A knowledge graph question answering reasoning method based on potential energy increment triggering

Abstract

The application provides a knowledge graph question and answer reasoning method based on potential energy increment triggering, which comprises the following steps: S1, obtaining a user question, constructing a neighborhood subgraph according to the user question, generating an initial candidate operation set based on the neighborhood subgraph, and screening a candidate operation set in the initial candidate operation set according to the executability judgment and the semantic similarity ranking; S2, inputting the user question and the candidate operation set into a large language model, and outputting the selected operation by the large language model; S3, calling a retrieval execution unit, executing the selected operation output by the large language model, and outputting the corresponding entity; S4, obtaining the subject entity quality and the virtual entity quality according to the entity, extracting the user question, and calculating the semantic distance; S5, calculating the potential energy of the current step, and calculating the potential energy increment, and performing a rollback operation when the potential energy increment exceeds a preset threshold value, until the final answer is output.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically a knowledge graph question-answering reasoning method based on potential energy increment triggering. Background Technology

[0002] Large-scale language models are used for temporal knowledge graph question answering, relying on their reasoning capabilities to retrieve answers from the graph. Temporal knowledge graphs explicitly model temporal evolution relationships, providing the knowledge and constraints required for temporal reasoning. Ideally, the model will follow the reasoning chain to arrive at a temporally consistent answer. However, in existing technologies, there is a bottleneck when models handle temporal problems due to breaks in the temporal reasoning chain. For example, CN120596645A proposes a data-intelligent question answering method and system based on retrieval-enhanced generation. Although it achieves high accuracy in knowledge reasoning and question answering, complex question handling capabilities, and answer generation quality, it is largely result-oriented and lacks a real-time monitoring mechanism for the reasoning process, making it unable to effectively remedy broken reasoning chains. However, existing technologies have the following problems when handling temporal knowledge graph question answering tasks: Existing technologies for handling temporal knowledge graph question answering tasks have two main problems: First, the real-time detection capability for reasoning breaks is insufficient. In complex temporal reasoning, the model lacks interpretable evaluation criteria, cannot monitor the reasoning state in real time, and is difficult to detect reasoning chain breaks in a timely manner, increasing the difficulty of subsequent error correction.

[0003] Second, there is a lack of a dynamic correction mechanism for the reasoning chain. After identifying a break or error in the reasoning, existing technologies struggle to proactively halt erroneous decision-making paths and backtrack to correct them. They cannot utilize error feedback to explore the correct path, thus affecting the robustness and reliability of the reasoning.

[0004] Early temporal knowledge graph question answering methods relied on rule decomposition or knowledge graph embedding techniques. Later, dedicated temporal modules improved performance, but the inference paths were static, lacking real-time monitoring and dynamic correction capabilities; errors in intermediate steps were difficult to recover from. Recent research utilizes the inference capabilities of large language models combined with retrieval tools, but inaccurate temporal understanding leads to retrieval noise. Some methods add separate temporal processing to improve sensitivity, but they remain result-oriented, lacking dynamic verification mechanisms and susceptible to cascading errors.

[0005] To enhance reasoning performance, existing technologies explore the self-correction capabilities of large language models, such as prompting self-correction or introducing evaluators. However, correction decisions rely on the model's own subjective evaluation, lack external objective standards, have poor interpretability of evaluation results, and are difficult to guarantee the accuracy of correction direction. Summary of the Invention

[0006] The purpose of this invention is to address the shortcomings of existing technologies by providing a knowledge graph question-answering reasoning method based on potential energy increment triggering.

[0007] This invention is achieved through the following technical solution: A knowledge graph question-answering reasoning method based on potential energy increment triggering includes: S1. Obtain the user's question, construct a neighborhood subgraph based on the user's question, generate an initial candidate operation set based on the neighborhood subgraph, and select several candidate operation sets within the initial candidate operation set based on executability judgment and semantic similarity ranking. S2. Input the user question and the set of candidate operations into a large language model, and the large language model outputs the selected operation; S3. Call the retrieval execution unit to perform the selected operation output by the large language model, so as to obtain the corresponding entity; S4. Obtain the subject entity quality and virtual entity quality based on the entities, extract the user question, and calculate the semantic distance; S5. Calculate the potential energy at the current step. And calculate the potential energy increment. When the potential energy increases Exceeding the preset threshold A rollback operation is performed until the final answer is output.

[0008] Furthermore, in step S1, in obtaining user questions... Next, first obtain the subject entity in the question. Then, search the underlying time-series knowledge graph. neighborhood subgraph ,in, Indicates a relationship. For the target entity, For this subject entity Valid time interval.

[0009] Furthermore, in step S1, Executability assessment includes: assessing the initial candidate operation set. In the current subgraph Whether it is legal or not, record it as ; Semantic relevance ranking includes: calculating the initial candidate operation set. User issues The semantic similarity is used to select the top N high scores based on semantic similarity.

[0010] Furthermore, in step S1, a candidate operation set is generated based on a preset operation template. The candidate operation template includes an entity-oriented operation set and a time-oriented operation set. The entity-oriented operation is used to obtain the target entity or relationship-related entity from the neighborhood subgraph, and the time-oriented operation is used to obtain the timestamp or time interval associated with the entity-relationship fact.

[0011] Furthermore, step S2 also includes historical reasoning information, and outputs the operation of the current step according to the following model: ; in, For the operation of the current step, For historical reasoning information, For user issues, For the candidate operation set, Large-scale language models.

[0012] Furthermore, between steps S3 and S4, the following steps are also included: The virtual entity particle is obtained by averaging the embedding vectors of multiple resulting entities. The virtual particle is represented as: ; in, It is a virtual particle embedding used for subsequent energy calculations. This represents the set of entities retrieved in the k-th step.

[0013] Furthermore, in step S4, semantic distance is defined using the embedding representation provided by TComplEx for time and the embedding representation provided by CronKGQA for the question. : ; in, and These represent the embeddings of the subject entity and the current virtual entity, respectively. It is a time embedding from TComplEx. It's an embedding from CronKGQA. Indicates respectively to Vectorize, This indicates the vectorized version. Perform a dot product operation and take the real part of the result; It is a small positive constant.

[0014] Furthermore, in step S5, the step potential energy Including content potential With time potential energy ,when At that time, take ;when At that time, take ; ; ; in, and These are question-guided query embeddings used for entity reasoning and time reasoning, respectively.

[0015] Furthermore, content maintains momentum. With time potential energy All are based on embedded interactive computation, where the embeddings include: topic entity embedding, virtual entity embedding, temporal embedding, and query embedding obtained by question guidance; and the similarity term is calculated through the real part of the multilinear inner product.

[0016] A knowledge graph question-answering reasoning system based on potential energy increment triggering includes: The candidate operation generation unit is used to generate a set of candidate operations based on the user question and the neighborhood subgraph, and then filter the candidate operation set. The decision unit is used to input the user's question, historical reasoning trajectory, and candidate operation set into a large language model to output the selected operation; The retrieval execution unit is used to perform the selected operation on the time-series knowledge graph and output intermediate results; The potential energy discrimination unit is used to calculate the potential energy in the current step and output the potential energy increment. The re-decision unit is used to roll back the current step when the potential energy increment exceeds a threshold, and inject the failed operation for re-decision.

[0017] The advantages and beneficial effects of this invention are as follows: By constructing a neighborhood subgraph and filtering candidate operation sets, the scope of reasoning is significantly reduced, invalid computation is decreased, and the efficiency of knowledge graph question-answering reasoning is improved; by combining large-scale language models to make operation decisions, the semantic understanding ability of the model is fully utilized to ensure the rationality and semantic fit of operation selection; by introducing potential energy and potential energy increment calculation, an objective quantitative evaluation standard for the reasoning process is established to realize real-time detection of reasoning chain breakage; when the potential energy increment exceeds the threshold, a rollback operation is performed to form a dynamic correction mechanism for the reasoning chain, timely termination of erroneous reasoning paths, avoidance of cascading errors, and significant improvement of the robustness and accuracy of question-answering reasoning. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart of the knowledge graph question-answering reasoning method based on potential energy increment triggering of the present invention.

[0020] Figure 2This is a model framework diagram of the knowledge graph question-answering reasoning method based on potential energy increment triggering of the present invention. Detailed Implementation

[0021] To enable those skilled in the art to better understand the present invention, the technical solution of the present invention will be further described below with reference to specific embodiments.

[0022] This embodiment provides a knowledge graph question-answering reasoning method based on potential energy increment triggering, such as... Figure 1 As shown, it includes: S1. Obtain the user's question, construct a neighborhood subgraph based on the user's question, generate an initial candidate operation set based on the neighborhood subgraph, and select several candidate operation sets within the initial candidate operation set based on executability judgment and semantic similarity ranking. S2. Input the user question and the set of candidate operations into a large language model, and the large language model outputs the selected operation; S3. Call the retrieval execution unit to perform the selected operation output by the large language model, so as to obtain the corresponding entity; S4. Obtain the subject entity quality and virtual entity quality based on the entities, extract the user question, and calculate the semantic distance; S5. Calculate the potential energy at the current step. And calculate the potential energy increment. When the potential energy increases Exceeding the preset threshold A rollback operation is performed until the final answer is output.

[0023] By constructing neighborhood subgraphs and filtering candidate operation sets, the scope of reasoning is significantly narrowed, invalid computations are reduced, and the efficiency of knowledge graph question-answering reasoning is improved. Operation decisions are made in conjunction with large-scale language models, fully utilizing the model's semantic understanding capabilities to ensure the rationality and semantic fit of operation selections. Potential energy and its increment are introduced to establish objective quantitative evaluation criteria for the reasoning process, enabling real-time detection of reasoning chain breaks. When the potential energy increment exceeds a threshold, a rollback operation is executed, forming a dynamic correction mechanism for the reasoning chain. This promptly terminates erroneous reasoning paths, avoids cascading errors, and significantly improves the robustness and accuracy of question-answering reasoning.

[0024] Furthermore, in step S1, in obtaining user questions... Next, first obtain the subject entity in the question. Then, search the underlying time-series knowledge graph. neighborhood subgraph ,in, Indicates a relationship. For the target entity, For the subject entity Valid time interval.

[0025] By clarifying the construction logic of the neighborhood subgraph and the plural composition of the temporal knowledge graph, we can accurately extract temporal knowledge related to the user's question subject entity, providing a precise and focused knowledge base for subsequent candidate operation generation and reasoning, avoiding interference from irrelevant graph information, and improving the pertinence and effectiveness of reasoning.

[0026] Furthermore, in step S1, Executability assessment includes: assessing the initial candidate operation set. In the current subgraph Whether it is legal or not, record it as ; Semantic relevance ranking includes: calculating the initial candidate operation set. User issues The semantic similarity is used to select the top N high scores based on semantic similarity.

[0027] The executability judgment filters invalid candidate operations from a legality perspective, excluding operations that cannot be implemented in the current subgraph, reducing invalid options for subsequent decisions and improving decision-making efficiency; semantic similarity ranking selects operations that are highly relevant to the user's question, ensuring the semantic matching degree between the candidate operation set and the question, reducing the probability of large language models selecting incorrect operations, and improving the accuracy of operation decisions.

[0028] Furthermore, such as Figure 1 and Figure 2 As shown, in step S1, a candidate operation set is generated according to a preset operation template. The candidate operation template includes an entity-oriented operation set and a time-oriented operation set. The entity-oriented operation is used to obtain the target entity or relationship-related entity from the neighborhood subgraph, and the time-oriented operation is used to obtain the timestamp or time interval associated with the entity-relationship fact.

[0029] Candidate operations are divided into two categories: entity-oriented and time-oriented. This aligns with the dual core features of entity and time in temporal knowledge graphs, enabling targeted retrieval and reasoning of entity and time information in the graph. This addresses the shortcomings of existing technologies in focusing on time-dimensional reasoning and improves the completeness and accuracy of time-based reasoning in temporal knowledge graph question answering.

[0030] Furthermore, such as Figure 1 and Figure 2 As shown, step S2 also includes historical reasoning information, and outputs the operation of the current step according to the following model: ; in, For the operation of the current step, For historical reasoning information, For user issues, For the candidate operation set, Large-scale language models.

[0031] By incorporating historical reasoning information into the decision-making process of large-scale language models, the models can make continuous and coherent operational choices based on past reasoning trajectories, forming a complete reasoning chain. This avoids the isolation of reasoning steps, reduces the possibility of models skipping steps or taking reasoning shortcuts, and improves the logic and coherence of the reasoning process.

[0032] Furthermore, such as Figure 1 and Figure 2 As shown, between step S3 and step S4, the following step is also included: The virtual entity particle is obtained by averaging the embedding vectors of multiple resulting entities. The virtual particle is represented as: ; in, It is a virtual particle embedding used for subsequent energy calculations. This represents the set of entities retrieved in the k-th step.

[0033] Virtual entity particles are generated by averaging the embedding vectors of the resulting entities, and the features of multiple entities are fused into a unified embedding representation. This provides a standardized and quantifiable feature basis for subsequent potential energy calculation, solves the problem that the results of multiple entities cannot directly participate in energy interaction calculation, and ensures the feasibility and accuracy of potential energy calculation.

[0034] Furthermore, such as Figure 1 and Figure 2 As shown, in step S4, semantic distance is defined using the embedding representation provided by TComplEx for time and the embedding representation provided by CronKGQA for the question. : ; in, and These represent the embeddings of the subject entity and the current virtual entity, respectively. It is a time embedding from TComplEx. It's an embedding from CronKGQA. Indicates respectively to Vectorize, This indicates the vectorized version. Perform a dot product operation and take the real part of the result. It is a small positive constant.

[0035] By integrating the temporal embedding of TComplEx and the question embedding of CronKGQA, the advantages of both technologies in temporal representation and question semantic representation are fully utilized to improve the accuracy of the embedding representation. Based on the semantic distance defined by the real part of the multilinear dot product, the semantic correlation between subject entities, virtual entities, questions and time is quantitatively calculated, providing a scientific and objective numerical basis for subsequent potential energy calculation and ensuring the accuracy of semantic similarity assessment.

[0036] Furthermore, such as Figure 1 and Figure 2 As shown, in step S5, the step potential energy Including content potential With time potential energy ,when At that time, take ;when At that time, take ; ; ; in, and These are question-guided query embeddings used for entity reasoning and time reasoning, respectively.

[0037] Potential energy is divided into content dimension and time dimension, respectively adapted to entity-oriented and time-oriented operation types, to achieve precise matching between potential energy calculation and operation type, so that potential energy can accurately reflect the rationality of different reasoning steps; based on embedded interaction and the real part of multilinear inner product, the potential energy calculation formula is designed to transform semantic relevance into potential energy value, so as to achieve objective quantification of the effectiveness of reasoning steps, and allow the potential energy increment to accurately represent the change of reasoning state, providing a reliable judgment basis for the detection of reasoning chain breakage.

[0038] Furthermore, content maintains momentum. With time potential energy All are based on embedded interactive computation, where the embeddings include: topic entity embedding, virtual entity embedding, temporal embedding, and query embedding obtained by question guidance; and the similarity term is calculated through the real part of the multilinear inner product.

[0039] The calculation basis and core methods of potential energy in the content dimension and time dimension are clearly defined to ensure the consistency and scientific nature of the two potential energy calculation logics. Through unified embedded interaction and multilinear inner product real part calculation, the potential energy results can truly reflect the semantic relationship of each element in the reasoning steps, thereby improving the accuracy and reliability of potential energy increment detection of reasoning chain breaks.

[0040] This embodiment also provides a knowledge graph question-answering reasoning system based on potential energy increment triggering, such as Figures 1-2 As shown, it includes: The candidate operation generation unit is used to generate a set of candidate operations based on the user question and the neighborhood subgraph, and then filter the candidate operation set. The decision unit is used to input the user's question, historical reasoning trajectory, and candidate operation set into a large language model to output the selected operation; The retrieval execution unit is used to perform the selected operation on the time-series knowledge graph and output intermediate results; The potential energy discrimination unit is used to calculate the potential energy in the current step and output the potential energy increment. The re-decision unit is used to roll back the current step when the potential energy increment exceeds a threshold, and inject the failed operation for re-decision.

[0041] Each unit has a clear functional division and performs its own duties, forming a complete reasoning closed loop of "generation-decision-execution-discrimination-correction," ensuring the implementation of the question-answering reasoning method based on potential energy increment. The candidate operation generation unit achieves accurate operation screening, the decision-making unit ensures the rationality of operation selection, the retrieval and execution unit completes knowledge extraction, the potential energy discrimination unit realizes real-time monitoring of the reasoning state, and the re-decision unit completes the correction of erroneous paths. All units work together to improve the efficiency, accuracy, and robustness of temporal knowledge graph question-answering reasoning as a whole. The re-decision unit injects failed operations for re-decision, which can use error feedback to eliminate invalid paths, guide the model to explore the correct reasoning direction, and further improve the effectiveness of reasoning.

[0042] The above description is only a preferred embodiment of the present invention. It should be noted that, for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A knowledge graph question-answering reasoning method based on potential energy increment triggering, characterized in that, include: S1. Obtain the user's question, construct a neighborhood subgraph based on the user's question, generate an initial candidate operation set based on the neighborhood subgraph, and select several candidate operation sets within the initial candidate operation set based on executability judgment and semantic similarity ranking. S2. Input the user question and the set of candidate operations into a large language model, and the large language model outputs the selected operation; S3. Call the retrieval execution unit to perform the selected operation output by the large language model, so as to obtain the corresponding entity; S4. Obtain the subject entity quality and virtual entity quality based on the entities, extract the user question, and calculate the semantic distance; S5. Calculate the potential energy at the current step. And calculate the potential energy increment. When the potential energy increases Exceeding the preset threshold A rollback operation is performed until the final answer is output.

2. The knowledge graph question-answering reasoning method based on potential energy increment triggering according to claim 1, characterized in that, In step S1, in obtaining user questions Next, first obtain the subject entity in the question. Then, search the underlying time-series knowledge graph. neighborhood subgraph ,in, Indicates a relationship. For the target entity, For the subject entity Valid time interval.

3. The knowledge graph question-answering reasoning method based on potential energy increment triggering according to any one of claims 1 or 2, characterized in that, In step S1, Executability assessment includes: assessing the initial candidate operation set. In the current subgraph Whether it is legal or not, record it as ； Semantic relevance ranking includes: calculating the initial candidate operation set. User issues The semantic similarity is used to select the top N high scores based on semantic similarity.

4. The knowledge graph question-answering reasoning method based on potential energy increment triggering according to claim 3, characterized in that, In step S1, a candidate operation set is generated based on a preset operation template. The candidate operation template includes an entity-oriented operation set and a time-oriented operation set. The entity-oriented operation is used to obtain the target entity or relationship-related entity from the neighborhood subgraph, and the time-oriented operation is used to obtain the timestamp or time interval associated with the entity-relationship fact.

5. The knowledge graph question-answering reasoning method based on potential energy increment triggering according to claim 1, characterized in that, Step S2 also includes historical reasoning information, and outputs the operation of the current step according to the following model: ； in, For the operation of the current step, For historical reasoning information, For user issues, For the candidate operation set, Large-scale language models.

6. The knowledge graph question-answering reasoning method based on potential energy increment triggering according to claim 1, characterized in that, Between steps S3 and S4, the following is also included: The virtual entity particle is obtained by averaging the embedding vectors of multiple resulting entities. The virtual particle is represented as: ; in, It is a virtual particle embedding used for subsequent energy calculations. This represents the set of entities retrieved in the k-th step.

7. The knowledge graph question-answering reasoning method based on potential energy increment triggering according to any one of claims 1 or 6, characterized in that, In step S4, semantic distance is defined using the embedding representation provided by TComplEx for time and the embedding representation provided by CronKGQA for the question. : ； in, and These represent the embeddings of the subject entity and the current virtual entity, respectively. It is a time embedding from TComplEx. It's an embedding from CronKGQA. Indicates respectively to Vectorize, This indicates the vectorized version. Perform a dot product operation and take the real part of the result; It is a small positive constant.

8. The knowledge graph question-answering reasoning method based on potential energy increment triggering according to claim 7, characterized in that, In step S5, the step potential energy Including content-driven potential Potential energy with time ,when At that time, take ;when At that time, take ; ； ； in, and These are question-guided query embeddings used for entity reasoning and time reasoning, respectively.

9. The knowledge graph question-answering reasoning method based on potential energy increment triggering according to claim 8, characterized in that, Content maintains potential Potential energy with time All are based on embedded interactive computation, where the embeddings include: topic entity embedding, virtual entity embedding, temporal embedding, and query embedding obtained by question guidance; and the similarity term is calculated through the real part of the multilinear inner product.

10. A knowledge graph question-answering reasoning system based on potential energy increment triggering, characterized in that, include: The candidate operation generation unit is used to generate a set of candidate operations based on the user question and the neighborhood subgraph, and then filter the candidate operation set. The decision unit is used to input the user's question, historical reasoning trajectory, and candidate operation set into a large language model to output the selected operation; The retrieval execution unit is used to perform the selected operation on the time-series knowledge graph and output intermediate results; The potential energy discrimination unit is used to calculate the potential energy in the current step and output the potential energy increment. The re-decision unit is used to roll back the current step when the potential energy increment exceeds a threshold, and inject the failed operation for re-decision.

Images