Build Chinese Emotion Lexicons Using A Graph-based Algorithm and Multiple Resources

Proceedings of the 23rd International Conference on Computational Linguistics (Coling 2010), pages 1209–1217, Beijing, August 2010 Build Chinese Emotion Lexicons Using A Graph-based Algorithm and Multiple Resources Ge Xu, Xinfan Meng, Houfeng Wang Key Laboratory of Computational Linguistics (Peking University), Ministry of Education Institute of Computational Linguistics, Peking University {xuge, mxf, wanghf}@pku.edu.cn Abstract For sentiment analysis, lexicons play an important role in many related tasks. In this paper, aiming to build Chinese emo- tion lexicons for public use, we adopted a graph-based algorithm which ranks words according to a few seed emotion words. The ranking algorithm exploits the simi- larity between words, and uses multiple similarity metrics which can be derived from dictionaries, unlabeled corpora or heuristic rules. To evaluate the adopted algorithm and resources, two independent judges were asked to label the top words of ranking list. It is observed that noise is almost un- avoidable due to imprecise similarity met- rics between words. So, to guarantee the quality of emotion lexicons, we use an iterative feedback to combine man- ual labeling and the automatic ranking al- gorithm above. We also compared our newly constructed Chinese emotion lexi- cons (happiness, anger, sadness, fear and surprise) with existing counterparts, and related analysis is offered. 1 Introduction Emotion lexicons have a great impact on the re- sults of related tasks. With high-quality emo- tion lexicons, systems using simple methods can achieve competitive performance. However, to manually build an emotion lexicon is time- consuming. Many research works in building lex- icons use automatic methods to assist the building procedure. Such works commonly rank words by the similarities to a set of seed words, then those words with high ranking scores are more likely to be added to the final lexicons or used as additional seed words. For Chinese, emotion lexicons are scarce re- sources. We can get a small set of emotion words from semantic dictionary (such as CCD, HowNet, synonym dictionaries) or directly from related pa- pers (Xu and Tao, 2003) (Chen et al. , 2009), but it is often not sufficient for practical systems. Xu et al. (2008) constructed a large-scale emotion on- tology dictionary, but it is not publicly available yet. In this paper, we adopted a graph-based algo- rithm to automatically rank words according to a few seed words. Similarity between words can be utilized and multiple resources are used to boost performance. Combining manual labeling with automatic ranking through an iterative feedback framework, we can produce high-quality emotion lexicons. Our experiments focused on Chinese, but the method is applicable to any other language as long as suitable resources exist. The remainder of this paper is organized as fol- lows. In Section 2, related works are introduced. In Section 3, we describe a graph-based algorithm and how to incorporate multiple resources. Sec- tion 4 gives the details of applying the algorithm on five emotions and shows how to evaluate the re- sults. Section 5 focuses on how to build and evalu- ate emotion lexicons, linguistic consideration and instruction …

Content truncated to 500 words