Research On Exploiting Rich Event Representation To Improve Event Causality Recognition

Event causality recognition is a new research field in natural language processing.Event causality identification is to judge whether there is a causal relationship between two events.At present,the existing research methods for identifying event causality mainly focus on combining basic neural network models,such as convolutional neural network models,with external causal knowledge,but lack of mining and using the rich semantic representation of event itself.However,events are more structured than plain text and have richer semantic information.We think that effective semantic representation of events can help recognize event causality.Therefore,we conduct an experimental study on how to use the semantic representation of events in the task of identifying Chinese event causality.Specifically,our work mainly includes the following three aspects:(1)A Chinese event causality identification model based on multi-column convolutional neural network,Multi-column Convolutional Bidirectional Gated Recurring Unit Nerual Networks,is designed and implemented.The Multi-column Convolutional Bidirectional Gated Recurring Unit Nerual Networks(MCBi-GRUs)have a convolutional layer and a bi-directional Gated Recurring Unit that can extract the local and global semantic features of text respectively.Simultaneously multicolumn neural networks acquire more multifaceted semantic features than singlecolumn neural networks.The experimental results show that the baseline model has a good classification ability under the condition of other relevant parameters being the same,which lays a foundation for the model construction in this paper.(2)In order to obtain the exact semantics contained in structured event text,an event semantic representation method based on tensor calculation is proposed.Based on the baseline system,the event combination tensor representation is added to obtain the semantic representation of the event by making full use of the information such as arguments and trigger.The experimental results show that the event causality recognition model with event tensor representation has better effect.Compared with the simple use of event text information,it can effectively use the structured information of event text.(3)In order to capture the evolution law of event pairs,interactive event representation and text-aware event representation are proposed.An interactive event representation is based on Interactive Attention and text-aware event representation is based on Gated-Attention.The model can take advantage of the interactive semantic representation between the two events and the contextual text that exists between the two events.The experimental results show that adding interaction information and discourse semantics between event pairs can significantly improve the effect of event causality recognition model.