Research On Construction Method Of Domain Event Graph

The event graph is essentially a knowledge base of affair logic,describing the evolutionary rules and patterns between events.Event graph has important application value in the fields of event prediction,common sense reasoning,and business recommendation.Structurally,the event graph is defined as a directed and cyclic graph.The nodes in the graph represent events,and the edges represent the logical relationships between events,mainly including obedient relationship and causal relationship.Constructing the event graph includes two tasks,event extraction and event relationship recognition.In the process of constructing event graph,existing natural language processing pre-training and fine-tuning models only encode the relationships between words,and ignore the internal meaning of the words themselves.This thesis integrates the sememe information into existing deep learning models,and conducts research on the construction of event graph based on the sememe fusion method.Event extraction is to identify the triggering word of a event and judge the actor and recipient of the triggering word.Aiming at the problem of event extraction,this paper uses the sequence labeling method based on Bert.Sememe is defined as the smallest semantic unit,labeling the meaning of each word.This section creatively fuses the sememe information into pre-traning and fine?tuning model,propose BBSMF method based on sememe matrix decomposition and BBSP method based on sememe prediction.The main idea is to train the sememe task and event extraction task at the same time,the sememe task can adjusts the encoding result of Bert,so as to obtain better results on the problem of event extraction.In this part of the experiment,the Bert and Bert relation models are used as baseline,and the above methods are combined respectively.Both results shows the F1-score of fusion method is better than that of baseline.Event relationship identification is to judge the specific relationship of the extracted events.Aiming at the problem of event dependency relationship recognition,this paper adopts the sentence-level classification method based on Bert.Since the two above-mentioned fusion methods are designed for words,they are not applicable to this section,so this section proposes a sentence-level sememe fusion method BBSSP.The main idea is to predict the sememe in the sequence through the flag of sentence classification of Bert,so that the two sentence sequences with dependency relationship and reverse dependency relationship are more similar in the feature space,and they are unrelated to the two sentence sequences withoutrelationship.In this part of the experiment,Bert is used as baseline,above method is combined to Bert.The result shows the F1-score of sememe fusion method is improved compared to baseline.