Research On Calculation Metods Of Document-level Event Factuality

In natural language,event is a basic semantic unit and a basic element of document composition.Event factuality describes whether the event is certain,negative or uncertain,and expresses people's attitudes or opinions on the content of the event,that is,the degree of truth of the event in the language text.Event factuality recognition is an important foundation of question answering system,text understanding and many other related tasks.At present,most of the researches on event factuality recognition are still at the sentence level,and there is almost no related research on document level event factuality.Therefore,this paper mainly focuses on the construction and recognition of document level event factuality,which includes the following three aspects:(1)The construction of document level event factuality corpusAiming at the problem that most of the related research is only at the sentence level,this paper constructs a document-level event factuality corpus DLEF(Document-level Event Factuality Identification)to promote the development of related research.This paper selects 4650 Chinese articles and 1730 English articles from China Daily and Sina Bilingual News.First of all,the corresponding articles are obtained on the website for word segmentation and other preprocessing;then,each article is labeled with one of the core events,including sentence level event factuality and document level event factuality;finally,the labeled corpus is statistically analyzed.(2)A method of document level event factuality recognition based on gated convolution networkAt present,there are few researches on document level event factuality recognition.In this paper,we proposes DEFI(Document-level Event Factuality Identification),a method based on gated convolutional network,to identify the factuality of a document level event.This method first extracts the semantic and syntactic information of events from sentences and syntactic paths using Gated Convolution Neural Network(GCNN),and then obtains the feature representation of each sequence's more important overall information relative to itself through self attention layer,so as to identify the document level event factuality.Experiments on Chinese and English corpus show that the performance of DEFI on macro average F1 and micro average F1 is better than that of reference system.(3)Joint recognition method of event sentence and document level event factuality based on BERTAt present,the existing research on the factuality identification of document level events is based on the known conditions of events and event sentences.To solve this problem,this paper transforms the problem of document level event factuality recognition into a problem of machine reading:first of all,it discusses the issue of event factuality by reading the article;then it finds out the event sentence and judges the document level event factuality according to the factuality question.It not only saves the workload of finding event sentences,but also retains the influence of all sentences on the factuality of document level event.Finally,multi task learning is used to identify event sentences and event factuality at the document level.Experiments on Chinese corpus show that compared with the reference system,this method improves both macro average F1 and micro average F1.Aiming at the problem of document level event factuality recognition,this paper puts forward an effective solution and achieves good performance,which will provide reference for further study of document level event factuality recognition.