Research And Application On Identification Of Chinese Event Factuality

Event factuality is the factual nature of the event in texts.It has been applied to many applications of Natural Language Processing(NLP),such as question answering system,information extraction and text semantic analysis,etc.At present,there is only a few research on identification of event factuality,especially in Chinese.This dissertation focuses on identifying Chinese event factuality and applying it to real systems as follows:(1)Identification of Chinese Event Factuality based on Factual Information.To mine factual information from lexical semantics,this dissertation proposes a feature-based approach to identify the Chinese event factuality which uses the sentence-level features,lexical-level features,predicate-level features,degree-level features,three-dimensional features and factuality features,etc.Experimental results on the Chinese event factuality corpus manifest that our approach improves the macro-F1 and micro-F1 by 5.80% and 3.35%,respectively,compared with a rule-based approach.(2)Identification of Chinese Event Factuality based on Convolutional Neural Networks.To identify event factuality from unlabeled texts,this dissertation proposes an effective CNN(Convolution Neural Networks)-based approach.It extracts the factual information from the event sentence and then regards them and their transformation as features.Meanwhile,it transfers these features to word vectors to construct the sentence-level word vector map.Finally,it inputs the word vector map to the CNN model to identify the event factuality.Experimental results on the Chinese event factuality corpus manifest that our approach improves the macro-F1 and micro-F1 by 12.59% and 1.11%,respectively.These results tackle the imbalanced data distribution problem in identification of event factuality.(3)A Stock Comment Analysis System based on Event Factuality.This dissertation applies the approach of event factuality identification to analyze the stock comments.This system has five functional modules,i.e.,the stock information crawling and processing module,the factual information extraction module,the features transformation module,the stock comment analysis module and performance analysis module.This dissertation proposes two effective approaches on identifying Chinese event factuality and applies the approaches to stock comments analysis.It demonstrates great value of research and application.