Research On Cross-lingual Textual Emotion Cause Detection

Most existing works on sentiment analysis focuses on the text sentiment or emotion classification,but there is few research on the textual emotion cause detection.Emotion cause detection research mainly focuses on the method for automatically identifying the factors or events in the text that trigger the emotions of groups or individuals.Most existing emotion cause detection methods may be camped into three major categories,namely rule based approach,statistical machine learning based approach and deep learning based approach.Rule based approach has the shortcoming of low coverage.Statistical machine learning based approach require a lot of artificial works on statistical features selection,which is time-consuming.Deep learning based approach is always puzzled by the lack of large-scale annotated data.To address these problems,this paper studies the methods to improve the performance of emotion cause detection from the perspectives of mono-lingual emotion cause detection and cross-lingual emotion cause detection,respectively.In view of the existing methods always ignoring the relation between emotion expression and emotion cause,this paper investigates an emotion cause detection method by incorporating gated recurrent unit,convolutional neural network and hierarchical attention mechanism.This method employs the gated recurrent unit and convolutional neural network to capture the semantic information of text sequence blocks,and incorporates the attention mechanism to model the relation between emotion expression and emotion cause.The experimental results on EMNLP2016 dataset show that the F1 performance of this method improves the memory network based method for 0.4%.Considering the small scale of available emotion cause detection annotated data,this paper studies the cross-lingual emotion cause detection approach.This approach aims to transfer the rich annotated data on the source language to the target language with poor resources for improving the performance of emotion cause detection.This paper investigates a cross-lingual emotion cause detection approach based on iterative method to augment data.This method obtains candidate training data on target language through machine translation and classifies the candidate training data by using the trained emotion cause classifier.The classifier is then re-trained by using the original data and new correctly predicted data.Next,the classifier is applied to the wrongly predicted data.Such data augmentation and classifier training are carried out iteratively to improve the performance of emotion cause detection.The experimental results on EMNLP2016 dataset show that the F1 performance of this method improves the memory network based method for 1.21%.Considering that the performance of cross-lingual emotion cause detection approach based on iterative method to augment data is influenced by the machine translation quality,this paper further investigates the cross-lingual emotion cause detection method based on adversarial training.This method extract the language-independent and task-related features through discriminator against game and gradient back-propagation of emotion cause classifier.The experimental results show that the F1 value of this approach is further increased for 1.29%,comparing to cross-lingual emotion cause detection method based on the iterative method to augment data.It achieves the known top performance on EMNLP2016 emotion cause dataset.