Research On Chinese Event Causality Extraction Based On Pre-trained Model

Event causality extraction is an important and challenging task in the field of information extraction,which aims to extract events from unstructured text and identify event pairs with causal relationships.Event causality extraction is the foundation of numerous artificial intelligence applications,and plays an important role in automatic question answering,knowledge graph construction,event prediction and other related fields.While the advent of pre-trained model technology has propelled the rapid advancement of information extraction techniques in recent years,the investigation into event causality extraction remains insufficient.Event causality extraction includes two steps: event extraction and event causality identification.Existing methods in event extraction are suffer from the uneven distribution of labeled data,resulting in low extraction accuracy.Additionally,in the research of event causality identification,there are problems of scarcity of labeled data and lack of common-sense knowledge.To overcome these challenges,this thesis conducts research on the two key steps of event causality extraction based on the BERT pre-trained model.The contents of this thesis are outlined as follows:(1)To address the problem of the low accuracy of existing Chinese event extraction methods,this thesis proposes a Chinese event extraction method based on BERT and label semantic extension.In this method,the semantic information of the event type label is firstly extended to obtain the semantic extension words of the label,and proposes an index to measure the quality of the extension words,which is used to screen the extension words.Second,the extended words and the original text are concatenated as the input of the model,and the semantic association between the extended words and the event triggers is used to enhance the ability of the model to recognize events.Finally,based on the BERT pre-trained model,a joint extraction framework that can simultaneously extract event triggers and event arguments is constructed.This framework eliminates the transmission error problem caused by the pipeline architecture and can effectively improve the accuracy.The experimental results show that the method proposed in this thesis outperforms the existing algorithms on the Chinese event extraction dataset.Meanwhile,ablation experiments and example studies demonstrate the effectiveness of the label semantic extension method.(2)To address the problems of scarcity of labeled data and lack of common-sense knowledge in Chinese event causality identification,this thesis proposes a Chinese event causality identification method based on commonsense knowledge enhancement.Based on the pre-trained BERT model,this method employs the prompt learning paradigm to stimulate the pre-training model’s few-shot learning ability and overcome the scarcity of labeled data.Additionally,by incorporating commonsense knowledge into the prompt template,the method addresses the lack of commonsense knowledge and effectively improves the causality reasoning of the model.Experiments on multiple public Chinese and English event causality datasets show that this method can effectively improve the accuracy of event causality identification,and is superior to existing algorithms in terms of generalization and few-shot learning ability.