Corpus Fusion for Emotion Classification

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 3287–3297, Osaka, Japan, December 11-17 2016. Corpus Fusion for Emotion Classification Suyang Zhu1, Shoushan Li1∗, Ying Chen2, Guodong Zhou1 1Natural Language Processing Lab, School of Computer Science and Technology, Soochow University, China 2College of Information and Electrical Engineering, China Agricultural University syzhu@stu.suda.edu.cn, lishoushan@suda.edu.cn, chenying@cau.edu.cn, gdzhou@suda.edu.cn Abstract Machine learning-based methods have obtained great progress on emotion classification. How- ever, in most previous studies, the models are learned based on a single corpus which often suf- fers from insufficient labeled data. In this paper, we propose a corpus fusion approach to address emotion classification across two corpora which use different emotion taxonomies. The objective of this approach is to utilize the annotated data from one corpus to help the emotion classifica- tion on another corpus. An Integer Linear Programming (ILP) optimization is proposed to refine the classification results. Empirical studies show the effectiveness of the proposed approach to corpus fusion for emotion classification. 1 Introduction Emotion classification aims to recognize human emotions, such as joy, anger or surprise in a given text. Emotion classification has a variety of applications including online chatting (Galik and Rank, 2012), news classification (Liu et al., 2013) and stock marketing (Bollen et al., 2011). In recent years, emotion classification in social media has been greatly popular in the Natural Language Processing (NLP) community (Chen et al., 2010; Purver and Battersby, 2012; Li et al., 2015). Because of the popularity of social media today, the analysis of short text on social media becomes more important (Kiritchenko et al., 2014; Wen and Wan, 2014; Wang et al., 2015). Users express their feelings and emotions on various social media platforms. Existing emotion classification approaches are based on corpus classification methods where human- annotated emotion corpora are leveraged to train a machine learning-based emotion classification models. Recently, several different emotion corpora have been proposed by different researchers, such as Yao et al. (2014) and Huang et al. (2015). However, the size of each existing labeled corpus might be rather limited due to the high cost of data annotation, which results low performance in traditional supervised emotion classification. In this paper, we propose a novel task, namely corpus fusion for emotion classification, which aims to leverage the data from different emotion corpora so as to alleviate the data deficiency problem. This task is motivated by the fact that although the emotion corpora are from different resources, they have the same objective of emotion classification. So it is easy to enlarge the size of the corpora by mixing the data of the same emotion category from two corpora. However, corpus fusion for emotion classification is challenging due to the following two factors: First, the emotion taxonomies are often different between two emotion corpora because of the lack of an accepted standard. As a result, similar instances which express similar or same emotion can be catego- rized into different types of emotions under different taxonomies. An example from two emotion …

Content truncated to 500 words