Information gain guidance-based retrieval enhanced thinking tree reasoning method and system

Abstract

The application discloses a retrieval enhanced thinking tree reasoning method and system based on information gain guidance, relates to the technical field of artificial intelligence, and mainly comprises the following steps: reconstructing a traditional linear generation process into a structured tree search process, introducing information gain as the basis for path evaluation and selection, actively pruning low-value branches, continuously expanding high-value paths, and realizing the deep integration of the retrieval and reasoning processes. The retrieval enhanced thinking tree reasoning method and system based on information gain guidance can improve the reasoning quality and efficiency.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and more specifically, to a retrieval-enhanced thinking tree reasoning method and system based on information gain guidance. Background Technology

[0002] With the rapid development of large-scale pre-trained language models (LLMs), retrieval-augmented generation (RAG) has gradually become an important method for improving the accuracy and interpretability of model knowledge. RAG effectively alleviates the knowledge obsolescence and illusion problems caused by the reliance on parameter memorization in traditional generative models by introducing relevant information from external knowledge bases during the generation stage, and has been widely applied in various knowledge-intensive scenarios. Existing RAG methods typically include three stages: retrieval, augmentation, and generation. First, the user query is mapped to a vector representation, and a similarity search is performed in the knowledge base. Then, the retrieved relevant text is input as context into the language model. Finally, the model generates the answer. However, these methods generally adopt a linear generation strategy during the reasoning process, that is, the reasoning process unfolds step by step based on a single path. Although this approach is computationally efficient, it is easily affected by early decision-making errors when facing complex reasoning tasks (such as multi-hop question answering, logical inference, etc.). Once a deviation occurs in a certain step of reasoning or retrieval, it is difficult to correct in subsequent processes, leading to inaccurate overall reasoning results.

[0003] Tree of Thoughts (ToT) is a generative framework proposed to address this problem. It models the reasoning process as a tree-structured search problem. Within this framework, the model can simultaneously expand multiple candidate "thought nodes" at each reasoning step and compare and select between different paths through an evaluation mechanism, thereby enhancing the model's exploration capabilities in complex problem spaces. However, existing ToT methods still have shortcomings when combined with RAG: they lack a quantitative characterization of "information effectiveness," focusing more on the semantic relevance between retrieval results and queries while ignoring the information gain value of newly added information relative to existing context. This leads to the repeated introduction of low-value or redundant information during reasoning, increasing computational overhead and affecting the quality and stability of the final reasoning path.

[0004] Expanding the reasoning space, improving the ability to solve complex problems, enhancing the quality and efficiency of reasoning, and increasing the reliability of results are urgent issues that need to be addressed. Summary of the Invention

[0005] The purpose of this invention is to provide a retrieval enhancement mind tree reasoning method and system based on information gain guidance, which can improve the quality and efficiency of reasoning.

[0006] This invention provides a retrieval enhancement mind tree reasoning method based on information gain guidance, comprising the following steps: S1: Generate a global logical plan based on the user's question using a language model; S2: Based on the user question and the global logical plan, initialize the reasoning search tree and the candidate node set; S3: Based on the inference search tree and the candidate node set, using a language model and combining the constraints of global logical planning, historical inference paths, and the current evidence set, generate multiple candidate node queries that cover those for which key evidence has not yet been obtained. Perform normalization and semantic deduplication on the candidate node queries to obtain the processed candidate node queries. S4: Based on each candidate node query, a new set of evidence is obtained from an external knowledge base using a retrieval method. The new set of evidence is then integrated with the historical evidence of the parent node to form a complete set of evidence for the child node. At the same time, the reasoning semantic state of the child node is updated to the reasoning result after integrating the evidence, thereby forming a new child node. S5: Calculate the cumulative path value score of child nodes using the path value function; S6: Sort the nodes according to their cumulative path value scores, and retain only the top B nodes with the highest scores as the candidate node set for the next layer. S7: Confirm that the depth of the current candidate node in the candidate node set has reached the maximum depth, select the node with the highest score as the optimal reasoning path, integrate all the evidence on the optimal reasoning path, generate the final answer based on the integrated optimal reasoning path, and output the corresponding reasoning path as the interpretability basis of the result.

[0007] Furthermore, each node of the inference search tree corresponds to a local inference state, and the expression of the node is:

[0008] in, Let i be the i-th node in the reasoning search tree. This represents the inference semantic state of node i. This indicates the retrieval query executed by the current node i. Denotes the set of evidence introduced by node i. This represents the cumulative score of node i. This indicates the depth of node i in the reasoning tree.

[0009] Furthermore, the expression for the candidate node query is:

[0010] in, For nodes The set of candidate node queries; For nodes The Query candidate nodes; This indicates the number of candidate queries generated by each node.

[0011] Furthermore, the formula for calculating the cumulative score of the path value is as follows:

[0012] in, For nodes The cumulative score of path value; Represents a node The parent node; Given a user question q and an existing set of evidence and newly introduced sets of evidence The path value function.

[0013] Furthermore, the formula for calculating the path value function is as follows:

[0014]

[0015]

[0016]

[0017]

[0018]

[0019] in, Given the original problem q and the existing set of evidence and newly introduced sets of evidence The path value function; Semantic relevance is used to measure the newly introduced set of evidence. The degree of semantic matching between the original question q and the original question q; For the novelty of information, used to measure new evidence. Compared to existing evidence along the path Information gain; and , which is a weighting coefficient used to balance the contribution between relevance and novelty; For the j-th document in the set; For document Semantic relevance to the original question q; For document Compared to existing evidence Information novelty; This indicates that the cross-encoder reordering model will be used to process paired queries. and documents Perform deep, full-attention interaction scoring The value truncated to the interval [0,1]; To check new documents Whether the source entity has already provided evidence The metrics that have appeared in the text; To measure new documents using Jaccard similarity Compared to existing evidence Redundancy at the text level; As the weight of entity novelty; For existing evidence Lexical elements in; This represents the set of tokens obtained after segmenting a document.

[0020] This invention also provides a retrieval-enhanced thinking tree reasoning system based on information gain guidance, the system comprising the following modules: The global logic planning module is used to generate a global logic plan based on the user's problem using a language model. The initialization module is used to: initialize the inference search tree and the candidate node set according to the user question and the global logical plan; The candidate node query generation module is used to: generate multiple candidate node queries covering those for which key evidence has not yet been obtained, based on the inference search tree and the candidate node set, using a language model and in conjunction with constraints of global logical planning, historical inference paths, and the current evidence set; and perform normalization and semantic deduplication on the candidate node queries to obtain the processed candidate node queries. The child node update module is used to: retrieve a new set of evidence from an external knowledge base based on each candidate node query using a retrieval method, integrate the new set of evidence with the historical evidence of the parent node to form a complete set of evidence for the child node, and update the inference semantic state of the child node to the inference result after integrating the evidence, thereby forming a new child node; The path value scoring module is used to calculate the cumulative path value score of child nodes using the path value function. The bundle search module is used to sort the nodes according to their cumulative path value scores and retain only the top B nodes with the highest scores as the candidate node set for the next layer. The optimal reasoning path module is used to: confirm that the depth of the current candidate node in the candidate node set has reached the maximum depth, select the node with the highest score as the optimal reasoning path, integrate all the evidence on the optimal reasoning path, generate the final answer based on the integrated optimal reasoning path, and output the corresponding reasoning path as the interpretability basis of the result.

[0021] Furthermore, each node of the inference search tree corresponds to a local inference state, and the expression of the node is:

[0022] in, Let i be the i-th node in the reasoning search tree. This represents the inference semantic state of node i. This indicates the retrieval query executed by the current node i. Denotes the set of evidence introduced by node i. This represents the cumulative score of node i. This indicates the depth of node i in the reasoning tree.

[0023] Furthermore, the expression for the candidate node query is:

[0024] in, For nodes The set of candidate node queries; For nodes The Query candidate nodes; This indicates the number of candidate queries generated by each node.

[0025] Furthermore, the formula for calculating the cumulative score of the path value is as follows:

[0026] in, For nodes The cumulative score of path value; Represents a node The parent node; Given a user question q and an existing set of evidence and newly introduced sets of evidence The path value function.

[0027] Furthermore, the formula for calculating the path value function is as follows:

[0028]

[0029]

[0030]

[0031]

[0032]

[0033] in, Given the original problem q and the existing set of evidence and newly introduced sets of evidence The path value function; Semantic relevance is used to measure the newly introduced set of evidence. The degree of semantic matching between the original question q and the original question q; For the novelty of information, used to measure new evidence. Compared to existing evidence along the path Information gain; and , which is a weighting coefficient used to balance the contribution between relevance and novelty; For the j-th document in the set; For document Semantic relevance to the original question q; For document Compared to existing evidence Information novelty; This indicates that the cross-encoder reordering model will be used to process paired queries. and documents Perform deep, full-attention interaction scoring The value truncated to the interval [0,1]; To check new documents Whether the source entity has already provided evidence The metrics that have appeared in the text; To measure new documents using Jaccard similarity Compared to existing evidence Redundancy at the text level; As the weight of entity novelty; For existing evidence Lexical elements in; This represents the set of tokens obtained after segmenting a document.

[0034] The information gain-guided retrieval enhancement mind tree reasoning method and system provided by this invention have the following beneficial effects: This invention reconstructs the traditional linear generation process into a tree-structured search process, introducing a multi-branch expansion mechanism in each inference step. This allows the model to explore multiple candidate inference paths simultaneously, significantly expanding the coverage of the inference space. This method effectively avoids the problem of overall inference failure due to early decision-making errors in single-path inference, enhancing the model's robustness and solution capability in complex multi-hop tasks, expanding the inference space, and improving the ability to solve complex problems.

[0035] This invention introduces information gain as a path evaluation metric, comprehensively considering the relevance and novelty of retrieved information during path selection, and dynamically scoring and filtering candidate paths. This mechanism prioritizes high-value inference paths and prunes low-value branches, achieving information gain-driven path optimization, reducing the introduction of redundant information and ineffective computational overhead, thereby improving inference accuracy while enhancing the overall system's operational efficiency.

[0036] This invention introduces logical rule constraints to uniformly model state transitions and path evolution in multi-path reasoning, ensuring overall semantic consistency across branches during expansion. This method effectively alleviates path divergence and semantic offset issues, reduces the risk of error propagation, and improves the coherence and interpretability of the reasoning chain, thereby enhancing the stability and reliability of the final generated results.

[0037] In summary, the Information Gain-Guided Enhanced Retrieval Tree Reasoning (IG-RA-ToT) method proposed in this invention employs bundle search to prevent the search space from growing exponentially with depth, retaining only a fixed number of candidate paths with the highest scores. This method reconstructs the traditional linear generation process into a structured tree search process, introduces information gain as a crucial basis for path evaluation and selection, actively prunes low-value branches during reasoning, and continuously expands high-value paths. This achieves a deep integration of the retrieval and reasoning processes, expands the reasoning space, enhances the ability to solve complex problems, improves reasoning quality and efficiency, and increases the reliability of the results. Attached Figure Description

[0038] The present invention will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings: Figure 1 This is a flowchart of the information gain-guided retrieval enhancement mind tree reasoning method provided by the present invention; Figure 2 This is a schematic diagram of the reasoning steps of the information gain-guided retrieval enhancement mind tree reasoning method provided by the present invention; Figure 3 This is a schematic diagram comparing the improved method provided by this invention with the ablation variant that removes mind tree reasoning on single-hop and multi-hop datasets. Figure 4 This is a schematic diagram comparing the EM and F1 indices of the improved method provided by this invention with the baseline method on six datasets. Detailed Implementation

[0039] To provide a clearer understanding of the technical features, objectives, and effects of the present invention, specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0040] Figure 1 A schematic diagram of the information gain-guided retrieval-enhanced mind tree reasoning method of this embodiment is shown. In this embodiment, the information gain-guided retrieval-enhanced mind tree reasoning method includes the following steps: S1: Generate a global logical plan based on the user's question using a language model; S2: Based on the user question and the global logical plan, initialize the reasoning search tree and the candidate node set; In one exemplary embodiment, each node of the inference search tree corresponds to a local inference state, and the expression of the node is:

[0041] in, Let i be the i-th node in the reasoning search tree. This represents the inference semantic state of node i. This indicates the retrieval query executed by the current node i. Denotes the set of evidence introduced by node i. This represents the cumulative score of node i. This indicates the depth of node i in the reasoning tree.

[0042] As an exemplary embodiment, in step S2, the multi-hop reasoning process is represented as a dynamically expanding reasoning search tree T. Each node in the tree corresponds to a local reasoning state, and the nodes form reasoning paths through parent-child relationships. Each path from the root node to the leaf node corresponds to a candidate reasoning chain, thereby giving the multi-hop reasoning process an explicit structured representation.

[0043] S3: Based on the inference search tree and the candidate node set, using a language model and combining the constraints of global logical planning, historical inference paths, and the current evidence set, generate multiple candidate node queries that cover those for which key evidence has not yet been obtained. Perform normalization and semantic deduplication on the candidate node queries to obtain the processed candidate node queries. In one exemplary embodiment, the expression for the candidate node query is:

[0044] in, For nodes The set of candidate node queries; For nodes The Query candidate nodes; This indicates the number of candidate queries generated by each node.

[0045] As an exemplary embodiment, the normalization process involves removing stop words or synonym unification. The semantic deduplication method uses semantic vector similarity calculation to remove highly similar queries, thus avoiding redundant retrieval and path loops.

[0046] S4: Based on each candidate node query, a new set of evidence is obtained from an external knowledge base using a retrieval method. The new set of evidence is then integrated with the historical evidence of the parent node to form a complete set of evidence for the child node. At the same time, the reasoning semantic state of the child node is updated to the reasoning result after integrating the evidence, thereby forming a new child node.

[0047] S5: Calculate the cumulative path value score of child nodes using the path value function; In one exemplary embodiment, the formula for calculating the cumulative path value score is:

[0048] in, For nodes The cumulative score of path value; Represents a node The parent node; Given a user question q and an existing set of evidence and newly introduced sets of evidence The path value function; In one exemplary embodiment, the formula for calculating the path value function is:

[0049]

[0050]

[0051]

[0052]

[0053]

[0054] in, Given the original problem q and the existing set of evidence and newly introduced sets of evidence The path value function; Semantic relevance is used to measure the newly introduced set of evidence. The degree of semantic matching between the original question q and the original question q; For the novelty of information, used to measure new evidence. Compared to existing evidence along the path Information gain; and , which is a weighting coefficient used to balance the contribution between relevance and novelty; For the j-th document in the set; For document Semantic relevance to the original question q; For document Compared to existing evidence Information novelty; This indicates that the cross-encoder reordering model will be used to process paired queries. and documents Perform deep, full-attention interaction scoring The value truncated to the interval [0,1]; To check new documents Whether the source entity has already provided evidence The metrics that have appeared in the text; To measure new documents using Jaccard similarity Compared to existing evidence Redundancy at the text level; As the weight of entity novelty; For existing evidence Lexical elements in; This represents the set of tokens obtained after segmenting a document.

[0055] S6: Sort the nodes according to their cumulative path value scores, and retain only the top B nodes with the highest scores as the candidate node set for the next layer. S7: Confirm that the depth of the current candidate node in the candidate node set has reached the maximum depth, select the node with the highest score as the optimal reasoning path, integrate all the evidence on the optimal reasoning path, generate the final answer based on the integrated optimal reasoning path, and output the corresponding reasoning path as the interpretability basis of the result.

[0056] This embodiment provides a retrieval-enhanced thinking tree reasoning system based on information gain guidance. The system includes the following modules: The global logic planning module is used to generate a global logic plan based on the user's problem using a language model. The initialization module is used to: initialize the inference search tree and the candidate node set according to the user question and the global logical plan; The candidate node query generation module is used to: generate multiple candidate node queries covering those for which key evidence has not yet been obtained, based on the inference search tree and the candidate node set, using a language model and in conjunction with constraints of global logical planning, historical inference paths, and the current evidence set; and perform normalization and semantic deduplication on the candidate node queries to obtain the processed candidate node queries. The child node update module is used to: retrieve a new set of evidence from an external knowledge base based on each candidate node query using a retrieval method, integrate the new set of evidence with the historical evidence of the parent node to form a complete set of evidence for the child node, and update the inference semantic state of the child node to the inference result after integrating the evidence, thereby forming a new child node; The path value scoring module is used to calculate the cumulative path value score of child nodes using the path value function. The bundle search module is used to sort the nodes according to their cumulative path value scores and retain only the top B nodes with the highest scores as the candidate node set for the next layer. The optimal reasoning path module is used to: confirm that the depth of the current candidate node in the candidate node set has reached the maximum depth, select the node with the highest score as the optimal reasoning path, integrate all the evidence on the optimal reasoning path, generate the final answer based on the integrated optimal reasoning path, and output the corresponding reasoning path as the interpretability basis of the result.

[0057] Furthermore, each node of the inference search tree corresponds to a local inference state, and the expression of the node is:

[0058] in, Let i be the i-th node in the reasoning search tree. This represents the inference semantic state of node i. This indicates the retrieval query executed by the current node i. Denotes the set of evidence introduced by node i. This represents the cumulative score of node i. This indicates the depth of node i in the reasoning tree.

[0059] Furthermore, the expression for the candidate node query is:

[0060] in, For nodes The set of candidate node queries; For nodes The Query candidate nodes; This indicates the number of candidate queries generated by each node.

[0061] Furthermore, the formula for calculating the cumulative score of the path value is as follows:

[0062] in, For nodes The cumulative score of path value; Represents a node The parent node; Given a user question q and an existing set of evidence and newly introduced sets of evidence The path value function.

[0063] Furthermore, the formula for calculating the path value function is as follows:

[0064]

[0065]

[0066]

[0067]

[0068]

[0069] in, Given the original problem q and the existing set of evidence and newly introduced sets of evidence The path value function; Semantic relevance is used to measure the newly introduced set of evidence. The degree of semantic matching between the original question q and the original question q; For the novelty of information, used to measure new evidence. Compared to existing evidence along the path Information gain; and , which is a weighting coefficient used to balance the contribution between relevance and novelty; For the j-th document in the set; For document Semantic relevance to the original question q; For document Compared to existing evidence Information novelty; This indicates that the cross-encoder reordering model will be used to process paired queries. and documents Perform deep, full-attention interaction scoring The value truncated to the interval [0,1]; To check new documents Whether the source entity has already provided evidence The metrics that have appeared in the text; To measure new documents using Jaccard similarity Compared to existing evidence Redundancy at the text level; As the weight of entity novelty; For existing evidence Lexical elements in; This represents the set of tokens obtained after segmenting a document.

[0070] In some embodiments, the above-described information gain-guided retrieval-enhanced mind tree reasoning method can also be implemented in the following ways.

[0071] like Figure 2 As shown, in this embodiment, the information gain-guided retrieval-enhanced mind tree reasoning method includes the following steps: 1. At the beginning of the reasoning stage, a global logical program π is generated first to characterize the potential reasoning structure and entity dependencies of the problem. The logical program is not only used to guide the generation of single-step retrieval queries, but also serves as a global constraint in the tree search process to keep the exploration direction of different branches consistent.

[0072] 2. Represent the multi-hop reasoning process as a dynamically expanding reasoning search tree T, where each node corresponds to a local reasoning state, formally defined as:

[0073] in, Let i be the i-th node in the reasoning tree. Represents the inferential semantic state of a node. This indicates the retrieval query being executed by the current node. This represents the set of evidence introduced by the node. This represents the cumulative score of the node. This represents the depth of a node in the reasoning tree. Nodes are connected by parent-child relationships to form reasoning paths. Each path from the root node to a leaf node corresponds to a candidate reasoning chain, thus providing an explicit, structured representation of the multi-hop reasoning process.

[0074] 3. During the mind tree reasoning process, the algorithm maintains a set of candidate nodes and expands it at each depth. Let the set of candidate nodes at the current level be... Then at depth t+1, for each node Generate multiple candidate node queries:

[0075] in For nodes The candidate query set is given by k, where k represents the number of candidate queries generated by each node. The generation of candidate queries is accomplished by the language model and constrained by logical programming π, historical reasoning paths, and the current evidence set to ensure that the queries can cover key evidence that has not yet been obtained.

[0076] For each candidate query The algorithm uses a retrieval module to obtain new sets of evidence from external knowledge bases. This data is then integrated with historical evidence along the path to form new child nodes. Simultaneously, to avoid redundant retrieval and path loops, the generated queries undergo normalization and semantic deduplication to prevent highly similar queries from being executed repeatedly.

[0077] 4. To evaluate the value of different reasoning paths, this invention constructs an information gain-inspired path value function: given the original question q and the existing evidence set... and newly introduced sets of evidence The value function is defined as follows:

[0078] in and These are weighting coefficients used to balance the contributions of relevance and novelty. The specific calculation methods for each of the two components are given below.

[0079] (1) Semantic relevance

[0080] This sub-item measures the newly introduced set of evidence. The degree of semantic matching between the document and the original question q. For each document in the set. The system invokes the Cross-Encoder re-ranking model to perform deep, full-attention interactive scoring on the query-document pairs:

[0081] Here, `clamp(·,0,1)` truncates the original scores to the [0,1] interval, and the relevance of the entire dataset is taken as the average score of all documents.

[0082] (2) Information novelty

[0083] This sub-item measures new evidence relative to existing evidence along the path. Information gain, this component encourages the model to explore new entities and factual information. For each new document Novelty is a weighted fusion of two dimensions: entity novelty and content novelty.

[0084] in As the weight of entity novelty, Check whether the source entity of the new document has appeared in existing evidence—if it is a completely new entity, assign a value of 1.0; if it is a duplicate entity, assign a lower base score of 0.3 to encourage the reasoning process to cover knowledge from more different sources. The redundancy at the text level is measured using Jaccard similarity:

[0085] in This represents the set of lexical units obtained after word segmentation of the document. The formula calculates the maximum Jaccard similarity between the new document and all existing documents, and takes its complement as the novelty score. The less lexical overlap between the new document and existing evidence, the higher the novelty. The novelty of the entire evidence set is also taken as the mean.

[0086] Based on information gain, the cumulative score of a node can be recursively defined as:

[0087] in Represents a node The parent node is defined such that the node score expresses the information gain of the entire reasoning path, thus providing an evaluation criterion for path selection.

[0088] 5. Based on the bundle search strategy, all newly generated nodes are first sorted according to the path value function, and only the top B nodes with the highest scores are retained as the candidate set for the next layer:

[0089] 6. During the inference termination phase, the algorithm selects the node with the highest score from the final candidate node set. This serves as the optimal reasoning path, and all evidence along this path is integrated:

[0090] in, Indicates from the root node to The set of all nodes on the path, This represents the evidence introduced by node v. The algorithm is then based on... Generate the final answer and output the corresponding reasoning path as the basis for interpretability of the result.

[0091] This embodiment provides a detailed experimental evaluation process.

[0092] 1. Evaluation Indicators (1) Classification performance indicators 1) Accuracy Accuracy measures the proportion of correct predictions made by a classifier across the entire dataset, directly reflecting its overall discriminative power. It is defined as:

[0093] in (·) is an indicator function that takes the value 1 when the condition is true and 0 otherwise.

[0094] 2) Macro average F1 score Because the distribution of problems of varying complexity is unbalanced in real-world scenarios, relying solely on overall accuracy is insufficient to fully characterize the performance of long-tail classes. Therefore, the Macro-F1 metric is introduced. For any class c∈C, its accuracy is... With recall rate Defined as:

[0095] in , and These represent the number of true positive, false positive, and false negative samples in this category, respectively. The corresponding category F1 score and the macro average of all categories are defined as follows: Macro-F1 assigns the same weights to the minority class and the majority class, which can more objectively measure the classifier's decision balance across classes of different complexities.

[0096] (2) Question and answer generation quality indicators 1) Precise Match Rate (EM) EM measures whether the predicted answer and the standard answer are completely consistent after standardization. A single-sample EM and its overall score on the test set are defined as follows:

[0097] Normalize(·) represents the normalization operation, which includes preprocessing steps such as lowercase conversion, removal of punctuation marks and extra spaces.

[0098] 2) F1 Score This metric is calculated based on the word-level overlap between the predicted and reference answers. The standardized predicted and reference answers are each segmented into word sets. and Its single-sample F1 score and overall score are defined as follows:

[0099] The F1 score is more lenient than the EM score, and can partially measure the local similarity of answers.

[0100] 3) Includes match accuracy (Acc) Considering that large language models tend to output complete natural language sentences rather than isolated entities, directly using EM and F1 metrics may penalize semantically correct results that differ in expression, thus failing to fully reflect the actual correctness of the generated results. Therefore, we introduce matching accuracy based on substring inclusion relationships as a metric:

[0101] in This indicates the inclusion relationship at the string level, with ∨ representing a logical OR operation. This metric effectively mitigates the misjudgment problem caused by the diverse forms of model expression, and more objectively characterizes the core factual accuracy of the generated results.

[0102] (3) System retrieval overhead indicators Average number of searches ( ): The average number of retrieval calls per query is used as a proxy metric for computational cost. Let the number of retrieval calls for the i-th sample be... The zero-search strategy corresponds to =0, single-hop retrieval strategy corresponds to =1, corresponding to multi-hop reasoning strategy This equals the actual number of iterative retrieval rounds performed. The average number of retrievals is defined as:

[0103] By combining this indicator with question-and-answer quality metrics, the specific benefits of improving methods in terms of quality-cost trade-offs can be intuitively quantified.

[0104] 2. Dataset To evaluate the performance of the improved method under different problem complexities, this invention selected six representative open-domain question answering datasets as benchmark sets. The selected datasets cover a variety of task types, from single-fact retrieval to multi-hop reasoning, and can well reflect the adaptability of the method under different knowledge dependency structures.

[0105] Based on the required depth of reasoning for each question, this invention categorizes the evaluation dataset into single-hop and multi-hop tasks. For single-hop fact-based tasks, SQuADv2.0, Natural Questions (NQ), and TriviaQA were selected. SQuADv2.0 primarily assesses the model's ability to accurately extract text fragments; NQ originates from real user search queries, exhibiting strong linguistic diversity and open-domain features; TriviaQA includes more complex entity relationships and long text contexts. These datasets are mainly used to evaluate the retrieval and generation capabilities of methods in single-retrieval scenarios.

[0106] In multi-hop reasoning tasks, this invention selected HotpotQA, 2WikiMultiHopQA, and Musique. HotpotQA requires the model to perform cross-document reasoning by bridging entities; 2WikiMultiHopQA involves attribute comparison and information aggregation of multiple entities; and Musique emphasizes multi-step combined reasoning. These datasets constitute an evaluation scenario for multi-hop reasoning strategies, used to analyze the model's planning and execution capabilities under complex knowledge-dependent conditions. Table 1 summarizes the basic characteristics and focuses of each dataset.

[0107] Table 1: Basic characteristics and focus of each dataset

[0108] To ensure comparability between different datasets, this invention standardizes the datasets: stratified sampling is performed from the official development sets of each dataset to construct evaluation subsets with consistent sample sizes, and the question text, answer format, and retrieval corpus structure are standardized. Each sample includes a unique identifier, question text, standard answer object, and context list to eliminate evaluation barriers caused by heterogeneous formats from different data sources.

[0109] 3. Method Testing This invention uniformly adopts the RAG configuration using a linear multi-hop strategy as the baseline, which can serve as a comparison starting point with complete structure and stable performance.

[0110] To examine the impact of the optimization algorithm on system performance, this invention employs ablation experiments. In the experiments, the system integrating the optimization module is used as a reference; the optimization algorithm is then removed, and the corresponding modules are replaced with the baseline algorithm implementation to observe performance changes. The actual contribution of this invention is further quantified by combining ablation experiments with comparative experiments.

[0111] 4. Experimental Results (1) Performance gains in single-hop and multi-hop scenarios Table 2: Comparison of exact matching rates of different methods on single-hop and multi-hop datasets

[0112] Table 2 shows a comparison of the exact matching rates of the improved method of this invention and the ablation variant that removes mind tree reasoning on single-hop and multi-hop datasets, and marks the absolute reduction on each dataset.

[0113] Figure 3 The improved method and the ablation variant that removes mind tree inference are compared on single-hop and multi-hop datasets. It can be seen that after removing mind tree inference, the average EM of the multi-hop dataset drops sharply, with 2WikiMultiHopQA showing the largest decrease. 13.8%), followed by HotpotQA ( (12.7%), while the single-hop dataset, although showing a smaller decrease, also showed a decline.

[0114] (2) End-to-end question-answering performance evaluation Figure 4 The left and right subplots in the figure show the EM and F1 scores of the improved method and the baseline method on six datasets, respectively. The improved method outperforms the baseline method on most datasets. In terms of macro-average metrics, the improved method improves the EM by approximately 9.41% and the F1 score by approximately 8.81% compared to the baseline.

[0115] Table 3: Detailed performance comparison of baseline and improved methods on various datasets

[0116] Table 3 presents a detailed performance comparison of the baseline and improved methods on various datasets. The performance improvement is more significant on multi-hop question-answering datasets. In tasks involving complex logic, the improved method achieves an average EM improvement of approximately 14.5%. These results demonstrate that the introduction of the mind tree reasoning algorithm enhances the stability of the method in long-chain reasoning scenarios, mitigating the potential logical interruptions and error accumulation issues that may occur in the baseline during multi-step reasoning.

[0117] (3) Analysis of performance gain and efficiency changes Table 4: Average EM and Average Number of Searches for the Baseline System and the Optimized System on Each Dataset

[0118] Table 4 presents the average EM and average number of searches for the baseline and optimized systems on various datasets. From an overall quality-cost efficiency perspective, the baseline system's efficiency ratio is... =19.46 / 3.64=5.35, the optimized system is as follows: =28.87 / 4.03=7.16, the unit retrieval cost efficiency of the optimized system improved by about 34%. This means that the overall number of searches only increased by about 11%, while the EM improved by 48%, and the increase in EM far exceeded the increase in retrieval cost, fully demonstrating the superiority of the proposed mechanism in the quality-cost trade-off.

[0119] (4) Comparative experiment Table 5: Cross-sectional F1 scores using multiple methods

[0120] The improved method was compared with various methods evaluated using the Llama3-8B-Instruct model on the FlashRAG benchmark platform in terms of F1 scores. As shown in Table 5, the improved method comprehensively outperformed all reference methods on both multi-hop datasets. On HotpotQA, the improved method achieved an F1 score of 54.4%, which is 12.9% higher than the previous best IRCoT; on 2WikiMultiHopQA, the improved method also outperformed all reference methods. Furthermore, the baseline method was basically on par with IRCoT on HotpotQA, but after the optimization algorithm of this invention, the performance achieved a leap from "on par" to "significantly outperformed", verifying the collaborative effectiveness of the proposed algorithm in multi-hop inference scenarios.

[0121] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A retrieval-enhanced mind tree reasoning method based on information gain guidance, characterized in that, Includes the following steps: S1: Generate a global logical plan based on the user's question using a language model; S2: Based on the user question and the global logical plan, initialize the reasoning search tree and the candidate node set; S3: Based on the inference search tree and the candidate node set, using a language model and combining the constraints of global logical planning, historical inference paths, and the current evidence set, generate multiple candidate node queries that cover those for which key evidence has not yet been obtained. Perform normalization and semantic deduplication on the candidate node queries to obtain the processed candidate node queries. S4: Based on each candidate node query, a new set of evidence is obtained from an external knowledge base using a retrieval method. The new set of evidence is then integrated with the historical evidence of the parent node to form a complete set of evidence for the child node. At the same time, the reasoning semantic state of the child node is updated to the reasoning result after integrating the evidence, thereby forming a new child node. S5: Calculate the cumulative path value score of child nodes using the path value function; S6: Sort the nodes according to their cumulative path value scores, and retain only the top B nodes with the highest scores as the candidate node set for the next layer. S7: Confirm that the depth of the current candidate node in the candidate node set has reached the maximum depth, select the node with the highest score as the optimal reasoning path, integrate all the evidence on the optimal reasoning path, generate the final answer based on the integrated optimal reasoning path, and output the corresponding reasoning path as the interpretability basis of the result.

2. The method according to claim 1, characterized in that, Each node in the inference search tree corresponds to a local inference state, and the expression of the node is: in, Let i be the i-th node in the reasoning search tree. This represents the inference semantic state of node i. This indicates the retrieval query executed by the current node i. Denotes the set of evidence introduced by node i. This represents the cumulative score of node i. This indicates the depth of node i in the reasoning tree.

3. The method according to claim 1, characterized in that, The expression for querying the candidate node is: in, For nodes The set of candidate node queries; For nodes The Query candidate nodes; This indicates the number of candidate queries generated by each node.

4. The method according to claim 1, characterized in that, The formula for calculating the cumulative score of the path value is as follows: in, For nodes The cumulative score of path value; Represents a node The parent node; Given a user question q and an existing set of evidence and newly introduced sets of evidence The path value function.

5. The method according to claim 1, characterized in that, The formula for calculating the path value function is as follows: in, Given the original question q and the existing set of evidence and newly introduced sets of evidence The path value function; Semantic relevance is used to measure the newly introduced set of evidence. The degree of semantic matching between the original question q and the original question q; For the novelty of information, used to measure new evidence. Compared to existing evidence along the path Information gain; and , which is a weighting coefficient used to balance the contribution between relevance and novelty; For the j-th document in the set; For document Semantic relevance to the original question q; For document Compared to existing evidence Information novelty; This indicates that the cross-encoder reordering model will be used to process paired queries. and documents Perform deep, full-attention interaction scoring The value truncated to the interval [0,1]; To check new documents Whether the source entity has already provided evidence The metrics that have appeared in the text; To measure new documents using Jaccard similarity Compared to existing evidence Redundancy at the text level; As the weight of entity novelty; For existing evidence Lexical elements in; This represents the set of tokens obtained after segmenting a document.

6. A retrieval-enhanced thinking tree reasoning system based on information gain guidance, characterized in that, The system includes the following modules: The global logic planning module is used to generate a global logic plan based on the user's problem using a language model. The initialization module is used to: initialize the inference search tree and the candidate node set according to the user question and the global logical plan; The candidate node query generation module is used to: generate multiple candidate node queries covering those for which key evidence has not yet been obtained, based on the inference search tree and the candidate node set, using a language model and in conjunction with constraints of global logical planning, historical inference paths, and the current evidence set; and perform normalization and semantic deduplication on the candidate node queries to obtain the processed candidate node queries. The child node update module is used to: retrieve a new set of evidence from an external knowledge base based on each candidate node query using a retrieval method, integrate the new set of evidence with the historical evidence of the parent node to form a complete set of evidence for the child node, and update the inference semantic state of the child node to the inference result after integrating the evidence, thereby forming a new child node; The path value scoring module is used to calculate the cumulative path value score of child nodes using the path value function. The bundle search module is used to sort the nodes according to their cumulative path value scores and retain only the top B nodes with the highest scores as the candidate node set for the next layer. The optimal reasoning path module is used to: confirm that the depth of the current candidate node in the candidate node set has reached the maximum depth, select the node with the highest score as the optimal reasoning path, integrate all the evidence on the optimal reasoning path, generate the final answer based on the integrated optimal reasoning path, and output the corresponding reasoning path as the interpretability basis of the result.

7. The system according to claim 6, characterized in that, Each node in the inference search tree corresponds to a local inference state, and the expression of the node is: in, Let i be the i-th node in the reasoning search tree. This represents the inference semantic state of node i. This indicates the retrieval query executed by the current node i. Denotes the set of evidence introduced by node i. This represents the cumulative score of node i. This indicates the depth of node i in the reasoning tree.

8. The system according to claim 6, characterized in that, The expression for querying the candidate node is: in, For nodes The set of candidate node queries; For nodes The Query candidate nodes; This indicates the number of candidate queries generated by each node.

9. The system according to claim 6, characterized in that, The formula for calculating the cumulative score of the path value is as follows: in, For nodes The cumulative score of path value; Represents a node The parent node; Given a user question q and an existing set of evidence and newly introduced sets of evidence The path value function.

10. The system according to claim 6, characterized in that, The formula for calculating the path value function is as follows: in, Given the original question q and the existing set of evidence and newly introduced sets of evidence The path value function; Semantic relevance is used to measure the newly introduced set of evidence. The degree of semantic matching between the original question q and the original question q; For the novelty of information, used to measure new evidence. Compared to existing evidence along the path Information gain; and , which is a weighting coefficient used to balance the contribution between relevance and novelty; For the j-th document in the set; For document Semantic relevance to the original question q; For document Compared to existing evidence Information novelty; This indicates that the cross-encoder reordering model will be used to process paired queries. and documents Perform deep, full-attention interaction scoring The value truncated to the interval [0,1]; To check new documents Whether the source entity has already provided evidence The metrics that have appeared in the text; To measure new documents using Jaccard similarity Compared to existing evidence Redundancy at the text level; As the weight of entity novelty; For existing evidence Lexical elements in; This represents the set of tokens obtained after segmenting a document.

Images