Search

CN-121860024-B - Event cause and effect mining method and system based on map constraint and time sequence optimal transmission

CN121860024BCN 121860024 BCN121860024 BCN 121860024BCN-121860024-B

Abstract

The invention provides an event causal mining method and system based on map constraint and time sequence optimal transmission, wherein the method comprises the steps of segmenting an emergency description text into text blocks, extracting atomic events from each text block, and formalizing each atomic event into a binary feature group Any two atomic events And Mapping to event nodes in preset knowledge graph And Computing event nodes And Constructing teacher model probability distribution according to the shortest path topological distance in the map, calculating vectors And The cosine distance between the two is used as basic semantic transmission cost, time sequence constraint is carried out on the basic semantic transmission cost, a time sequence transmission cost matrix is constructed, an optimal transmission matrix which minimizes the total transmission cost is searched, and the event causal relation which satisfies time sequence optimal transmission under the constraint of the domain knowledge graph is obtained through the joint loss function of the minimized teacher probability distribution and the optimal transmission matrix.

Inventors

  • WU SHUNJIE
  • CHEN HUANHUAN

Assignees

  • 中国科学技术大学

Dates

Publication Date
20260512
Application Date
20260317

Claims (10)

  1. 1. An event causal mining method based on graph constraint and time sequence optimal transmission, which is characterized by comprising the following steps: cutting the emergency description text to be processed into text blocks with preset lengths, and reserving an overlapping area between adjacent text blocks; atomic event extraction is carried out on each text block to form an atomic event group, and each atomic event is formed into a binary characteristic group , Semantic content feature vectors for characterizing atomic events, A timestamp for characterizing the atomic event timing; any two atomic events in the atomic event group And Mapping to corresponding event nodes in a preset domain knowledge graph And And calculates event nodes And Shortest path topological distance in the domain knowledge graph, constructing teacher model probability distribution according to the shortest path topological distance corresponding to any two atomic events, wherein the teacher model probability distribution is used for indicating the events according to human common knowledge Resulting in an event Is a priori probability of (2); Calculating vectors And Cosine distance between them and take it as event To event Performing time sequence constraint on the basic semantic transmission cost to obtain time sequence transmission cost, constructing a time sequence transmission cost matrix according to the time sequence transmission cost of any two atomic events in the atomic event group, and searching an optimal transmission matrix for minimizing the total transmission cost by adopting Sinkhorn iteration algorithm to obtain student model prediction probability distribution; constructing a combined loss function of the teacher model probability distribution and the student model prediction probability distribution, and obtaining event causal relations meeting time sequence optimal transmission under the constraint of the domain knowledge graph through minimizing the combined loss function.
  2. 2. The method of claim 1, wherein constructing a teacher model probability distribution based on shortest path topology distances corresponding to any two atomic events comprises: Constructing teacher model probability distribution by utilizing Softmax function with temperature coefficient according to the shortest path topological distance The calculation formula is as follows: ; Wherein, the Event nodes on domain knowledge graph And Shortest path topological distance in the map; is a temperature coefficient and is used for adjusting the smoothness of the distribution; As the total number of events in the atomic event group, For indicating events based on common sense of human Resulting in an event Is a priori probability of (c).
  3. 3. The method of claim 1, wherein timing constraints on the base semantic transmission costs result in timing transmission costs, comprising: superimposing a preset time sequence physical constraint function on the basis of the basic semantic transmission cost to construct a final time sequence transmission cost : ; Wherein, the As a time sequential physical constraint function, it is defined as a piecewise function: 。
  4. 4. the method of claim 1, wherein constructing a joint loss function of a teacher model probability distribution and a student model predictive probability distribution comprises: constructing optimal transmission loss based on timing mask And knowledge distillation loss for measuring the difference between student model prediction probability and teacher model distribution probability ; Calculating optimal transmission loss And knowledge distillation loss The joint loss function of (2) is as follows: Wherein, the The prediction probability of the student model is output by Sinkhorn algorithm; the teacher model distribution probability is generated based on the knowledge graph; Is a numerical stability constant; Is a balance super parameter; At the cost of time series transmission.
  5. 5. The method according to claim 1, wherein the method further comprises: Constructing metatask distribution, wherein each metatask comprises a support set and a query set; when event causal mining is performed on a new field, global parameters after system pre-training is completed are obtained Computing joint loss functions using samples in metatask support sets of the current domain And performs one or more gradient descent updates to obtain temporary parameters adapted to the current domain : ; Wherein, the For the learning rate of the inner layer cycle, In order to support the collection of objects, As a loss function Is a gradient of (a).
  6. 6. The method of claim 1, wherein after segmenting the incident description text to be processed into text blocks of a preset length, the method further comprises: the absolute time description information and the relative time description information in the text block are identified, the absolute time description information and the relative time description information are converted into a unified timestamp format, and the normalized time characteristics are calculated.
  7. 7. The method according to any one of claims 1-6, further comprising: Screening out that the weight is higher than a preset threshold value from the student model prediction probability distribution As candidate causal edges, constructing a global directed graph based on the candidate causal edges; performing topological sorting on the global directed graph to convert the global directed graph into a sequence constraint graph, and detecting and removing a circulating edge through a topological sorting algorithm to ensure that the sequence constraint graph meets the acyclic performance; The betweenness centrality of each node is calculated based on the sequence constraint graph, key blocking points on the propagation path are identified according to the betweenness centrality, corresponding emergency blocking suggestions are generated according to the key blocking points, and betweenness centrality is calculated The calculation formula is as follows: ; Wherein, the Representing slave source nodes To the target node Is provided for the number of all the shortest causal paths, Representing the passage of nodes therein Is a number of paths of the network.
  8. 8. An event causal mining system based on graph constraints and time-series optimal transmission, the system comprising: the preprocessing module is used for segmenting the emergency description text to be processed into text blocks with preset lengths, and an overlapping area is reserved between the adjacent text blocks; the event extraction module is used for extracting the atomic events of each text block to form an atomic event group and formalizing each atomic event into a binary feature group , Semantic content feature vectors for characterizing atomic events, A timestamp for characterizing the atomic event timing; The common sense atlas causal mining module is used for mining any two atomic events in the atomic event group And Mapping to corresponding event nodes in a preset domain knowledge graph And And calculates event nodes And Shortest path topological distance in the domain knowledge graph, constructing teacher model probability distribution according to the shortest path topological distance corresponding to any two atomic events, wherein the teacher model probability distribution is used for indicating the events according to human common knowledge Resulting in an event Is a priori probability of (2); A time sequence causal mining module for calculating vectors And Cosine distance between them and take it as event To event Performing time sequence constraint on the basic semantic transmission cost to obtain time sequence transmission cost, constructing a time sequence transmission cost matrix according to the time sequence transmission cost of any two atomic events in the atomic event group, and searching an optimal transmission matrix for minimizing the total transmission cost by adopting Sinkhorn iteration algorithm to obtain student model prediction probability distribution; And the joint optimization module is used for constructing a joint loss function of the teacher model probability distribution and the student model prediction probability distribution, and obtaining event causal relations meeting time sequence optimal transmission under the constraint of the domain knowledge graph through minimizing the joint loss function.
  9. 9. The system of claim 8, wherein the system further comprises: the domain adaptation module is used for constructing meta-task distribution, each meta-task comprises a support set and a query set, and when event causal mining is carried out on a new domain, global parameters of system pretraining completion are obtained Computing joint loss functions using samples in metatask support sets of the current domain And performs one or more gradient descent updates to obtain temporary parameters adapted to the current domain : ; Wherein, the For the learning rate of the inner layer cycle, In order to support the collection of objects, As a loss function Is a gradient of (a).
  10. 10. The system of claim 8, wherein the system further comprises: The logic verification module is used for screening out that the weight is higher than a preset threshold value from the student model prediction probability distribution The global directed graph is topologically ordered to convert the global directed graph into a sequence constraint graph, and cyclic edges are detected and removed through a topological ordering algorithm to ensure that the sequence constraint graph meets the acyclic performance; The decision output module is used for calculating the betweenness centrality of each node based on the sequence constraint graph, identifying key blocking points on the propagation path according to the betweenness centrality, and generating corresponding emergency blocking suggestions according to the key blocking points The calculation formula is as follows: ; Wherein, the Representing slave source nodes To the target node Is provided for the number of all the shortest causal paths, Representing the passage of nodes therein Is a number of paths of the network.

Description

Event cause and effect mining method and system based on map constraint and time sequence optimal transmission Technical Field The invention relates to the technical field of artificial intelligence and natural language processing, in particular to an event causal mining method and system based on atlas constraint and time sequence optimal transmission. Background Emergency response and decision-making of emergencies are highly dependent on a deep understanding of the law of evolution of the event, with its associated text data usually in unstructured or semi-structured form, covering complex causal logic. However, the data format is various and the information is scattered, so that the cost of manual analysis and causal extraction is high, the efficiency is low, and human errors are easy to introduce. With the rapid development of artificial intelligence and natural language processing technology, the technology based on deep learning shows strong capability in the fields of text understanding and information extraction, and provides a new idea for intelligent analysis of unstructured data. Nevertheless, challenges remain in how to mine deep causal links from complex text, and in particular how to guarantee causal inference chronological order and physical knowledge at the algorithm level. In the text-oriented knowledge unit association relation mining technology proposed in the prior art, candidate knowledge unit pairs are screened out by calculating term frequency and distance characteristics of knowledge units, and association mining is carried out by using statistical indexes such as support degree and the like. The method realizes automatic association of text knowledge to a certain extent, provides statistical basis for preliminary analysis of massive texts, and is suitable for shallow association analysis in the general field. In the prior art, information verification technology of a specific field is also proposed, and verification is carried out on the authenticity of text content by constructing a knowledge graph and feature engineering of the specific field, but the technology is only suitable for detecting the text of a specific vertical category. With the development of large model technology, the prior art also provides event extraction technology based on a large model, and by utilizing the semantic understanding capability of a large language model, event trigger words and arguments are identified from unstructured texts and are converted into structured data, so that the dependence on a traditional rule template is reduced. Although the large model improves the semantic generalization capability of event extraction, the large model has serious limitation in processing causal mining tasks, namely, a general language model usually lacks strict physical constraint on time dimension and is easy to connect according to semantic similarity, so that inverse time sequence logic errors of 'result cause'. In summary, the prior art has the drawbacks of lacking hard constraints of physical timing and lacking efficient guidance of common sense logic when dealing with causal mining of emergencies. Most of the existing statistical mining or semantic matching algorithms are time-insensitive, logic spurious of reverse timing sequences is difficult to eradicate from an algorithm bottom layer, and the existing methods cannot effectively utilize domain knowledge patterns as prior distribution to calibrate a deep learning model, so that the reasoning capacity and logic interpretability of the model are insufficient when the model faces an implicit causal or small sample scene. Disclosure of Invention In view of the above, the present invention has been made in order to provide an event causal mining method and system based on graph constraint and time-series optimal transmission, which overcome the above problems. The invention provides an event causal mining method based on graph constraint and time sequence optimal transmission, which comprises the following steps: cutting the emergency description text to be processed into text blocks with preset lengths, and reserving an overlapping area between adjacent text blocks; atomic event extraction is carried out on each text block to form an atomic event group, and each atomic event is formed into a binary characteristic group ,Semantic content feature vectors for characterizing atomic events,A timestamp for characterizing the atomic event timing; any two atomic events in the atomic event group AndMapping to corresponding event nodes in a preset domain knowledge graphAndAnd calculates event nodesAndShortest path topological distance in the domain knowledge graph, constructing teacher model probability distribution according to the shortest path topological distance corresponding to any two atomic events, wherein the teacher model probability distribution is used for indicating the events according to human common knowledgeResulting in an eventIs a priori probabi