A multi-hop RAG question-answering method and system based on path exploration and hyperbolic refinement

Abstract

This invention relates to a multi-hop RAG question-answering method and system based on path exploration and hyperbolic refinement, addressing the problems of rigid reasoning paths and evidence noise interference in multi-hop question-answering processes. The method includes receiving complex multi-hop queries, employing a multi-level path exploration search strategy, constructing a hierarchical reasoning tree through Monte Carlo tree search, and sequentially performing atomic operations such as direct answering, retrieved answering, and sub-question decomposition to explore the optimal reasoning path; performing hyperbolic refinement on candidate evidence retrieved during the construction process, constructing an evidence graph, and utilizing a graph attention network for information propagation, thereby dynamically reordering the evidence and suppressing noise; recursively aggregating the processed intermediate question-answering results from bottom to top, and combining them with the original query to call a language model to generate the final answer. This invention effectively improves the coherence and accuracy of multi-hop reasoning and is suitable for natural language processing applications such as complex question answering and knowledge reasoning.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, and in particular to a multi-hop RAG question-answering method and system based on path exploration and hyperbolic refinement. Background Technology

[0002] Currently, with the continuous advancement of artificial intelligence and natural language processing technologies, multi-hop question answering has significant application value in various fields such as information retrieval, intelligent customer service, medical diagnosis, and legal consultation. Many practical problems involve cross-domain knowledge and hierarchical reasoning processes, making single-step question answering insufficient to meet users' needs for solving complex problems. Therefore, the concept of Retrieval-Augmented Generation (RAG), which combines external retrieval evidence with the powerful language understanding capabilities of generative models, has become an important technical approach for addressing complex multi-hop question answering.

[0003] However, multi-hop RAGs face numerous technical challenges in their implementation. Key technical difficulties include effectively decomposing complex queries, constructing reasonable multi-level reasoning paths, and balancing the relationship between directly generated evidence and evidence retrieved from retrieval sites. Furthermore, the diversity and potential noise of candidate evidence significantly interfere with the continuity of reasoning paths and the accuracy of answers, making it difficult to guarantee the overall system stability.

[0004] Current multi-hop RAG technologies generally rely on predefined rules or static knowledge graphs to construct reasoning paths, resulting in limited ability to capture the inherent logical relationships within complex queries. Furthermore, existing methods typically lack effective sorting and noise filtering mechanisms for processing retrieved evidence, leading to redundant information being mixed into the final answer, thus affecting question-answering performance and semantic coherence. These shortcomings make it difficult for existing technologies to fully meet the needs of practical applications when dealing with highly complex problems involving multiple overlapping pieces of information. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the existing technology by providing a multi-hop RAG question-answering method and system based on path exploration and hyperbolic refinement, so as to solve or partially solve the problems of fixed and rigid reasoning paths and the existence of evidence noise interference.

[0006] The objective of this invention can be achieved through the following technical solutions:

[0007] According to one aspect of the present invention, a multi-hop RAG question answering method based on path exploration and hyperbolic refinement is provided, comprising the following steps:

[0008] The system acquires multi-hop query input, constructs a multi-level reasoning tree based on a recursive path exploration search strategy, and obtains the optimal subtree through Monte Carlo tree search. The path exploration search strategy includes direct answering, answering after retrieval, and sub-question decomposition.

[0009] During the path exploration search process, for the node that provides the answer after retrieval for the current query, multiple candidate evidences retrieved are mapped to hyperbolic space for hyperbolic refinement, and evidence with high relevance to the current query is selected. Combined with the current query, an intermediate answer is generated as the intermediate node of the reasoning tree.

[0010] For each node in the optimal subtree, the final answer is obtained through recursive aggregation.

[0011] As a preferred technical solution, the process of mapping the retrieved multiple candidate evidences to hyperbolic space for hyperbolic refinement includes the following steps:

[0012] An evidence graph structure is constructed based on multiple retrieved candidate evidence;

[0013] Based on the aforementioned evidence graph structure, the embeddings of the query and noise-suppressed Euclidean space are mapped to hyperbolic space. Then, a graph attention network is used for graph propagation learning to propagate and update the information of the nodes in the evidence graph structure. The evidence is then sorted according to the relative characteristics of the node embedding vectors in the evidence graph structure.

[0014] The hyperbolic space embedding mapping after query and evidence ranking and noise suppression is used to filter the multiple pieces of evidence with the highest relevance to the current query through similarity calculation.

[0015] As a preferred technical solution, the reasoning tree includes nodes of the following types that alternate according to the tree's hierarchy:

[0016] Operation status nodes include information about the path exploration and search strategies used;

[0017] Text-based nodes include the current query, evidence, and answer information.

[0018] As a preferred technical solution, the root node of the reasoning tree is a text-type node. After selecting a path exploration search strategy for the query in the text-type node, an operation state-type node is obtained. Based on the selected path exploration search strategy, the corresponding text-type child nodes are further obtained.

[0019] As a preferred technical solution, the recursion is stopped during the recursive construction of the reasoning tree when a preset tree height limit is reached, or when all subproblems are single-hop problems.

[0020] As a preferred technical solution, the Monte Carlo tree search process includes the following steps:

[0021] Selection: Based on the number of times a node is visited and its cumulative score, the optimal node N to be expanded is selected from the current tree using the upper confidence bound criterion;

[0022] Exploration: Multiple path exploration search strategies are used for node N;

[0023] Extension: Insert the new nodes obtained in the exploration steps and their intermediate answers into the reasoning tree;

[0024] Backtracking: Update the statistics of the ancestral nodes based on the evaluation score of the new node to guide subsequent selections.

[0025] As a preferred technical solution, the process of obtaining the query output through recursive aggregation includes:

[0026] The nodes in the optimal subtree are recursively aggregated from bottom to top, and combined with the original query input, the final query output is generated by calling the language model. For each text-type node in the optimal subtree, the current query and the corresponding answer are processed through a prompting process to allow the large language model to summarize the answer of the current query layer by layer, and finally aggregated into the answer of the initial query.

[0027] As a preferred technical solution, the process of generating intermediate answers includes the following steps:

[0028] Based on the filtered evidence that is highly relevant to the current query and the current query, an intermediate answer is generated using a language model.

[0029] As a preferred technical solution, the path exploration search strategy also includes using a language model to directly answer simple or single-hop problems.

[0030] Another aspect of the present invention provides a multi-hop RAG question answering system based on path exploration and hyperbolic refinement, for implementing the aforementioned multi-hop RAG question answering method based on path exploration and hyperbolic refinement, wherein the multi-hop RAG question answering system includes:

[0031] A multi-level path exploration module is used to obtain multi-hop query inputs, construct a multi-level reasoning tree based on a recursive path exploration search strategy, and obtain the optimal subtree through Monte Carlo tree search. The path exploration search strategy includes direct answering, answering after retrieval, and sub-question decomposition.

[0032] The hyperbolic evidence refinement module is used during the path exploration and search process to map multiple candidate evidence retrieved to the hyperbolic space for hyperbolic refinement for nodes that provide answers to the current query. It filters out evidence that is highly relevant to the current query and generates intermediate answers based on the current query, which serve as intermediate nodes of the reasoning tree.

[0033] The answer aggregation unit is used to recursively aggregate each node in the optimal subtree to obtain the final answer.

[0034] Compared with the prior art, the present invention has at least one of the following beneficial effects:

[0035] (1) Improved coherence and accuracy of multi-hop reasoning: This invention effectively solves the problems of rigid reasoning paths and noise interference of evidence in the prior art by dynamically constructing reasoning trees and hyperbolic refining of candidate evidence. The method effectively improves the coherence and accuracy of multi-hop reasoning and is suitable for natural language processing applications such as complex multi-hop question answering and knowledge reasoning.

[0036] (2) The relevance and robustness of evidence screening are optimized: The present invention first maps the retrieved candidate evidence to hyperbolic space, and then refines it by combining graph attention network. By taking advantage of the hierarchical nature of hyperbolic geometry, the evidence is dynamically reordered and noise is suppressed, which effectively improves the relevance and quality of the screened evidence, thereby enhancing the robustness and accuracy of the question answering system in complex scenarios.

[0037] (3) Enhanced interpretability and flexibility of reasoning path: The present invention dynamically constructs a hierarchical reasoning tree based on Monte Carlo tree search, and alternates between operation state nodes and text nodes to clearly present the strategy selection and answer generation process at each step; this structure not only improves the transparency of the reasoning process, but also allows for flexible adjustment of the tree depth and branches according to the query complexity, thereby improving the overall system controllability. Attached Figure Description

[0038] Figure 1 The flowchart shows the multi-hop RAG question-answering method based on path exploration and hyperbolic refinement in the embodiment.

[0039] Figure 2 This is a schematic diagram of the framework of the multi-hop RAG question-answering method in the embodiment;

[0040] Figure 3 This is a schematic diagram of a multi-hop RAG question-answering system based on path exploration and hyperbolic refinement in the embodiment;

[0041] Figure 4 The bar chart shows the test results of the multi-hop RAG method for ablation experiments based on the exact match (EM) value in the embodiment.

[0042] Figure 5 This is a bar chart showing the test results of the multi-hop RAG method for ablation experiments based on the F1 score in the embodiments.

[0043] Figure 6 This is a schematic diagram of the electronic device in the embodiment. Detailed Implementation

[0044] 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, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0045] Example 1

[0046] To address the problems existing in the aforementioned prior art, this embodiment provides a multi-hop RAG question-answering method based on path exploration and hyperbolic refinement, see [link to relevant documentation]. Figure 1 The method includes the following steps:

[0047] Step S1: Receive input containing complex multi-hop queries and process the queries using a multi-level path exploration search strategy.

[0048] Specifically, the system receives input containing complex multi-hop queries and processes the queries using a multi-level path exploration search strategy. The exploration strategy includes three atomic operations: direct answer, search-based answer, and sub-question decomposition. A hierarchical reasoning tree is constructed based on the Monte Carlo Tree Search (MCTS) strategy. The node with the highest score in the current path is selected according to the score of each node N in the MCTS strategy, and finally, an optimal subtree is formed as the multi-hop solution path.

[0049] Step S2: Map the retrieved candidate evidence to hyperbolic space and perform hyperbolic refinement.

[0050] Specifically, this includes: constructing evidence graphs, using Graph Attention Network (GAT) for graph propagation learning, and using hypersphere mapping to map Euclidean embeddings to hyperbolic space, thereby achieving dynamic reordering of evidence and noise suppression.

[0051] Step S3 is performed recursively to generate intermediate answers.

[0052] Specifically, the evidence highly relevant to the current question obtained in step S2 is combined with the current query, and an intermediate answer is generated by calling the language model, which serves as an intermediate node in the reasoning tree.

[0053] In step S4, the nodes in the best answer tree are recursively aggregated from bottom to top to generate the final answer.

[0054] Specifically, the nodes in the best answer tree obtained in steps S1 and S3 are recursively aggregated from bottom to top, and combined with the original query to generate the final answer by calling the language model.

[0055] In one embodiment, such as Figure 2 As shown, the atomic operations under the multi-level exploration path strategy include: direct answering, which directly calls the language model to generate an answer for simple or single-hop queries; retrieval-based answering, which calls the language model to generate an answer after appending retrieval evidence information to the query; and sub-problem decomposition, which performs semantic decomposition on complex problems to generate several sub-problems and recursively executes atomic operations until a complete reasoning path is formed.

[0056] Specifically, for direct answers to atomic operations (e.g., subquery: What country was Maxwell from?), the Prompt project is used to input the question into a Large Language Model (LLM). LLMs include, but are not limited to, Mistral-7B, Qwen2-7B, and Qwen2-72B. The LLM then directly generates the answer to this query. This is a preferred approach, fully utilizing the rich prior knowledge within the LLM while avoiding the introduction of additional noise that could negatively impact performance. For answers after retrieval of atomic operations (e.g., subquery: When was the first Nobel Prize in Physics awarded to the United States?), this method integrates the RAG workflow, a pipeline for generating results after retrieval, to handle simpler queries requiring additional knowledge. This is achieved by initiating an initial retrieval tool, such as Best Matching 25. (BM25) Retrieves evidence relevant to the current query from a graph-structured knowledge base (built through sentence keyword matching or semantic similarity connections). This evidence is then further refined using a hyperbolic refinement method to select the highest-quality, most relevant (Top-K) evidence. A Prompt process is then used to generate the current answer through the LLM. For sub-question decomposition of atomic operations (e.g., initial query: "Which country first won the Nobel Prize in Physics, the country that proposed the theory of electromagnetic waves or the United States?"), the current query Prompt process allows the LLM to break down complex multi-hop problems into several logically simpler questions for step-by-step solution. All sub-questions need to be answered here. As a preferred approach, the traditional RAG process focuses more on retrieval and cannot better address the reasoning needs of multi-hop problems. Decomposing multi-hop problems into sub-questions effectively improves reasoning performance. It is important to note that the templates in the Prompt process described above are carefully designed.

[0057] In one embodiment, such as Figure 2As shown, the hierarchical reasoning tree structure includes two different types of nodes: operation state nodes and question-and-answer text nodes. Specifically, operation state nodes contain the selected atomic operation type, while question-and-answer text nodes contain the actual current query, evidence, and answer information. These two types of nodes alternate according to the tree's hierarchy. The root node of the tree is a question-and-answer text node. After performing atomic operations on the query contained in the current node, a child node of the operation state type is obtained. Subsequently, based on the different atomic operations, corresponding child nodes of the question-and-answer text type are obtained. The child node obtained from a direct answer or a response after retrieval is a single node, which should contain the query answer. The sub-problem decomposition of atomic operations leads to tree branching, resulting in several child nodes each containing a sub-problem. It is important to note that sub-problem decomposition will proceed recursively until a preset tree height limit is reached, or all sub-problems are confirmed as single-hop problems.

[0058] In one embodiment, such as Figure 2 As shown, the MCTS search strategy includes: selection of nodes in the hierarchical reasoning tree; exploration; expansion; and backpropagation. Specifically, it includes four steps: Selection: Based on the number of times a node is visited and its cumulative score, the optimal node N to be expanded is selected in the current tree using the Upper Confidence Bound (UCB) criterion; Exploration: The three atomic operations mentioned above are attempted on node N; Expansion: The new node obtained in the exploration step and its intermediate answers are inserted into the reasoning tree; Backpropagation: The statistics of the ancestor nodes are updated based on the evaluation score of the new node to guide subsequent selections. The above process is repeated until the stopping condition is met (such as reaching the maximum number of nodes or obtaining a subtree with a sufficiently high score), and the "best answer subtree" T is selected from the tree.

[0059] In one embodiment, such as Figure 2 As shown, hyperbolic refinement of candidate evidence includes: constructing a graph structure for multiple retrieved pieces of evidence, using a graph attention network to propagate and update information between nodes in the graph to dynamically adjust the correlation between each piece of evidence, and sorting and suppressing noise based on the relative characteristics of the embedding vectors.

[0060] Specifically, all the evidence initially retrieved is combined with the current query to construct a fully connected graph. This means that we assume all the current evidence and the query are highly correlated with each other. Each piece of evidence and the query is treated as a node in the fully connected graph (this graph is not related to the graph structure in the knowledge base). This node includes: content (evidence or query), graph ID in the knowledge base (if any), neighbors in the knowledge base (if any), and content features (such as embeddings). This fully connected graph is then learned through graph attention network (GAT). After multiple rounds of information propagation, the features of each node in the previously obtained fully connected graph are updated, and the weights of each edge change. At this point, the GAT network has learned the correlation between the features of the evidence and the query in the graph through the attention mechanism. The top-K features of the evidence with the highest weights that are most similar to the query features are then selected, i.e., k pieces of evidence highly correlated with the query are found, further reducing noise interference to LLM. It is worth noting that the embedding model here uses a multifactor dimensionality reduction (MDR) model trained when building the knowledge base.

[0061] In one embodiment, such as Figure 2 As shown, hyperspherical mapping involves mapping the embeddings of queries and evidence from Euclidean space to hyperbolic space, such as the Poincaré sphere model. This avoids the embedding distortion problem inherent in Euclidean space (where, in tree or graph structures with excessive hop counts, embeddings in Euclidean space exhibit mismatches such as close Euclidean distances but low semantic similarity in actual sentences), better capturing the semantic hierarchy information between embeddings. Specifically, in the embedding selection process, the Euclidean space embeddings are hyperbolic mapped using a special calculation formula between the Poincaré sphere and Euclidean space. It's important to note that the Möbius multiplication and Möbius addition operations in hyperbolic space differ somewhat from multiplication and addition in Euclidean space.

[0062] In one embodiment, such as Figure 2 As shown, bottom-up recursive aggregation includes: aggregating the sub-questions and corresponding results of each intermediate question and answer in a bottom-up manner according to the hierarchical order of the best answer tree, recursively obtaining the answer to the parent question, until the final answer is generated. Specifically, for each question and answer text class node in the tree, the current query and its corresponding answer are used to enable the LLM to summarize the answer to the current query layer by layer through the Prompt project, and finally aggregated into the answer to the initial query.

[0063] In summary, compared with existing technologies, this application proposes a general multi-hop RAG question answering method and framework. First, it maps questions and candidate evidence to a hyperbolic space and refines them hierarchically using a graph attention network, transforming complex knowledge graph reasoning into a simple iterative exploration process. Then, it dynamically generates and evaluates reasoning paths through Monte Carlo tree search, selecting the optimal subtree using a comprehensive scoring function composed of relevance, language model confidence, and path length, achieving efficient and accurate answer derivation. This framework supports plug-and-play (PnP) replacement of retrieval tools, models, and graph algorithms, meeting the need for rapid cross-domain portability. Simultaneously, it introduces a large language model for iterative verification and supervised feedback of intermediate answers, continuously improving multi-hop reasoning capabilities. Finally, by integrating the natural language interaction advantages of the large language model with the structured reasoning capabilities of graph reasoning algorithms, it provides knowledge graph users with a natural, easy-to-use, and highly interpretable question answering experience.

[0064] Example 2

[0065] Building upon Example 1, this example provides a multi-hop RAG question-answering system based on path exploration and hyperbolic refinement, such as... Figure 3 ,include:

[0066] The multi-level path exploration module 302 is used to process inputs containing complex multi-hop queries and generate hierarchical reasoning trees based on Monte Carlo tree search. This module supports atomic operations such as direct answering, searched answering, and sub-question decomposition.

[0067] The hyperbolic evidence refinement module 304 is used to construct graph structures, update graph attention propagation, and transform embeddings for candidate evidence retrieved by the multi-level path exploration module, thereby achieving dynamic reordering of evidence and noise suppression.

[0068] Answer aggregation unit 306 is used to recursively aggregate the intermediate results generated by the multi-level path exploration module and the hyperbolic refinement module from bottom to top, and combine them with the original query to generate the final answer by calling the language model.

[0069] For specific constraints on multi-hop RAG systems based on path exploration and hyperbolic refinement, please refer to the constraints on multi-hop RAG methods based on path exploration and hyperbolic refinement mentioned above, which will not be repeated here.

[0070] The modules in the aforementioned multi-hop RAG system based on path exploration and hyperbolic refinement can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in hardware or independent of the processor in a computer device, or stored in software within the computer device's memory, allowing the processor to invoke and execute the corresponding operations of each module.

[0071] To verify the effectiveness of the present invention, experimental tests were conducted on the multi-hop RAG method and system based on path exploration and hyperbolic refinement. Benchmark tests and ablation experiments were carried out, and the experimental results were analyzed in depth to verify the effectiveness of the method and the contribution of each module.

[0072] In one embodiment, the experiment was conducted in the following environment: hardware consisted of one NVIDIA A800 GPU (80 GB) and an Intel Xeon CPU; software consisted of Ubuntu 22.04, Python 3.10, PyTorch 2.1, and Transformers 4.35.

[0073] The experiments included benchmarking and ablation experiments. Specifically, the benchmarking experiments used four publicly available multi-hop question-answering datasets: HotpotQA, 2WikiMultiHopQA, MuSiQue, and MultiHopRAG; the comparison methods included BeamRetrieval, GenGround, OPEN-RAG, and TRACE; and the evaluation metrics were EM (Exact Match) and F1.

[0074] Based on the complete method, the ablation experiments removed the following key modules for comparison: path exploration module, hyperbolic refinement module, and answer aggregation unit. The rest of the process remained unchanged, and the benchmark test experimental steps were repeated to obtain the EM and F1 scores for each ablation version.

[0075] Referring to Table 1, the experimental results are satisfactory. Specifically, our method outperforms the comparison methods on all datasets. On the MuSiQue dataset, EM and F1 scores reach 48.4% and 61.9%, respectively, representing improvements of approximately 6.7 and 6.6 percentage points compared to the OPEN-RAG method. On the 2WikiMultiHopQA dataset, EM scores improve from 51.6% to 62.6%, and F1 scores improve from 61.0% to 70.2%.

[0076] Table 1. Benchmark Test Results

[0077]

[0078] Reference Figure 4 and Figure 5 Based on the results of the ablation experiment, Figure 4 and Figure 5 It can be known that:

[0079] (1) After removing MCTS, EM decreased by about 8% and F1 decreased by about 6% on the HotpotQA dataset; EM decreased by about 12% and F1 decreased by about 10% on the MuSiQue dataset;

[0080] (2) After removing the hypersphere mapping, the EM on the 2WikiMultiHopQA dataset decreased by about 7%, and the F1 decreased by about 8%;

[0081] (3) After removing recursive aggregation, the EM on the MultiHopRAG dataset decreased by about 5%, and the F1 decreased by about 4%.

[0082] Experiments fully demonstrate the significant role of the multi-level path exploration, hyperbolic evidence refinement, and bottom-up recursive aggregation key technologies described in this invention in improving the performance of complex multi-hop question answering. Specifically, the path exploration strategy is a preferred solution; MCTS can balance exploration and utilization in the multi-hop space, effectively recover cross-segment logic, and avoid the propagation of erroneous paths. The hyperbolic refinement mechanism is also a preferred solution; hypersphere mapping combined with GAT propagation enhances hierarchical semantic modeling, giving highly relevant evidence higher ranking weights and improving retrieval accuracy. The recursive aggregation process is also indispensable; bottom-up aggregation maintains the contextual consistency of the answers to each sub-question, reduces information loss and conflicts, and achieves a more coherent final answer. All of the above demonstrate that the multi-hop RAG method and system based on path exploration and hyperbolic refinement have high versatility and robustness: this invention performs excellently in entity-related, wiki-like, complex combinatorial, and open-domain multi-hop scenarios, verifying the wide applicability of each module design.

[0083] In summary, this invention effectively solves the problems of rigid reasoning paths and evidence noise interference in the prior art by dynamically constructing reasoning trees and performing hyperbolic refinement on candidate evidence. The method effectively improves the coherence and accuracy of multi-hop reasoning and is suitable for natural language processing applications such as complex multi-hop question answering and knowledge reasoning.

[0084] Example 3

[0085] Based on the foregoing embodiments, this embodiment provides an electronic device, including: one or more processors and a memory, wherein the memory stores one or more programs, the one or more programs including instructions for executing the multi-hop RAG question answering method based on path exploration and hyperbolic refinement as described in Embodiment 1.

[0086] like Figure 6 At the hardware level, the electronic device includes a processor, internal bus, network interface, memory, and non-volatile memory, and may also include other hardware required for the business operations. The processor reads the corresponding computer program from the non-volatile memory into memory and then runs it to achieve the above-mentioned functions. Figure 1The method described herein. Of course, in addition to software implementation, this invention does not exclude other implementation methods, such as logic devices or a combination of hardware and software, etc. That is to say, the execution subject of the following processing flow is not limited to each logic unit, but can also be hardware or logic devices.

[0087] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0088] Example 4

[0089] This embodiment provides a computer-readable storage medium including one or more programs executable by one or more processors of an electronic device, the one or more programs including instructions for performing the multi-hop RAG question-answering method based on path exploration and hyperbolic refinement of Embodiment 1.

[0090] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0091] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A multi-hop RAG question-answering method based on path exploration and hyperbolic refinement, characterized in that, Includes the following steps: The system acquires multi-hop query input, constructs a multi-level reasoning tree based on a recursive path exploration search strategy, and obtains the optimal subtree through Monte Carlo tree search. The path exploration search strategy includes direct answering, answering after retrieval, and sub-question decomposition. During the path exploration search process, for the node that provides the answer after retrieval for the current query, multiple candidate evidences retrieved are mapped to hyperbolic space for hyperbolic refinement, and evidence with high relevance to the current query is selected. Combined with the current query, an intermediate answer is generated as the intermediate node of the reasoning tree. For each node in the optimal subtree, the final answer is obtained through recursive aggregation. The process of mapping the retrieved candidate evidence to hyperbolic space for hyperbolic refinement includes the following steps: An evidence graph structure is constructed based on multiple retrieved candidate evidence; Based on the aforementioned evidence graph structure, the embeddings of the query and noise-suppressed Euclidean space are mapped to hyperbolic space. Then, a graph attention network is used for graph propagation learning to propagate and update the information of the nodes in the evidence graph structure. The evidence is then sorted according to the relative characteristics of the node embedding vectors in the evidence graph structure. The hyperbolic space embedding map, after query and evidence ranking and noise suppression, is used to filter the multiple pieces of evidence with the highest relevance to the current query through similarity calculation. The inference tree includes nodes of the following types that alternate according to the tree's hierarchy: Operation status nodes include information about the path exploration and search strategies used; Text-based nodes include the current query, evidence, and answer information. The Monte Carlo tree search process includes the following steps: Selection: Based on the number of times a node is visited and its cumulative score, the optimal node N to be expanded is selected from the current tree using the upper confidence bound criterion; Exploration: Multiple path exploration search strategies are used for node N; Extension: Insert the new nodes obtained in the exploration steps and their intermediate answers into the reasoning tree; Backtracking: Update the statistics of the ancestral nodes based on the evaluation score of the new node to guide subsequent selections.

2. The multi-hop RAG question-answering method based on path exploration and hyperbolic refinement according to claim 1, characterized in that, The root node of the inference tree is a text-type node. After selecting a path exploration search strategy for the query in the text-type node, an operation state-type node is obtained. Based on the selected path exploration search strategy, the corresponding text-type child nodes are further obtained.

3. The multi-hop RAG question-answering method based on path exploration and hyperbolic refinement according to claim 1, characterized in that, During the recursive construction of the reasoning tree, the recursion stops when the preset tree height limit is reached, or when all subproblems are single-hop problems.

4. The multi-hop RAG question-answering method based on path exploration and hyperbolic refinement according to claim 1, characterized in that, The process of obtaining the query output through recursive aggregation includes: The nodes in the optimal subtree are recursively aggregated from bottom to top, and combined with the original query input, the final query output is generated by calling the language model. For each text-type node in the optimal subtree, the current query and the corresponding answer are processed through the Prompt project to allow the large language model to summarize the answer of the current query layer by layer, and finally aggregated into the answer of the initial query.

5. The multi-hop RAG question-answering method based on path exploration and hyperbolic refinement according to claim 1, characterized in that, The process of generating intermediate answers includes the following steps: Based on the filtered evidence that is highly relevant to the current query and the current query, an intermediate answer is generated using a language model.

6. The multi-hop RAG question-answering method based on path exploration and hyperbolic refinement according to claim 1, characterized in that, The path exploration search strategy also includes using a language model to directly answer simple or single-hop problems.

7. A multi-hop RAG question-answering system based on path exploration and hyperbolic refinement, characterized in that, For implementing the multi-hop RAG question answering method based on path exploration and hyperbolic refinement as described in any one of claims 1-6, the multi-hop RAG question answering system comprises: A multi-level path exploration module is used to obtain multi-hop query inputs, construct a multi-level reasoning tree based on a recursive path exploration search strategy, and obtain the optimal subtree through Monte Carlo tree search. The path exploration search strategy includes direct answering, answering after retrieval, and sub-question decomposition. The hyperbolic evidence refinement module is used during the path exploration and search process to map multiple candidate evidence retrieved to the hyperbolic space for hyperbolic refinement for nodes that provide answers to the current query. It filters out evidence that is highly relevant to the current query and generates intermediate answers based on the current query, which serve as intermediate nodes of the reasoning tree. The answer aggregation unit is used to recursively aggregate each node in the optimal subtree to obtain the final answer.

Images