Research On End-to-end Sentence-level Event Factuality Identification

Event factuality refers to the factuality of events in the text for the event source.At present,event factuality identification is lack of research on raw corpus,and only considers the event factuality under the default source.Therefore,this dissertation studies the end-toend sentence-level event factuality identification to identify the events,event sources and corresponding event factuality in the raw text,which has more practical significance.Firstly,most of the existing researches on event factuality rely on annotation information,do not take into account the performance of event factuality identification in raw corpus,and ignore the impact of event source on event factuality.To solve these problems,this dissertation proposes an end-to-end event factuality identification method based on hybrid model.This method fuses the information of events,event sources,cues and related dependent path,and uses the hybrid model BiLSTM+GCN to identify the event factuality from end-to-end.Experiments show that this method is superior to the benchmark system,and Micro-F1 and Macro-F1 are improved by 1.79%and 8.67%respectively.Secondly,because the process of the method based on hybrid model is complex,timeconsuming and serious cascading error.Aiming at these problems,this dissertation proposes the method of joint model based on GCN,which aims to capture the relationship between events,event sources and event factuality.At the same time,aiming at the problem of polysemy of event,this dissertation proposes to use BERT to encode sentences,supplemented by linguistic features(part of speech,dependency,lemma)to strengthen the semantic representation of words.This dissertation GCN uses GCN to capture the syntactic and semantic features of events,and integrate the event and event source information for event factuality identification.Experiments show that this method is superior to the benchmark system,and micro-F1 and Macro-F1 are improved by 2.31%and 1.44%respectively.Finally,the lack of expression of English data directly leads to the problem that it is difficult for the model to correctly identify the factuality of events.this dissertation proposes to use cross-lingual information to improve the performance of the event factuality model.In order to explore the impact of cross-lingual information on event factuality,this dissertation proposes to use the methods based on cross-lingual word vector,cross-lingual data augmentation and cross-lingual features for end-to-end event factuality identification.The experimental results show that cross-lingual information can effectively improve the performance of the model,and the method based on cross-lingual features is the most effective.Compared with the benchmark model,micro-F 1 and Macro-F 1 are improved by 4.11%and 8.32%respectively.Aiming at the problems existing in the research of end-to-end sentence level event factuality identification,this dissertation puts forward corresponding solutions to improve the performance of the model,which is helpful to the further research of end-to-end event factuality identification.