Study On Emotion Cause Detection From Chinese Text

With the rapid development of Internet and social media,a large number of subjective texts,which are published by users,are flooding the Internet.These texts cover the users' product review and the responses to the hot social incidents,and so on.Thus,opinion mining,public opinion analysis and text emotion analysis in the field of nature language processing attract the concerns of a lot of researchers.Currently,the exiting works on text emotion computing majorly focus on emotion classification,but lack the deeply research on the emotion cause analysis.Thus,in this thesis,we focus on detecting the reason from the text which triggers the generation of emotion.The existing works on text emotion analysis may be camped into rule-based methods and statistical-based methods.The former requires manually constructed rule bases,but the completement of the rule base is hard to guarantee.The latter normally employs lexical and phrasal features related to emotion expression.Its performance relies on large annotated corpus.Target to the shortages of present methods,we investigate emotion cause detection methods based on memory networks and multi-kernel support vector machine,respectively.Based on these,we further investigate an ensemble learning based method to incorporate these two methods for improving the performance of emotion cause detection.The main work of this thesis includes two parts:Being aimed at the lack of consideration of the relationship between the emotion keywords and their corresponding causes,we propose emotion cause detection method based memory networks with attention mechanism.By encoding the relationship between the input emotion keywords and candidate emotion causes with stacking computation of attention,this method may local the core words of the emotion causes using the attention mechanism and next identify the clause with emotion cause.The experimental results on EMNLP 2016 dataset show that this method achieves 69.55% on F score,which outperforms the baseline method by 1.99%.Considering that the memory networks encodes the emotion-related cause-effect relationship in text but neglects the textual features to emotion cause expressions,we extract the sequential features of emotion cause expression and the syntactic tree-structural features of emotion cause.We then designed a multi-kernel support vector machines based emotion cause detection method by incorporating polynomial kernel function and convolutional kernel function.On this basis,we design ensemble learning based method,which uses the memory networks based method and the multi-kernel support vector machine based method as base classifiers with bagging based ensemble.The experimental results on EMNLP 2016 dataset show that the ensemble method outperforms the baseline by7.69% F score.To best of our knowledge,this method achieves the highest performance on this dataset.