processor in an electronic device to implement the multi-hop reasoning method based on dynamic reasoning guidance and multi-level self-feedback retrieval.

[0019] Fifthly, a computer program product is provided, wherein when the instructions in the computer program product are executed by the processor of an electronic device, the electronic device performs the multi-hop reasoning method based on dynamic reasoning guidance and multi-level self-feedback retrieval.

[0020] Compared with the prior art, the present invention has at least the following beneficial effects: This invention is based on a multi-hop reasoning method using dynamic reasoning guidance and multi-level self-feedback retrieval. By introducing an explicit thought chain (CoT) guidance mechanism and a three-level self-feedback retrieval framework, it effectively solves key problems in existing technologies such as low retrieval efficiency, easily broken reasoning logic, and noise accumulation. Specifically, it includes the following aspects: 1. Significantly improves reasoning efficiency and reduces retrieval costs Achieving goal-oriented and precise retrieval: Existing methods often perform blind or global searches, resulting in a large number of invalid queries. This invention guides retrieval through dynamic reasoning, enabling the Large Language Model (LLM) to dynamically generate a sequence of sub-questions that precisely matches the reasoning progress based on the current reasoning state and explicit thought chain. This makes the retrieval requirements highly targeted.

[0021] Significantly reduced retrieval count: Due to the precise and efficient retrieval targets, the method of this invention can complete multi-hop reasoning tasks with the lowest number of retrievals. According to experimental results, the method of this invention achieves optimal reasoning performance while having the lowest number of retrievals, thereby significantly reducing the large amount of computation time and resource consumption caused by the retrieval process.

[0022] Reduced average time per question: Due to the reduction in the number of retrievals, the method of this invention maintains a relatively low average time per question. Although each round of retrieval may involve calling a large language model, the overall efficiency is significantly superior to existing methods that rely on multiple retrievals due to the reduction in the total number of retrievals.

[0023] 2. Enhance the robustness and accuracy of reasoning paths. Effectively avoids logical breaks in reasoning: Existing methods are susceptible to noise interference or intermediate errors during the reasoning process, leading to breaks in the reasoning chain. This invention employs a three-level collaborative retrieval mechanism to obtain related documents, dynamically selecting appropriate retrieval strategies and performing quality assessment and self-correction on the retrieved evidence. By returning negative feedback signals to the dynamic guidance module, the model can promptly correct intermediate reasoning errors, increasing the probability of obtaining accurate information and reducing the probability of reasoning proceeding in an erroneous direction.

[0024] Achieving a high signal-to-noise ratio for information: By refining information from related documents through a large language model, redundant information and noise are filtered out, and only refined, coherent, and high signal-to-noise ratio key facts are used for the next step of thought chain generation. This effectively solves the problem of noise accumulation in traditional multi-hop reasoning and ensures the coherence of the reasoning chain and the accuracy of the facts.

[0025] 3. Improve the interpretability and generalization ability of the model. Transparent Reasoning Process (High Interpretability): The method of this invention uses an explicit thought chain (CoT) mechanism to intuitively demonstrate the complete retrieval and reasoning process of the model from question to answer. Users can clearly see the effective content provided by the retrieved information and the logical derivation steps of the model, greatly enhancing the credibility and interpretability of the model.

[0026] Adaptable to complex multi-hop reasoning tasks: The method of this invention is designed to handle different types of multi-hop question-answering tasks. Experiments demonstrate that the method of this invention achieves state-of-the-art results (F1 score and accuracy) on multiple multi-hop question-answering datasets, reflecting its superior generalization ability and high performance for complex reasoning patterns.

[0027] In summary, this invention improves upon existing technologies in terms of accuracy, efficiency, and robustness. Attached Figure Description

[0028] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0029] Figure 1 A schematic diagram of a multi-hop reasoning system architecture based on dynamic reasoning guidance and multi-level self-feedback retrieval according to an embodiment of the present invention; Figure 2 A flowchart illustrating the process of obtaining related documents using a three-level collaborative retrieval mechanism in this embodiment of the invention; Figure 3 This invention provides an example of a multi-hop reasoning method based on dynamic reasoning guidance and multi-level self-feedback retrieval, illustrating its application. Detailed Implementation

[0030] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, those skilled in the art can obtain other embodiments without creative effort.

[0031] This invention provides a multi-hop reasoning method based on dynamic reasoning guidance and multi-level self-feedback retrieval, comprising the following steps: S1. Deconstruct the multi-hop reasoning process into a goal-oriented sequential decision problem, and use a large language model for dynamic reasoning guidance to generate a sequence of sub-problems that match the reasoning progress in real time, guiding the retrieval of target documents and dynamically generating sub-problems; S2. Based on the generated sub-questions, a three-level collaborative retrieval mechanism is used to obtain related documents; S3. Refine the information of related documents through a large language model, integrate the refined information into the reasoning chain, and generate the answer through reasoning.

[0032] In one possible implementation, step S1 mainly includes the following process: Input and State Encoding: In Reasoning i Step, receive the current original question. and accumulated reasoning information (i.e., the evidence retrieved in the first i-1 steps and the generated thought chain) are used as input; Explicit thought chain generation: Before generating the final answer, the large language model is guided by the explicit thought chain (CoT) mechanism, which forces the generation of intermediate reasoning steps; Dynamic sub-problem generation: Based on the currently generated thought chain content, the large language model dynamically and proactively analyzes the core information or missing bridge entities required for the next reasoning step. Accordingly, the model generates a precise, goal-oriented sub-problem. Through sub-problems By clearly defining the target of the next search step, the blind or inefficient global search in traditional methods can be avoided.

[0033] in, Representing the i The reasoning state of the step, These are sub-questions that guide the next step of the search. This mechanism ensures that each search step is accurate and purposeful, thereby significantly reducing the overall number of searches and average search time.

[0034] In one possible implementation, in view of the fact that existing methods have a single retrieval strategy and lack the ability to correct low-quality evidence, this invention designs a three-level self-feedback retrieval mechanism, which, together with an information refinement module, ensures that the retrieved evidence is both comprehensive and accurate.

[0035] Step S2, which describes obtaining related documents using a three-level collaborative retrieval mechanism based on the generated sub-questions, includes: Level 1 (Basic Retrieval Layer): A dense retrieval system is used to retrieve documents related to the sub-question from the vector knowledge base. A relevance judge (a fine-tuned classification model) is used to filter out irrelevant noise and retain documents with the required relevance. Second level (large model knowledge retrieval layer): If the first level retrieval fails, the second level calls the large language model (such as Llama3-8B) to generate relevant knowledge, and avoids illusions by using the "knowledge boundary awareness" instruction (to determine whether the model has actually mastered the knowledge), and supplements common sense or long-tail information not covered by the knowledge base. Level 3 (Key Entity Retrieval Layer): If the first two levels of retrieval fail, the third level extracts key entities from the sub-question (such as "Calfreud"), and re-retrieves based on the key entities to expand information coverage and avoid retrieval failures caused by the expression bias of the factor question.

[0036] In one possible implementation, the retrieved original documents may contain redundant information, disrupting the coherence of the reasoning chain. In this embodiment, step S3 combines the historical reasoning chain, retaining only key facts directly related to the current reasoning step (e.g., extracting only "career experience" from "Melania's biographical documents"), and standardizing terminology and logical expressions to ensure that the refined information can be directly integrated into the reasoning chain, avoiding irrelevant information interfering with subsequent reasoning.

[0037] Furthermore, the multi-hop reasoning method based on dynamic reasoning guidance and multi-level self-feedback retrieval in this embodiment of the invention achieves multi-hop reasoning by iteratively executing steps S1 to S3: reasoning to generate sub-problems → retrieving accurate information → refining and purifying information → continuing reasoning based on new information until the final answer is generated, with the number of iterations not exceeding a preset upper limit (such as 6 times) to balance effect and efficiency.

[0038] The multi-hop reasoning method based on dynamic reasoning guidance and multi-level self-feedback retrieval in this invention can be applied to multi-hop question answering tasks on datasets such as HotpotQA, 2WikiMultihopQA, or StrategyQA.

[0039] This invention also proposes a multi-hop reasoning system based on dynamic reasoning guidance and multi-level self-feedback retrieval, comprising: The sub-problem generation module is used to deconstruct the multi-hop reasoning process into a goal-oriented sequence decision problem, and uses a large language model for dynamic reasoning guidance to generate a sequence of sub-problems that matches the reasoning progress in real time, guiding the retrieval of target documents and dynamically generating sub-problems. The three-level collaborative retrieval module is used to retrieve related documents based on the generated sub-questions using a three-level collaborative retrieval mechanism. The Information Refinement and Reasoning module is used to refine information from related documents using a large language model, integrate the refined information into the reasoning chain, and generate answers through reasoning.

[0040] For multi-hop reasoning tasks, this invention proposes a multi-hop reasoning model based on dynamic reasoning-guided multi-level retrieval (DRMR), such as... Figure 1The diagram illustrates the process of DRMR in solving the multi-hop problem. The model employs a modular architecture, implemented using three components: the inference module, the core, is responsible for problem decomposition and inference decisions; the retrieval module handles information retrieval; and the refinement module ensures the quality of information integration. These three modules interact iteratively, breaking down the complex multi-hop problem into simpler subtasks, executing them iteratively until the complex multi-hop problem is answered. The inference and refinement modules are implemented based on a large language model, while the retrieval module supplements traditional dense retrieval techniques with a large language model.

[0041] The model's reasoning process uses an incremental reasoning mechanism, iteratively executing reasoning steps to form a process like... Figure 1 The reasoning chain shown on the right illustrates the retrieval reasoning process and addresses the multi-hop problem. Each iteration typically includes three key steps: First, the reasoning module performs reasoning analysis based on the current knowledge state, dynamically determines information needs, and generates precise sub-question queries. This enables the large language model to dynamically determine when and what specific content to retrieve based on historical reasoning information, significantly improving the targeting and quality of sub-question generation. Figure 1 The reasoning chain shown indicates that the reasoning and sub-problems are generated by this module. The reasoning part is mainly divided into information integration reasoning (represented by red text) and demand information judgment (represented by green text). Based on the reasoning results, the model generates corresponding sub-problems for querying. Secondly, the retrieval module employs a multi-level feedback mechanism to execute queries, selecting the optimal retrieval method through a self-feedback mechanism, effectively reducing retrieval noise. Figure 1 As shown in the diagram, the generated sub-questions are sent to retrieve relevant information, corresponding to the parts highlighted in red in the reasoning chain. Finally, the refinement module performs structured extraction and noise reduction on the retrieved information, ensuring that key information is retained while maintaining logical consistency with the reasoning chain, corresponding to the refined information content in the reasoning chain. This iterative process continues until the reasoning module determines that enough evidence has been accumulated to generate the final answer. The reasoning module then organizes and answers the question, obtaining the summary in the reasoning chain, i.e., the answer required for the question. The algorithm for the iterative process is shown in the table below:

[0042] A typical human strategy for dealing with complex problems is to solve them step by step: first, identify the initial sub-problems to be solved, retrieve relevant information by searching knowledge, and then perform reasoning based on that information; subsequently, determine subsequent sub-problems based on the current progress, and repeat the information integration and reasoning steps until the problem is completely solved. This mechanism reduces the overall difficulty of solving the problem through hierarchical decomposition, while simultaneously achieving dynamic correlation and deep fusion of multi-source information. As a classic paradigm for task decomposition, Chain-of-Thought (CoT) technology is highly compatible with the logic of human problem-solving for complex issues—first constructing intermediate reasoning paths, and then deriving the final conclusion. Its core lies in guiding the model to output the thought process step by step rather than directly generating the answer, reducing the task difficulty and enhancing the interpretability of the results by explicitly modeling the reasoning steps.

[0043] This invention draws on this cognitive paradigm and uses thought chain generation technology to simulate the cognitive process of solving multi-hop problems in humans.

[0044] In the proposed DRMR framework, the reasoning module serves as the core component for constructing the thought chain, guiding its extension through a sub-problem generation mechanism enhanced by dynamic reasoning. This module utilizes structured prompting engineering to guide the large language model to complete tasks through contextual learning, achieving deep synergy between the dynamic reasoning process and knowledge retrieval. Details of the prompts are shown in Table 1.

[0045] Specifically, for a given multi-hop problem Reasoning module The system will first analyze the question, determine the information needed to answer it, and generate a reasoning process. Explain the content to be retrieved in the current step, and then generate sub-questions to be retrieved based on this reasoning process. The initial reasoning process can be expressed as follows:

[0046] After completing the initial reasoning step, DRMR will use the generated subproblems. Interact with the search and refinement modules to obtain relevant information. And add it to the mind chain =[ [In the context of] the updated thought chain, the reasoning model Based on the multi-hop issue It reasones with the thought chain containing previous reasoning steps and related information, and decomposes the problem based on the reasoning to generate new sub-problems. The model will iteratively perform the above reasoning process until the [number]th iteration. The reasoning process of the steps can be expressed as follows:

[0047] The model will iteratively perform the above reasoning process until the [number]th iteration. The reasoning process of the first step, at this time the reasoning module is based on the thought chain. =[ We now have enough information to reason our way to the answer. Or the set search limit has been reached. The formula is expressed as:

[0048] Based on the process described above, the inference module's iterative problem decomposition process based on known information can be described as follows:

[0049]

[0050] in, Representing the The next iteration process uses The relevant information obtained through the retrieval and refinement modules Representing the During the round of iterative retrieval, there is already enough information to reason and obtain the answer. Or the set search limit has been reached. .

[0051] In designing the prompts for the reasoning module, several key principles were followed to ensure the model could accurately execute the task. As shown in Table 1, firstly, the stages were clearly defined: three clear stages guided the thinking process (reasoning analysis → sub-question generation → retrieval → reasoning analysis → sub-question generation → retrieval → ... → reasoning analysis → conclusion), and a structured thought chain was established using format markers (Reasoning / Sub-question / Information). Secondly, the cognitive process in the thought chain was visualized: the separation of known / unknown information was mandatory, the progressive construction of the evidence chain was emphasized, and the specific dimensions of missing information were required to be explained. Through these constraints, the model can accurately decompose the problem based on known information and ask for the information needed to answer the question. This enhances the ability of the large model to generate accurate responses based on existing information, thereby reducing the risk of illusion.

[0052] Table 1

[0053] The retrieval module is an important module in the DRMR method for retrieving relevant knowledge based on the sub-problems generated by the reasoning planning module. Its performance is directly related to the performance of the entire model, and many studies have experimentally proven this conclusion.

[0054] Traditional retrieval-based question answering systems obtain information through query matching. - A collection of paragraphs The information is then input into a question-answering model for a response. However, the complexity of multi-hop questions makes direct retrieval difficult to obtain all the necessary information, resulting in missing information and noise interference. To address this issue, the present invention employs a dynamic reasoning-enhanced sub-question generation mechanism in the reasoning module proposed earlier. This mechanism decomposes the original question and generates easier-to-retrieve and answer sub-questions through analysis and reasoning, reducing the difficulty of accurate retrieval. However, this iterative retrieval process based on question decomposition may lead to two derivative problems: 1. The generated sub-question formulation is not conducive to retrieval, causing retrieval failure. For example, when answering the multi-hop question "Are the director of the 1995 film *Coolie 1* and its director Karl Floyd the same nationality?" in the 2WikiMQA dataset, the sub-question "Was Karl Floyd an Indian director?" retrieves the information "In this way, Modu Bosch became the first filmmaker from India to work closely with the director in a major Hollywood setting. He also worked with Karl Floyd in Ufa." However, this text is irrelevant to the question. The reason for the failure to retrieve relevant information is that Karl Floyd was not from India. Using the question "Was Karl Floyd an Indian director?" makes it difficult to find Karl Floyd's actual nationality, leading to a significant amount of noise in the search. 2. Lack of corresponding information in the knowledge base causes empty searches. For example, when processing the common-sense multi-hop question "Do piranhas eat in crossroads kitchens?" in the StrategyQA dataset, the sub-question "Can piranhas enter kitchens?" will fail to retrieve the information because there are no relevant records in the knowledge base. Previous work based on iterative retrieval ignored these problems, opting to directly use a search engine, resulting in poor search performance. This invention addresses these issues by proposing a multi-level self-feedback retrieval module. This module provides feedback based on whether the currently retrieved content meets the query requirements, employing different retrieval methods accordingly. Specifically, for example... Figure 2 As shown, this invention designs a three-level self-feedback retrieval architecture.

[0055] 1) Basic Search Layer For the sub-question queried from the input in the search engine The retrieval module will first use the current query. Through traditional dense search instruments In the vector knowledge base The search is conducted to ensure the reliability of the information obtained. Regarding the search results... Because it may contain noisy information, this invention uses a finely tuned correlation determiner. To determine the search paragraph and subproblems of the query Is it relevant? If it is relevant, it is included in the search results; otherwise, it is discarded as noise. This can be expressed as:

[0056]

[0057] in, The search results output by the retrieval module represent those processed by the relevance assessor in the first-level retrieval layer. This determines the set of all search fragments relevant to the subquestion of the query. If all retrieved documents are considered irrelevant, then... This indicates that the search was incorrect or that the relevant knowledge was not found in the knowledge base. At this point, a self-feedback mechanism is triggered to perform a second-level search.

[0058] 2) Large-scale model knowledge retrieval layer Because large language models have a large number of parameters and store massive amounts of knowledge, and can better understand query semantics compared to dense retrieval models, this approach can fully leverage the capabilities of large language models to generate relevant knowledge documents when the retrieval layer fails. For the aforementioned question, "Can a man-eating shark enter the kitchen?", which defies common sense, the large language model can easily determine the answer as "no." Therefore, this invention uses it as a second-layer feedback retrieval scheme to supplement the information retrieval scheme when the basic retrieval layer fails. However, large language models suffer from a serious illusion problem, easily generating incorrect or fabricated knowledge. To further prevent the large model from generating illusions that affect the fidelity of the generated answers, this invention references a knowledge boundary awareness algorithm for large language models, allowing the large language model to judge knowledge boundaries, with specific instructions... As shown in Table 2. Knowledge documents are generated only when the model confirms that it possesses relevant knowledge. The retrieval process can be represented by the following formula:

[0059] If the large model cannot answer the current question, that is... If it is still empty, proceed to the third level of search.

[0060] 3) Key Entity Retrieval Layer When both levels of retrieval fail, it indicates a sub-problem. This may contain long-tail knowledge that is difficult to apply directly to retrieval. In this case, it is necessary to consider sub-problems. The goal is to retrieve as much relevant information as possible, rather than information that is entirely related to the entire sub-question. Therefore, the granularity of the search content can be reduced, using only key entities contained within the question to retrieve more relevant information. For example, for the question "Was Karl Floyd an Indian director?", a large model can be used to extract key entities, extracting the key entity "Karl Floyd," and then performing a dense search to obtain information about Karl Floyd's profession and nationality, thus easily determining that "Karl Floyd was not an Indian director." The large model extracts key words to indicate key information. As shown in Table 3: Table 2

[0061] The retrieval process can be represented by the following formula:

[0062] Through the aforementioned multi-level self-feedback retrieval steps, the model can select the appropriate retrieval method for the current scenario, ensuring that relevant content is retrieved while mitigating the impact of retrieval noise to some extent. Furthermore, the retrieval device proposed in this embodiment is a pluggable structure, allowing for adjustments to the hierarchical structure and retrieval device type during actual use. For example, when the required information is typically non-specialized domain knowledge or long-tail knowledge, the model can use a large-model retrieval device as the first layer, typically using a large model for question answering. The second layer is set to a dense retrieval device, used as a supplement when the question reaches the knowledge boundary of the large model and an accurate answer is difficult to determine. For certain scenarios where an overly complex retrieval process is unnecessary, a single retrieval device can be used to ensure retrieval efficiency.

[0063] Table 3

[0064] While the retrieval module helps large language models fill in missing knowledge in reasoning, research has found that directly inserting complete documents may disrupt the coherence of the thought chain due to their length and redundancy, and the excessive length of the thought chain may cause the model to forget knowledge. To overcome this problem, the information refinement module of this invention uses a large language model to selectively integrate relevant and concise information into the reasoning chain through a separate generation process, resulting in prompt words as shown in Table 4.

[0065] Table 4

[0066] This module processes the retrieved documents to align them with the model's specific reasoning needs, transforming raw information into refined, relevant knowledge while maintaining the coherence and logical consistency of the main reasoning chain. This can be formally represented as follows:

[0067] in, Indicates the first The information refined by the information refinement module in each iteration Representing the The information retrieved by the step-by-step retrieval module =[ [] represents the thought process that has been constructed in the current problem reasoning process.

[0068] Figure 3 This paper demonstrates the reasoning process and results of the DRMR method on a multi-hop problem in the StrategyQA dataset. It shows that the model generates sub-questions to be retrieved with the help of reasoning information, enabling it to retrieve relevant documents more accurately for the current step. Furthermore, the intuitive retrieval and reasoning process within the thought chain allows users to clearly see the effective content provided by the retrieved information and the model's reasoning process, demonstrating the model's interpretability and effectiveness. During retrieval, the model selects an appropriate retrieval method through a self-feedback mechanism, allowing it to obtain more comprehensive information. For the question "Is there a common the Quran?", the model uses a large model for knowledge generation, providing accurate knowledge to aid in the reasoning process.

[0069] Another embodiment of the present invention also proposes a computer-readable storage medium storing at least one instruction, which is executed by a processor in an electronic device to implement the multi-hop reasoning method based on dynamic reasoning guidance and multi-level self-feedback retrieval.

[0070] Another embodiment of the present invention also proposes a computer program product, wherein the instructions in the computer program product, when executed by the processor of an electronic device, cause the electronic device to perform the multi-hop reasoning method based on dynamic reasoning guidance and multi-level self-feedback retrieval.

[0071] For example, the instructions stored in the memory can be divided into one or more modules / units. These modules / units are stored in a computer-readable storage medium and executed by the processor to complete the multi-hop inference method based on dynamic inference guidance and multi-level self-feedback retrieval described in this invention. The one or more modules / units can be a series of computer-readable instruction segments capable of performing specific functions, which describe the execution process of the computer program in the server.

[0072] The electronic device may be a smartphone, laptop, PDA, or cloud server, among other computing devices. It may include, but is not limited to, a processor and memory. Those skilled in the art will understand that the electronic device may also include more or fewer components, or combinations of certain components, or different components; for example, it may also include input / output devices, network access devices, buses, etc.

[0073] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor.

[0074] The memory can be an internal storage unit of the server, such as a hard drive or RAM. It can also be an external storage device, such as a plug-in hard drive, Smart Media Card (SMC), Secure Digital (SD) card, or FlashCard. Furthermore, the memory can include both internal and external storage units. The memory is used to store computer-readable instructions and other programs and data required by the server. It can also be used to temporarily store data that has been output or will be output.

[0075] It should be noted that the information interaction and execution process between the above-mentioned module units are based on the same concept as the method embodiment. For details on their specific functions and technical effects, please refer to the method embodiment section. They will not be repeated here.

[0076] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0077] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of this application can be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include at least: any entity or device capable of carrying the computer program code to a photographing device / terminal device, a recording medium, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium. Examples include USB flash drives, portable hard drives, magnetic disks, or optical disks.

[0078] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0079] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A multi-hop reasoning method based on dynamic reasoning guidance and multi-level self-feedback retrieval, characterized in that, include: The multi-hop reasoning process is deconstructed into a goal-oriented sequential decision problem, and a large language model is used for dynamic reasoning guidance to generate a sequence of sub-problems that match the reasoning progress in real time, guiding the retrieval of target documents and dynamically generating sub-problems. Based on the generated sub-questions, a three-level collaborative retrieval mechanism is used to obtain related documents; By refining information from related documents using a large language model, the refined information is integrated into the reasoning chain to generate the answer.

2. The multi-hop reasoning method based on dynamic reasoning guidance and multi-level self-feedback retrieval according to claim 1, characterized in that, The process of deconstructing multi-hop reasoning into a goal-oriented sequential decision problem and using a large language model for dynamic reasoning guidance, generating a sequence of sub-problems in real time that matches the reasoning progress to guide target document retrieval, includes the following steps: during reasoning... i Step, receive the current original question. and accumulated reasoning information As input; Before generating the final answer, the large language model is guided by an explicit thought chain mechanism, which forces the generation of intermediate reasoning steps. Based on the currently generated thought chain content, the large language model analyzes the core information or missing bridge entities required for the next reasoning step, and generates a sub-problem accordingly. Through sub-problems Define the target of the next search step, as shown in the following expression: in, Representing the i The reasoning state of the step, This is a sub-question that guides the next step of the retrieval.

3. The multi-hop reasoning method based on dynamic reasoning guidance and multi-level self-feedback retrieval according to claim 2, characterized in that, The accumulated reasoning information For the reasoning i -1. Evidence retrieved and thought chain generated in step 1.

4. The multi-hop reasoning method based on dynamic reasoning guidance and multi-level self-feedback retrieval according to claim 1, characterized in that, The steps for obtaining related documents using a three-level collaborative retrieval mechanism based on the generated sub-questions include: The first level uses a dense search engine to retrieve relevant documents for sub-questions from the vector knowledge base, and a relevance judge filters out irrelevant noise and retains documents with the required relevance. If the first-level retrieval fails, the second level calls the large language model to generate relevant knowledge, and uses the "knowledge boundary awareness" instruction to avoid illusions and supplement common sense or long-tail information not covered by the knowledge base. If the first two levels of retrieval fail, the third level extracts the key entities in the sub-questions and re-retrieves based on the key entities to expand information coverage and avoid retrieval failures caused by deviations in the expression of factor questions.

5. The multi-hop reasoning method based on dynamic reasoning guidance and multi-level self-feedback retrieval according to claim 1, characterized in that, In the step of refining information from related documents using a large language model, integrating the refined information into the reasoning chain, and generating an answer, the historical reasoning chain is combined to retain only the key facts directly related to the current reasoning step, and to unify terminology and logical expression to ensure that the refined information can be directly integrated into the reasoning chain and to avoid irrelevant information interfering with subsequent reasoning.

6. The multi-hop reasoning method based on dynamic reasoning guidance and multi-level self-feedback retrieval according to claim 1, characterized in that, The steps for generating the answer through reasoning involve iteratively generating sub-questions, using a three-level collaborative retrieval mechanism to obtain related documents, refining the information in the related documents through a large language model, and integrating the refined information into the reasoning chain to achieve multi-hop reasoning until the final answer is generated, with the number of iterations not exceeding a preset limit.

7. An application of the multi-hop reasoning method based on dynamic reasoning guidance and multi-level self-feedback retrieval as described in any one of claims 1 to 6 in multi-hop question answering tasks on the HotpotQA, 2WikiMultihopQA, or StrategyQA datasets.

8. A multi-hop reasoning system based on dynamic reasoning guidance and multi-level self-feedback retrieval, characterized in that, include: The sub-problem generation module is used to deconstruct the multi-hop reasoning process into a goal-oriented sequence decision problem, and uses a large language model for dynamic reasoning guidance to generate a sequence of sub-problems that matches the reasoning progress in real time, guiding the retrieval of target documents and dynamically generating sub-problems. The three-level collaborative retrieval module is used to retrieve related documents based on the generated sub-questions using a three-level collaborative retrieval mechanism. The Information Refinement and Reasoning module is used to refine information from related documents using a large language model, integrate the refined information into the reasoning chain, and generate answers through reasoning.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores at least one instruction, which is executed by a processor in an electronic device to implement the multi-hop reasoning method based on dynamic reasoning guidance and multi-level self-feedback retrieval as described in any one of claims 1 to 6.

10. A computer program product, characterized in that, When the instructions in the computer program product are executed by the processor of the electronic device, the electronic device performs the multi-hop reasoning method based on dynamic reasoning guidance and multi-level self-feedback retrieval as described in any one of claims 1 to 6.

Images