Research On Causal Relation Extraction Based On Deep Learning And Graph Attention Networks

Causal relationship is an important relationship type in natural language texts and plays a vital role in many fields such as relational reasoning.Therefore,extracting causal relationship is a basic task in text mining.However,causal relationship extraction is still a new field at present without mature research system and publicly evaluated dataset.The existing research results all have their own research system and cannot be compared horizontally,which is one of the key factors that hinder the progress of causal relationship extraction.Moreover,the existing related concepts,research methods,datasets and labeling methods are scattered in studies without systematic summary,which makes it impossible for researchers to obtain a complete picture of the field and conduct in-depth research.Based on the comprehensive summary and multi-angle innovation of these aspects,we propose a new research system for causal relation extraction,which can serve as a reference for the follow-up researches.The existing researches mainly use the three methods of text classification,relation extraction and sequence labeling to explore the causality in text.We focuses on the sequence labeling mothed to extract causal entity and identify the direction of causality,without relying on the feature engineering or causal background knowledge.For the other two methods,we make a supplementary exploration.The main contributions of this paper can be summarized as follows:(1)Based on the basic concept,we systematically summarizes the related concepts and the types of causality scattered in each research;(2)In terms of research methods,we summarizes three commonly used research fields,and comprehensively summarizes other research methods,so that readers can understand the overall picture of this field;(3)As for the algorithm model,we extend syntactic dependency tree to the syntactic dependency graph,adopt graph attention networks to natural language processing,and introduce the concept of S-GAT(graph attention networks base on syntactic dependency graph).Combining the deep learning model and S-AGT,the Bi-LSTM+CRF+S-GAT model for causality extraction is proposed,which generates causal label of each word in the sentence according to the input word vectors;(4)In terms of dataset,we comprehensively summarize the existing causal datasets,exploring the practicability and extensibility from multiple perspectives.The SemEval dataset is modified and extended,rules are made to re-label experimental data according to the defects,proposing a new causality extraction dataset ESC(Extended SemEval based on causality);(5)The existing labeling methods of causal sequence are summarized,concluding the labeling criteria,and propose the causal labeling method of “core word”.In view of the labeling controversy,multiple candidate labeling sequences are set,the dataset E-SCIFI(extended SCIFI)is proposed to explore the optimal causal labeling method through experiments.The experimental results show that the labeling method of “causal core word” is the best in experimental results.(6)Extensive experiments are conducted on the ESC,which shows that our model achieves 6.4% improvement over the-state-of-art model Bi-LSTM+CRF+self-ATT in terms of prediction accuracy.In addition,based on the two classification research methods,supplementary experiments are conducted on Altlex and SemEval with the expanded model to comprehensively explore causality,conducting a comprehensive study of causality extraction and the extensibility of the model we proposed.