A. Farzindar and V. Keselj (Eds.): Canadian AI 2010, LNAI 6085, pp. 40–50, 2010. © Springer-Verlag Berlin Heidelberg 2010 Hierarchical Approach to Emotion Recognition and Classification in Texts Diman Ghazi 1 , Diana Inkpen 1 , and Stan Szpakowicz 1,2 1 School of Information Technology and Engineering, University of Ottawa 2 Institute of Computer Science, Polish Academy of Sciences {dghaz038,diana,szpak}@site.uottawa.ca Abstract. We explore the task of automatic classification of texts by the emotions expressed. We consider how the presence of neutral instances affects the performance of distinguishing between emotions. Another facet of the evaluation concerns the relation between polarity and emotions. We apply a novel approach which arranges neutrality, polarity and emotions hierarchically. This method significantly outperforms the corresponding “flat” approach which does not take into account the hierarchical information. We also compare corpus-based and lexical-based feature sets and we choose the most appropriate set of features to be used in our hierarchical classification experiments. Keywords: Sentiment analysis, emotion in text, emotion recognition, text classification, hierarchical classification. 1 Introduction In recent years there has been a growing interest in automatic identification and extraction of opinions, emotions, and sentiment in text. Motivations for this task include the desire to provide tools for information analysts in government, commercial, and political domains, who want to automatically track attitudes and feelings in on-line forums [1]. Emotions, an important element of human nature, have also been widely studied in psychology and behavioral sciences. They have also attracted the attention of researchers in computer science and particularly in computational linguistics. This paper looks at the categorization of a sentence into six basic emotions (defined as emotions with universally accepted distinctive facial expressions). Those are happiness, sadness, fear, anger, disgust, and surprise [2]. We added a class non- emotional for the sentences which bear no emotion. These seven classes are common in much of the previous work [3, 4, 5, 6]. There has been progress in research on polarity and sentiment analysis, but little work has been done in automatic recognition of emotion in text. We assume that emotions carried by a sentence are not independent of their polarity; therefore we try to find a link between them and we want to apply classification methods, which consider these connections.