Research Of Emotion Analysis On Heterogeneous Texts

Emotion analysis is a type of fine-grained sentiment analysis task.The machine learning based emotion analysis approach aims to train machine a model to automatically annotate the input text with the pre-defined emotion category(i.e.,emotion classification),or the emotion score(i.e.,emotion regression).The heterogeneous text denotes a set of texts which contains textual resources in different languages,domains,or annotation systems.Traditional emotion analysis task needs a large-scale corpus in single language,domain,and annotation system to train the model for better performance,while this kind of corpus is often hard to get.Thus,the research of modeling the relationship between heterogeneous texts,and leveraging large-scale heterogeneous texts for model training is always a research focus in emotion analysis.In this paper,we focus on the research of emotion analysis on heterogeneous texts which aims to model the relationship between heterogeneous texts,and to leverage the heterogeneous texts to train better emotion analysis models.Specifically,our work in this paper is threefold:(1)Research of cross-corpus emotion classificationThe existing emotion classification corpora often use different classification systems and annotation guideline due to the lack of an admitted standard.While different emotion classification systems include different emotion categories in both types and numbers,it is hard to directly leverage the resources in one corpus to another one.To address this problem,in this paper,we propose a new task namely corpus fusion which aims to leverage resources in various corpora,and introduce integer linear programming to address this task.Specifically,we first construct the relationship between two emotion classification systems of two corpora,and then train two classifier on two emotion classification corpora independently.Finally,we optimize the results of two classifier on the same testing set through integer linear propagation we the pre-defined constraints which describe the relationship between two emotion classification systems.Empirical study shows that our proposed approach can notably improve the classification accuracy comparing with models trained on a single corpus.(2)Research of cross-lingual semi-supervised emotion classificationAnnotated large-scale lingual resources in a rich-resource language(e.g.,English)are easy to acquire,while such resources in a low-resource language(e.g.,Chinese)are hard to obtain.Thus the research of leveraging multi-lingual resources for model training is always a hotspot in emotion analysis.In this paper,we propose an approach to address the cross-lingual emotion classification with an adversarial neural network.Our approach introduces adversarial learning to the neural network in order to leverage both the rich-resource English annotated corpus and Chinese unannotated corpus to help the emotion classification task on the low-resource Chinese benchmark dataset.Empirical study shows that our proposed approach can notably outperformance not only the traditional supervised and semi-supervised approaches,but also several state-of-the-art cross-lingual approaches on two Chinese benchmark datasets.(3)Research of cross-domain multi-dimensional emotion regressionComparing with emotion classification,emotion regression task is more suitable for conducting fine-grained emotion analysis.Emotion regression aims to rate an input text with emotion scores in three dimensions,i.e.,Valence,Arousal,and Dominance.However,existing researches on emotion regression mostly focus on one dimension,or independently train multiple models for different dimensions.To address the problem,in this paper,we propose an approach to multi-dimensional emotion regression task based on an adversarial neural network.The proposed approach can not only learn more dimension-specific features in the text through adversarial learning,but also leverage the lingual resources in other domains to improve the regression performance.Empirical study shows that our proposed approach can notably improve the regression performances in all emotion dimensions compared with several state-of-the-art regression baselines on a large-scale multi-domain emotion regression corpus.