Portable Features for Classifying Emotional Text

2012 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 587–591, Montr´eal, Canada, June 3-8, 2012. c⃝2012 Association for Computational Linguistics Portable Features for Classifying Emotional Text Saif Mohammad National Research Council Canada Ottawa, Canada, K1A 0R6 saif.mohammad@nrc-cnrc.gc.ca Abstract Are word-level affect lexicons useful in de- tecting emotions at sentence level? Some prior research finds no gain over and above what is obtained with ngram features—arguably the most widely used features in text classifica- tion. Here, we experiment with two very dif- ferent emotion lexicons and show that even in supervised settings, an affect lexicon can pro- vide significant gains. We further show that while ngram features tend to be accurate, they are often unsuitable for use in new domains. On the other hand, affect lexicon features tend to generalize and produce better results than ngrams when applied to a new domain. 1 Introduction Automatically identifying emotions expressed in text has a number of applications, including track- ing customer satisfaction (Bougie et al., 2003), de- termining popularity of politicians and government policies (Mohammad and Yang, 2011), depression detection (Osgood and Walker, 1959; Pestian et al., 2008; Matykiewicz et al., 2009; Cherry et al., 2012), affect-based search (Mohammad, 2011), and improving human-computer interaction (Vel´asquez, 1997; Ravaja et al., 2006). Supervised methods for classifying emotions ex- pressed in a sentence tend to perform better than un- supervised ones. They use features such as unigrams and bigrams (Alm et al., 2005; Aman and Szpakow- icz, 2007; Neviarouskaya et al., 2009; Chaffar and Inkpen, 2011). For example, a system can learn that the word excruciating tends to occur in sentences la- beled with sadness, and use this word as a feature in classifying new sentences. Approaches that do not rely on supervised train- ing with sentence-level annotations often use affect lexicons. An affect lexicon, in its simplest form, is a list of words and associated emotions and senti- ments. For example, the word excruciating may be associated with the emotions of sadness and fear. Note that such lexicons are at best indicators of probable emotions, and that in any given sentence, the full context may suggest that a completely differ- ent emotion is being expressed. Therefore, it is un- clear how useful such word-level emotion lexicons are for detecting emotions and meanings expressed in sentences, especially since supervised systems re- lying on tens of thousands of unigrams and bigrams can produce results that are hard to surpass. For ex- ample, it is possible that classifiers can learn from unigram features alone that excruciating is associ- ated with sadness and fear. In this paper, we investigate whether word– emotion association lexicons can provide gains in addition to those already provided by ngram fea- tures. We conduct experiments with different affect lexicons and determine their usefulness in this ex- trinsic task. We also conduct experiments to deter- mine how portable the ngram features and the emo- tion lexicon features are to a new domain. 2 Affect Lexicons The WordNet …

Content truncated to 500 words