Ranking Reader Emotions Using Pairwise Loss Minimization and Emotional Distribution Regression

Proceedings of the 2008 Conference on Empirical Methods in Natural Language Processing, pages 136–144, Honolulu, October 2008. c⃝2008 Association for Computational Linguistics Ranking Reader Emotions Using Pairwise Loss Minimization and Emotional Distribution Regression Kevin Hsin-Yih Lin and Hsin-Hsi Chen Department of Computer Science and Information Engineering National Taiwan University No. 1 Roosevelt Rd. Sec. 4, Taipei, Taiwan {f93141, hhchen}@csie.ntu.edu.tw Abstract This paper presents two approaches to ranking reader emotions of documents. Past studies assign a document to a single emotion cate- gory, so their methods cannot be applied di- rectly to the emotion ranking problem. Furthermore, whereas previous research ana- lyzes emotions from the writer’s perspective, this work examines readers’ emotional states. The first approach proposed in this paper minimizes pairwise ranking errors. In the sec- ond approach, regression is used to model emotional distributions. Experiment results show that the regression method is more ef- fective at identifying the most popular emo- tion, but the pairwise loss minimization method produces ranked lists of emotions that have better correlations with the correct lists. 1 Introduction Emotion analysis is an increasingly popular re- search topic due to the emergence of large-scale emotion data on the web. Previous work primarily studies emotional contents of texts from the writer's perspective, where it is typically assumed that a writer expresses only a single emotion in a document. Unfortunately, this premise does not hold when analyzing a document from the reader's perspective, because readers rarely agree unani- mously on the emotion that a document instills. Figure 1 illustrates this phenomenon. In the figure, 0% 10% 20% 30% 40% Heartwarming Happy Sad Surprising Angry Boring Awesome Useful Emotion % of Readers Figure 1. Emotional responses of 626 people after read- ing a Yahoo! News article about an Iranian refugee mother and her two children who finally reunited with their family in the March of 2007 after been stranded in a Moscow airport for 10 months due to false passports. readers’ responses are distributed among different emotion categories. In fact, none of the emotions in Figure 1 has a majority (i.e., more than 50%) of the votes. Intuitively, it is better to provide a ranking of emotions according to their popularity rather than associating a single reader emotion with a document. As a result, current writer-emotion analysis techniques for classifying a document into a single emotion category are not suitable for ana- lyzing reader emotions. New methods capable of ranking emotions are required. Reader-emotion analysis has potential applica- tions that differ from those of writer-emotion analysis. For example, by integrating emotion ranking into information retrieval, users will be able to retrieve documents that contain relevant contents and at the same time produce desired feel- ings. In addition, reader-emotion analysis can as- sist writers in foreseeing how their work will influence readers emotionally. 136 In this paper, we present two approaches to ranking reader emotions. The first approach is in- spired by the success of the pairwise loss minimi- zation framework used in information retrieval to rank documents. Along …

Content truncated to 500 words