Emotion Classification Using Massive Examples Extracted from the Web

Proceedings of the 22nd International Conference on Computational Linguistics (Coling 2008), pages 881–888 Manchester, August 2008 Emotion Classification Using Massive Examples Extracted from the Web Ryoko TOKUHISA†‡ †Toyota Central R&D Labs., INC. Nagakute Aichi JAPAN tokuhisa@mosk.tytlabs.co.jp Kentaro INUI‡ ‡Nara Institute of Science and Technology Ikoma Nara JAPAN {ryoko-t,inui,matsu}@is.naist.jp Yuji MATSUMOTO‡ Abstract In this paper, we propose a data-oriented method for inferring the emotion of a speaker conversing with a dialog system from the semantic content of an utterance. We first fully automatically obtain a huge collection of emotion-provoking event in- stances from the Web. With Japanese cho- sen as a target language, about 1.3 million emotion provoking event instances are ex- tracted using an emotion lexicon and lexi- cal patterns. We then decompose the emo- tion classification task into two sub-steps: sentiment polarity classification (coarse- grained emotion classification), and emo- tion classification (fine-grained emotion classification). For each subtask, the collection of emotion-proviking event in- stances is used as labelled examples to train a classifier. The results of our ex- periments indicate that our method signif- icantly outperforms the baseline method. We also find that compared with the single- step model, which applies the emotion classifier directly to inputs, our two-step model significantly reduces sentiment po- larity errors, which are considered fatal er- rors in real dialog applications. 1 Introduction Previous research into human-computer interac- tion has mostly focused on task-oriented dialogs, where the goal is considered to be to achieve a c⃝ 2008. Licensed under the Creative Commons Attribution-Noncommercial-Share Alike 3.0 Unported li- cense (http://creativecommons.org/licenses/by-nc-sa/3.0/). Some rights reserved. given task as precisely and efficiently as possi- ble by exchanging information required for the task through dialog (Allen et al., 1994, etc.). More recent research (Foster, 2007; Tokuhisa and Terashima, 2006, etc.), on the other hand, has been providing evidence for the importance of the af- fective or emotional aspect in a wider range of di- alogic contexts, which has been largely neglected in the context of task-oriented dialogs. A dialog system may be expected to serve, for example, as an active listening 1 partner of an el- derly user living alone who sometimes wishes to have a chat. In such a context, the dialog system is expected to understand the user’s emotions and sympathize with the user. For example, given an utterence I traveled far to get to the shop, but it was closed from the user, if the system could infer the user’s emotion behind it, it would know that it would be appropriate to say That’s too bad or That’s really disappointing. It can be easily imag- ined that such affective behaviors of a dialog sys- tem would be beneficial not only for active listen- ing but also for a wide variety of dialog purposes including even task-oriented dialogs. To be capable of generating sympathetic re- sponses, a dialog system needs a computational model that can infer the user’s emotion behind his/her utterence. There have been a range of stud- ies for building a model for classifying a …

Content truncated to 500 words