Causal Relation Extraction Based On Reinforcement Learning And Large Language Representation Models

The research on causality has a long history,but the research on causality extraction in natural language processing is still an emerging field.At first,causality extraction was only one of the relation extraction tasks,but as the importance of causality extraction tasks gradually emerged,more and more researches have been devoted to the causal relationship extraction task.As a relatively new research,we often encounter the problem that there is insufficient datasets,and the solution is to use meta-learning or data augmentation methods.Most of the current data enhancement methods use a set of rule-based processing methods,which do not make targeted enhancements for different tasks.This paper summarizes the existing data augmentation methods and proposes a novel task-adaptive data augmentation model based on adversarial training.On the other hand,in the task of causality extraction,the identification of implicit causality has always been a major obstacle in research.Due to its diverse manifestations in different corpora,few studies have made progress on implicit causality.The adaptive adversarial model proposed in this paper has another effect.First the ACMM model can specifically select the words that are more important for causality.And it is proved by experiments that in sentences with explicit causal connectives,the ACMM model has a higher probability of removing causal connectives.Therefore,we believe that in the implicit causal relationship without obvious causal connectives,the words or phrases selected by the model with a higher probability are more likely to be implicit causal connectives.Through this study,implicit causality can be viewed from another perspective,which can provide a reference for subsequent research on implicit causality.Experiments on two public datasets show that the model with adversarial training has better performance than the original model,which proves the effectiveness of this model.The comparative experiments with different word deletion strategies show that the strategy generated by the ACMM model is better and has higher stability.We also demonstrate the ability of the model to recognize causal connectives by visualizing the output of the model,and our model also has some ability of discriminating implicit causal connectives.