A time sequence knowledge graph reasoning method and system based on double-system cooperation

Abstract

This invention belongs to the field of knowledge graph reasoning technology, and particularly relates to a temporal knowledge graph reasoning method and system based on dual-system collaboration. It includes constructing concurrent event interaction representations within a query timestamp, generating a candidate set, and performing preliminary ranking of candidate entities in the candidate set; retrieving key evidence chains from cross-timestamp history, inputting the key evidence chains and candidate re-ranking results into a large language model for structured reasoning, and obtaining candidate deep re-ranking results; performing dynamic routing based on confidence, and obtaining the final ranking result based on the preliminary and deep re-ranking results; and selecting the top-ranked result from the final ranking results as the temporal knowledge graph reasoning result. This invention explicitly models concurrent event interactions within the same timestamp, avoiding prediction bias caused by the smoothing out of concurrent interactions, focusing and compressing key evidence from long histories, and mitigating evidence dilution caused by contextual redundancy in long histories.

Description

Technical Field

[0001] This invention belongs to the field of knowledge graph reasoning technology, and in particular relates to a temporal knowledge graph reasoning method and system based on dual-system collaboration. Background Technology

[0002] Temporal knowledge graphs typically represent the dynamic evolution of interactions between entities in the real world using timestamped quadruples, such as (e h ,r,e t ,t), where e h and e t Let r be the entity, r be the relation, and t be the timestamp. Due to the unavoidable omissions and noise in the data collection and annotation of time-series knowledge graphs, it is necessary to reason and complete unknown facts based on limited factual information.

[0003] To address this need, temporal knowledge graph reasoning completion aims to predict and sort missing elements based on timestamped structural information. Typically, given a quadruple query (?, r, et, t) or (eh, r, ?, t), temporal knowledge graph reasoning completion needs to sort the candidate entity set and predict missing elements.

[0004] In recent years, research on reasoning completion of temporal knowledge graphs has evolved from early learning methods based on temporal embeddings to explicit modeling of temporal dependencies through recursive models, and then to temporal modeling frameworks that incorporate attention mechanisms combined with transformers. With the development of large language models, the idea of ​​using the semantic understanding capabilities of large language models to reason and complete temporal knowledge graphs has gradually gained attention.

[0005] However, existing temporal knowledge graph reasoning completion methods based on large language models still have two problems: On the one hand, existing methods often linearize historical events in chronological order or simply aggregate multiple events that occur within the same timestamp. When there are a large number of concurrent events within the same timestamp and there are interactions such as constraints, collaboration, and substitution between the events, the model can often only learn coarse-grained time trends and find it difficult to capture the key mechanisms that drive subsequent evolution through simultaneous interactions.

[0006] On the other hand, real-world TKGs often have a long span and high fact density. If a large amount of historical facts are directly input into a large language model for reasoning and completion, the key evidence will be diluted due to contextual redundancy, making it difficult for the model to stably focus on the few historical clues that truly determine the current prediction. Summary of the Invention

[0007] To overcome the shortcomings of the prior art, this invention provides a temporal knowledge graph reasoning method and system based on dual-system collaboration. Through dual-system collaborative design, concurrent event interactions within the same timestamp are explicitly modeled, avoiding prediction bias caused by the smoothing of concurrent interactions. Key evidence in long history is focused and compressed, alleviating evidence dilution caused by redundancy in long historical context.

[0008] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions: The first aspect of this invention provides a temporal knowledge graph reasoning method based on dual-system collaboration.

[0009] A temporal knowledge graph reasoning method based on dual-system collaboration includes the following steps: Concurrent event interaction representations are constructed within the query timestamp, a candidate set is generated, and the candidate entities in the candidate set are initially sorted to obtain the preliminary candidate reordering results and confidence levels. Retrieve key evidence chains from cross-timestamp history, input the key evidence chains and candidate rearrangement results into a large language model, perform structured reasoning, and obtain candidate deep rearrangement results; Dynamic routing is performed based on confidence levels, and the final ranking result is obtained based on the preliminary candidate re-ranking results and the candidate depth re-ranking results. The result that appears first in the final sorting is selected as the result of the temporal knowledge graph inference.

[0010] A second aspect of the present invention provides a temporal knowledge graph reasoning system based on dual-system collaboration.

[0011] A time-series knowledge graph reasoning system based on dual-system collaboration includes: The preliminary reordering module is configured to: construct a concurrent event interaction representation within the query timestamp, generate a candidate set, perform preliminary sorting on the candidate entities in the candidate set, and obtain the preliminary candidate reordering results and confidence levels; The deep reordering module is configured to: retrieve key evidence chains from cross-timestamp history, input the key evidence chains and candidate reordering results into the large language model, perform structured reasoning, and obtain candidate deep reordering results; The dynamic routing module is configured to perform dynamic routing based on confidence and obtain the final ranking result based on the preliminary candidate re-ranking result and the candidate depth re-ranking result. The reasoning module is configured to: filter the top result in the final sorting results as the reasoning result for the time-series knowledge graph.

[0012] A third aspect of the present invention provides a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the steps of a time-series knowledge graph reasoning method based on dual-system collaboration as described in the first aspect of the present invention.

[0013] A fourth aspect of the present invention provides an electronic device, including a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps in a time-series knowledge graph reasoning method based on dual-system collaboration as described in the first aspect of the present invention.

[0014] The above one or more technical solutions have the following beneficial effects: This invention provides a temporal knowledge graph reasoning method and system based on dual-system collaboration. It enhances the expressive power of simultaneous timestamp interaction through concurrent graph modeling, reducing prediction bias caused by concurrent information loss; it improves reasoning efficiency by first shrinking and then re-ranking the candidate space, transforming full entity prediction into small-scale candidate rearrangement; it enhances reasoning stability by compressing historical chain evidence, increasing the density of key evidence under long historical conditions; and it achieves a balance between efficiency and accuracy by using a gating decision mechanism to realize the collaboration between fast branching and verification branches.

[0015] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0016] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0017] Figure 1 This is a diagram of the overall framework for dual-system collaborative reasoning in Example 1.

[0018] Figure 2 This is a flowchart of the dual-system collaborative reasoning method in Example 1. Detailed Implementation

[0019] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0020] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of the present invention.

[0021] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0022] Terminology Explanation: Temporal Knowledge Graph (TKG): Based on the knowledge graph, a timestamp (t) is introduced, and a quadruple ((e_s, r,e_o, t)) is used to represent the fact that occurs at time (t).

[0023] Temporal Knowledge Graph Reasoning and Completion (TKGR): Given historical records, predict or sort missing facts. Typical queries are ((e_s, r, ?, t)) or ((?, r, e_o, t)).

[0024] Concurrent events: A set of four events that occur within the same timestamp.

[0025] Concurrent Interaction Graph (TIME-slice): A subgraph constructed in units of a time slice (t), where nodes are entities and edges are the relationships between events that occur within that time slice.

[0026] Historical Evidence Chain: A set of time-stamped evidence associated with a query.

[0027] Large Language Model (LLM): A pre-trained language model capable of generative reasoning.

[0028] Parameter Efficient Fine-Tuning (PEFT): A fine-tuning method for task adaptation without updating or with minimal updates to the base model parameters.

[0029] Confidence: A measure of the reliability of a model’s predictions, which can be expressed as a score interval, distribution entropy, calibration probability, or learned confidence header output.

[0030] Example 1 This embodiment discloses a time-series knowledge graph reasoning method based on dual-system collaboration. Through dual-system collaborative design, concurrent event interactions within the same timestamp are explicitly modeled, key evidence in long history is focused and compressed, prediction bias caused by the smoothing out of concurrent interactions is avoided, and evidence dilution caused by the redundancy of long historical context is alleviated.

[0031] like Figure 1 and Figure 2 As shown, a time-series knowledge graph reasoning method based on dual-system collaboration includes the following steps: Concurrent event interaction representations are constructed within the query timestamp, a candidate set is generated, and the candidate entities in the candidate set are initially sorted to obtain the preliminary candidate reordering results and confidence levels. Retrieve key evidence chains from cross-timestamp history, input the key evidence chains and candidate rearrangement results into a large language model, perform structured reasoning, and obtain candidate deep rearrangement results; Dynamic routing is performed based on confidence levels, and the final ranking result is obtained based on the preliminary candidate re-ranking results and the candidate depth re-ranking results. The result that appears first in the final sorting is selected as the result of the temporal knowledge graph inference.

[0032] This embodiment constructs a dual-system that works in tandem with each other, including a fast intuition system and a slow thinking system. The fast intuition system is used to obtain the preliminary candidate reordering results and performs the above-mentioned steps: constructing a concurrent event interaction representation within the query timestamp, generating a candidate set, and performing a preliminary sorting of the candidate entities in the candidate set to obtain the preliminary candidate reordering results and confidence levels. The slow-thinking system is used to obtain candidate deep rearrangement results by performing the following steps: retrieving key evidence chains from cross-timestamp history, inputting the key evidence chains and candidate rearrangement results into a large language model, performing structured reasoning, and obtaining candidate deep rearrangement results.

[0033] Overall, the dual-system collaborative reasoning framework adopted in this embodiment, such as Figure 1 As shown, it includes three processing steps: ① A fast, intuitive system for concurrent interaction awareness: used to construct concurrent event interaction representations within query timestamps, generate candidate sets and perform fast initial ranking, while outputting confidence scores; ② A slow-thinking system focusing on long-historical evidence: used to retrieve a small number of key evidence chains from cross-timestamp history, organize the evidence in an instructional manner and input it into a large language model to complete structured reasoning, and perform deep rearrangement of the candidate set; ③ Gated fusion module: used for dynamic routing based on confidence level and fusion of fast and slow system outputs to obtain the final ranking result.

[0034] Among them, the fast intuition system is responsible for efficiently screening candidates and giving confidence levels, the slow thinking system is responsible for evidence verification and deep re-ranking of low-confidence samples, and the gating fusion module is responsible for unified decision-making and outputting the final ranking.

[0035] Furthermore, a concurrent event interaction representation is constructed within the query timestamp to generate a candidate set, specifically including: Collect historical quadruple facts and group them by timestamp, and construct a concurrent graph within each timestamp; Get the query timestamp and the corresponding concurrency graph; In the concurrent graph corresponding to the query timestamp, a reverse edge is added to each edge, and the inbound edge count of the relation-target node is normalized. The messages are weighted during the convolution stage of the relation graph. Select a context time slice within the historical window prior to the query time, and perform relational graph convolution to update the entity representation on the concurrent graph.

[0036] More specifically, a fast intuitive system that is aware of concurrent interactions can use any of the following to achieve concurrent structure modeling and initial candidate ranking: graph convolution, graph attention, temporal graph Transformer, or hypergraph coding.

[0037] Furthermore, a candidate set is generated, specifically including: Calculate the structure score and historical prior score of the query separately, normalize and fuse the structure score and historical prior score to obtain the comprehensive score of the candidate entity; Candidate entities are sorted in descending order of their overall scores, and the top [entities] are selected. Each entity forms a candidate set.

[0038] The structural scoring of the fast intuition system can be achieved by using any one of the following methods: trilinear scoring, translational scoring, bilinear scoring, or neural matching scoring to score and rank candidates.

[0039] The candidate generation of the fast intuition system can adopt any of the following strategies to form a candidate set: fixed Top-K, adaptive Top-K, or multi-way recall union.

[0040] Furthermore, a conditional log-likelihood score is calculated for each candidate entity in the candidate set, and the preliminary ranking is completed according to the conditional log-likelihood score.

[0041] Furthermore, key chains of evidence are retrieved from cross-timestamp history, specifically including: In the interval Internally retrieve historical facts related to the head entity and candidate entities to construct the target evidence chain; in, This indicates the timestamp corresponding to the query to be predicted. Indicates the length of the history backtracking window, the interval [ , ) indicates from the current query time Backtracking W time units, excluding the current time. The half-open time interval; Chain expansion is scored based on entity connectivity and temporal proximity: ; in, It is a positive weighting constant; Based on chain expansion scoring, the obtained target evidence chains are filtered to obtain key evidence chains.

[0042] More specifically, slow-thinking systems that focus on long-term historical evidence can use any of the following methods to obtain key historical evidence: multi-hop graph retrieval, vector retrieval, rule retrieval, or a combination thereof.

[0043] Furthermore, the key evidence chains and candidate rearrangement results are input into a large language model for structured reasoning to obtain candidate deep rearrangement results, specifically including: Organize evidence in an instructional manner using any of the following methods: chain, tree, subgraph, tabular, or structured text. The evidence organized in an instructional manner and the preliminary candidate rearrangement results are input into the large language model. The candidate depth ranking is achieved by using one of the following methods: candidate-by-candidate scoring, pairwise comparison rearrangement, or list-based rearrangement, and the candidate depth rearrangement results are obtained.

[0044] Furthermore, dynamic routing is performed based on confidence levels, and the final ranking result is obtained based on the preliminary candidate re-ranking results and the candidate depth re-ranking results, specifically including: First, the fast system outputs the preliminary candidate rearrangement results and their confidence levels; When the confidence level is not lower than the gate threshold, the preliminary candidate rearrangement result is directly output as the final ranking result; When the confidence level is lower than the gating threshold, the preliminary candidate rearrangement results, key evidence chains, and query descriptions are input into the Slow Thinking system to obtain the candidate deep rearrangement results.

[0045] In the serial gating mode, the candidate depth rearrangement result is used as the final ranking result; or, in the parallel fusion mode, the candidate preliminary rearrangement result and the candidate depth rearrangement result are weighted fusion or inverse sorting fusion to obtain the final ranking result.

[0046] The output fusion of fast and slow systems can be achieved by using any of the following methods: linear weighted fusion, inverse ranking fusion, Borda fusion, or learning-based fusion to obtain the final ranking result.

[0047] The specific steps are as follows: Collect historical quadruple facts and group them by timestamp. Construct a concurrent graph within each timestamp, defined as:

[0048] in, Represents any timestamp, Represents a set of entity nodes. Represents timestamp The set of concurrent edges below, Indicates the head entity. Indicates the relation type, Indicates the tail entity. This indicates that the timestamp is... A historical fact projected onto a concurrent graph The resulting directed edge.

[0049] To enhance the concurrent relationship propagation capability, a reverse edge is added to each edge, and relationship-target node incoming edge count normalization is adopted. Messages are weighted during the relationship graph convolution stage.

[0050] Select a context time slice within the history window prior to the query time, and perform relational graph convolution to update the entity representation on the concurrent graph:

[0051] in, This represents the entity representation of node v in the l-th layer of the relational graph convolution. This represents the entity representation of node v in the (l+1)th layer of the relational graph convolution. Let (u,r,v) represent the self-loop transformation matrix corresponding to the l-th layer. This represents a relation edge in the concurrent graph corresponding to timestamp τ that points to the target node v, where u is the source node, v is the target node, and r is the relation type. σ represents the number of incoming edges from the target node v to the relation r, and σ represents the non-linear activation function. This represents the entity representation of node u in the l-th layer relation graph convolution.

[0052] Furthermore, for each original edge (h,r,t) under timestamp τ, the original directed edge h→t is retained, and a reverse edge t→h is added, with the reverse edge corresponding to the reverse relationship identifier r. rev This allows relational information to propagate bidirectionally, both forward and backward.

[0053] By adding reverse edges to the original edges, relational information within the same timestamp can be propagated bidirectionally, thereby enhancing the expressive power of inter-entity interaction dependencies. By normalizing the inbound edge count of relation-target nodes, high-frequency relations and high-in-degree nodes can be prevented from occupying excessive weight in message aggregation, improving the stability of entity representation updates under different relational patterns. By selecting context time slices within the historical window before the query time and performing relational graph convolution, concurrent interaction information at the same moment and recent historical structural evolution information can be retained simultaneously during the candidate generation stage, improving the accuracy of structural candidate ranking.

[0054] For any query to be completed (For example, ) and candidate set C( Candidate entities in ) Calculate candidate entities respectively In the query The structural score and the historical prior score corresponding to the missing location are obtained, and the two types of scores are normalized and fused to obtain the comprehensive score of the candidate entity. :

[0055] in, Indicates a query to be predicted. Indicates candidate entities, This indicates that candidate entity e is in the query The structure-normalized score is calculated from the concurrent graph context using a structure scoring model. Indicates candidate entities In the query The historical prior normalized score is obtained based on historical facts consistent with the query header entities and relations, combined with time decay statistics; λ represents the fusion weight of the structural score and the historical prior score, with a value range of [0,1]. Indicates candidate entities In the query The overall score below; according to Sort the candidate entities in descending order and select the top K candidate entities to form a candidate set or the final candidate ranking result.

[0056] Concurrent evidence, short historical evidence, and candidate sets are organized according to a fixed template and output uniformly according to constraints to ensure that the model output is comparable.

[0057] Calculate the conditional log-likelihood score for each candidate entity in the candidate set and perform an initial sort:

[0058] in, This represents the first-stage input context of the fast intuitive system. This represents the candidate text representation corresponding to candidate entity e. Indicates candidate text The corresponding mark length; This represents the parameters of the large language model corresponding to the fast intuition system; This represents the first-stage length-normalized conditional log-likelihood score of candidate entity e under query q. The score is calculated separately for each candidate entity in the candidate set. The candidates are then initially rearranged in descending order of their scores.

[0059] This step outputs the candidate rearrangement results and their confidence features, which serve as the basis for subsequent gating decisions.

[0060] In the interval Internally, historical facts related to the head entity and candidate entities are retrieved, a small number of key evidence chains are constructed, and chain expansion is scored based on entity connectivity and temporal proximity.

[0061] in, It is a positive weighting constant.

[0062] The historical chain evidence and the output of the fast intuition system are jointly encoded and then input into the slow thinking model to perform deep reordering of the candidate set and obtain the ranking result.

[0063] After jointly encoding historical chain evidence and the output of the fast intuition system, the data is input into the slow thinking model to perform a deep reordering of the candidate set, specifically: The query description, the prediction results and their confidence levels from the Quick Intuition system, concurrent evidence, key evidence chains, and candidate sets are assembled according to a unified template to form the input context for the second stage. For each candidate entity e in the candidate set, compute its candidate text representation. In context Length normalized conditional log-likelihood score under the given conditions:

[0064] in, This represents the second-stage input context of the slow-thinking system. This represents the parameters of the large language model corresponding to the slow-thinking system. The candidate text representation of candidate entity e. This represents the second-stage length-normalized conditional log-likelihood score of candidate entity e under query q.

[0065] Then according to The candidate entities are sorted in descending order to obtain the candidate depth reordering result.

[0066] Calculate the confidence level based on the difference between the top two candidate entities and the candidate distribution entropy in the fast intuition system:

[0067] in, This represents the difference between the scores of the top-ranked candidate entity and the second-ranked candidate entity in the fast intuition system. The information entropy representing the score distribution of candidate entities is preferably calculated based on the normalized probability distribution of the top M candidate entities. This indicates the confidence level of the fast intuitive system in the current query.

[0068] Dynamic routing is performed based on confidence levels, and the final ranking result is obtained based on the preliminary candidate re-ranking results and the candidate depth re-ranking results, specifically including: When the confidence level is not lower than the gate threshold, the preliminary candidate rearrangement result is directly output as the final ranking result; When the confidence level is lower than the gating threshold, the final ranking result is obtained by using either serial gating mode or parallel fusion mode. The final sorting result is obtained by using either a series gating mode or a parallel fusion mode, specifically as follows: In the serial gating mode, the candidate depth rearrangement result is used as the final sorting result; In the parallel fusion mode, the preliminary candidate rearrangement results and the deep candidate rearrangement results are fused by weighted fusion or reciprocal sorting to obtain the final ranking result.

[0069] The final output includes a unified sorting result and corresponding confidence information.

[0070] This embodiment has the following advantages: (1) Improve the ability to express the interaction of the same timestamp by using concurrent graph modeling, and reduce the prediction bias caused by the loss of concurrent information; (2) By first shrinking the candidate space and then refining it, the prediction of all entities is transformed into a small-scale candidate rearrangement, thereby improving the reasoning efficiency; (3) By compressing historical chain evidence, the density of key evidence under long historical conditions is increased, thereby enhancing the stability of reasoning; (4) The fast branch and the verification branch are coordinated through the gating decision mechanism. When the confidence of the fast branch output is not lower than the preset threshold, the fast branch result is directly output. When the confidence of the fast branch output is lower than the preset threshold, the result is switched to the verification branch output, thereby achieving a balance between efficiency and accuracy. (5) In the parallel collaborative implementation, by jointly judging and fusing the fast branch result and the verification branch result, when the confidence level of the fast branch output is not lower than the preset threshold, the fast branch result is directly used as the final output result. When the confidence level of the fast branch output is lower than the preset threshold, the fast branch result and the verification branch result are fused and the final result is output, thereby further improving the accuracy and stability of the final sorting result in complex query scenarios.

[0071] Example 2 This embodiment discloses a time-series knowledge graph reasoning system based on dual-system collaboration.

[0072] A time-series knowledge graph reasoning system based on dual-system collaboration includes: The preliminary reordering module is configured to: construct a concurrent event interaction representation within the query timestamp, generate a candidate set, perform preliminary sorting on the candidate entities in the candidate set, and obtain the preliminary candidate reordering results and confidence levels; The deep reordering module is configured to: retrieve key evidence chains from cross-timestamp history, input the key evidence chains and candidate reordering results into the large language model, perform structured reasoning, and obtain candidate deep reordering results; The dynamic routing module is configured to perform dynamic routing based on confidence and obtain the final ranking result based on the preliminary candidate re-ranking result and the candidate depth re-ranking result. The reasoning module is configured to: filter the top result in the final sorting results as the reasoning result for the time-series knowledge graph.

[0073] Example 3 The purpose of this embodiment is to provide a computer-readable storage medium.

[0074] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of a time-series knowledge graph reasoning method based on dual-system collaboration as described in Embodiment 1 of this disclosure.

[0075] Example 4 The purpose of this embodiment is to provide an electronic device.

[0076] An electronic device includes a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps in a time-series knowledge graph reasoning method based on dual-system collaboration as described in Embodiment 1 of this disclosure.

[0077] The steps and methods involved in the apparatuses of Embodiments 2, 3, and 4 above correspond to those in Embodiment 1. For specific implementation details, please refer to the relevant description section of Embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood as including any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and enabling the processor to perform any of the methods in this invention.

[0078] Those skilled in the art will understand that the modules or steps of the present invention described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, thereby allowing them to be stored in a storage device for execution by a computer device, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. The present invention is not limited to any particular combination of hardware and software.

[0079] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A time-series knowledge graph reasoning method based on dual-system collaboration, characterized in that, Includes the following steps: Concurrent event interaction representations are constructed within the query timestamp, a candidate set is generated, and the candidate entities in the candidate set are initially sorted to obtain the preliminary candidate reordering results and confidence levels. Retrieve key evidence chains from cross-timestamp history, input the key evidence chains and candidate rearrangement results into a large language model, perform structured reasoning, and obtain candidate deep rearrangement results; Dynamic routing is performed based on confidence levels, and the final ranking result is obtained based on the preliminary candidate re-ranking results and the candidate depth re-ranking results. The result that appears first in the final sorting is selected as the result of the temporal knowledge graph inference.

2. The time-series knowledge graph reasoning method based on dual-system collaboration as described in claim 1, characterized in that, Construct concurrent event interaction representations within the query timestamp and generate a candidate set, specifically including: Collect historical quadruple facts and group them by timestamp, and construct a concurrent graph within each timestamp; Get the query timestamp and the corresponding concurrency graph; In the concurrent graph corresponding to the query timestamp, a reverse edge is added to each edge, and the inbound edge count of the relation-target node is normalized. The messages are weighted during the convolution stage of the relation graph. Select a context time slice within the historical window prior to the query time, and perform relational graph convolution to update the entity representation on the concurrent graph; Entities in the concurrent graph corresponding to the query timestamp are selected as candidate entities.

3. The time-series knowledge graph reasoning method based on dual-system collaboration as described in claim 1, characterized in that, Generating a candidate set specifically includes: Calculate the structure score and historical prior score of the query separately, normalize and fuse the structure score and historical prior score to obtain the comprehensive score of the candidate entity; Candidate entities are sorted in descending order of their overall scores, and the top [entities] are selected. Each entity forms a candidate set; The formula for calculating the comprehensive score of candidate entities is as follows: ； in, Indicates a query to be predicted; Indicates candidate entities; This indicates that candidate entity e is in the query The structure-normalized fractions below; Indicates candidate entities In the query The historical prior normalized score is given by λ; λ represents the fusion weight between the structural score and the historical prior score, and its value ranges from [0,1]. Indicates candidate entities In the query The overall score below.

4. The time-series knowledge graph reasoning method based on dual-system collaboration as described in claim 1, characterized in that, Calculate the conditional log-likelihood score for each candidate entity in the candidate set, and perform preliminary sorting based on the conditional log-likelihood score.

5. The time-series knowledge graph reasoning method based on dual-system collaboration as described in claim 1, characterized in that, Retrieving key chains of evidence from cross-timestamped history, specifically including: In the interval Internally retrieve historical facts related to the head entity and candidate entities to construct the target evidence chain; among which, This indicates the timestamp corresponding to the query to be predicted. Indicates the length of the history backtracking window, the interval [ , ) indicates from the current query time Backtracking W time units, excluding the current time. The half-open time interval; Chain expansion is scored based on entity connectivity and temporal proximity: ； in, It is a positive weighting constant; Based on chain expansion scoring, the obtained target evidence chains are filtered to obtain key evidence chains.

6. The time-series knowledge graph reasoning method based on dual-system collaboration as described in claim 1, characterized in that, The key evidence chain and candidate rearrangement results are input into a large language model for structured reasoning to obtain candidate deep rearrangement results, specifically including: Organize evidence in an instructional manner using any of the following methods: chain, tree, subgraph, tabular, or structured text. The evidence organized in an instructional manner and the preliminary candidate rearrangement results are input into the large language model. The candidate depth ranking is achieved by using one of the following methods: candidate-by-candidate scoring, pairwise comparison rearrangement, or list-based rearrangement, and the candidate depth rearrangement results are obtained.

7. The time-series knowledge graph reasoning method based on dual-system collaboration as described in claim 1, characterized in that, Dynamic routing is performed based on confidence levels, and the final ranking result is obtained based on the preliminary candidate re-ranking results and the candidate depth re-ranking results, specifically including: When the confidence level is not lower than the gate threshold, the preliminary candidate rearrangement result is directly output as the final sorting result; When the confidence level is lower than the gating threshold, the final ranking result is obtained by using either serial gating mode or parallel fusion mode. The final sorting result is obtained by using either a series gating mode or a parallel fusion mode, specifically as follows: In the serial gating mode, the candidate depth rearrangement result is used as the final sorting result; In the parallel fusion mode, the preliminary candidate rearrangement results and the deep candidate rearrangement results are fused by weighted fusion or reciprocal sorting to obtain the final ranking result.

8. A time-series knowledge graph reasoning system based on dual-system collaboration, characterized in that, include: The preliminary reordering module is configured to: construct a concurrent event interaction representation within the query timestamp, generate a candidate set, perform preliminary sorting on the candidate entities in the candidate set, and obtain the preliminary candidate reordering results and confidence levels; The deep reordering module is configured to: retrieve key evidence chains from cross-timestamp history, input the key evidence chains and candidate reordering results into the large language model, perform structured reasoning, and obtain candidate deep reordering results; The dynamic routing module is configured to perform dynamic routing based on confidence and obtain the final ranking result based on the preliminary candidate re-ranking result and the candidate depth re-ranking result. The reasoning module is configured to: filter the top result in the final sorting results as the reasoning result for the time-series knowledge graph.

9. A computer-readable storage medium having a program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps in the time-series knowledge graph reasoning method based on dual-system collaboration as described in any one of claims 1-7.

10. An electronic device comprising a memory, a processor, and a program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the time-series knowledge graph reasoning method based on dual-system collaboration as described in any one of claims 1-7.

Images