Contextual Valence Shifters Supporting Affect Analysis of Utterances in Japanese Michal Ptaszynski Pawel Dybala Rafal Rzepka Kenji Araki Graduate School of Information Science and Technology, Hokkaido University {ptaszynski, paweldybala, kabura, araki}@media.eng.hokudai.ac.jp Abstract: The paper presents a support method for affect analysis of utterances in Japanese. One of the problems in the system for affect analysis developed by us before was confusing the valence of emotion types in the final stage of analysis. The cause of this problem was extracting from the utterance only the emotive expression keyword without its grammatical context. To solve this problem we enhance the emotion types extraction procedure in the baseline system with grammatical analysis using Contextual Valence Shifters (CVS). CVS are words, or phrases such as "not", "very much" "not quite", which determine the semantic orientation of the valence of emotive expressions. Keywords: Affect Analysis, Contextual Valence Shifters. 1. Introduction Research in the field of Affective Computing has been gathering popularity of researchers since being initiated only a little over ten years ago [1]. The interest in such research is usually focused on recognizing the emotions of users in human-computer interaction. In the most popular methods the emotions are recognized from: facial expressions [2], voice [3] or biometric data [4]. However, these methods, usually based on behavioral approach, ignore the semantic context of emotions. Therefore, although achieving good results in laboratory, such methods become useless in real life. A system for recognition of emotions from facial expressions, assigning “sadness” when a user is crying would be critically mistaken, if the user was e.g. cutting an onion in the kitchen. This led to formation of Affect Analysis - a field focused on developing natural language processing techniques for estimating the emotive aspect of text. There were several attempts to achieve this goal for the Japanese language. For example, Tsuchiya et al [5] tried to estimate emotive aspect of utterances with a use of association mechanism. On the other hand, Tokuhisa et al [6] used a large number of examples from the Web. However, none of the present methods is capable to perform a deep contextual analysis. Ptaszynski et al [7] proposed a pioneer method for Affect Analysis of utterances basing on the idea of two-step classification of emotive content – general emotiveness and specific emotion types. However, one of the problems with this method was confusing the valence polarity of emotive expressions in the last step of analysis. To solve this problem and to push Ptaszynski’s method one step towards a deeper contextual analysis of emotive content we decided to apply the idea of Contextual Valence Shifters to the baseline system to enhance the specific emotion types determination. 2. Contextual Valence Shifters The idea of using Contextual Valence Shifters (CVS) in Sentiment Analysis has been first proposed by Polanyi and Zaenen [8]. They distinguish two kinds of CVS: negations and intensifiers. The group of negations contains words and phrases like “not”, “never”, and “not quite”, which change the valence polarity of semantic orientation of an evaluative word they stick to. The group of intensifiers contains words like “very”, “very much”, and “deeply”, which intensify the semantic orientation of an evaluative word. So far the idea of CVS analysis was successfully applied to the field of Sentiment Analysis of texts in English [9]. A few attempts on Japanese ground [10] show that it is also applicable for the Japanese language. Examples of negations in the Japanese language are grammatical structures such as: -nai 1 (not-), amari -nai (not quite-), mattaku -nai (not at all-), or sukoshi mo -nai (not even a bit-). Intensifiers are represented by such grammatical structures as: totemo- (very much-), sugoku- (-a lot), or kiwamete- (extremely). However, till now there were no attempts to apply CVS in the field of Affect Analysis neither in English nor in Japanese. This paper presents the first pioneer attempt of that kind. 3. Definition and Classification of Emotions Nakamura [11] defines emotions as every temporary state of mind, feeling or emotional state evoked by experiencing different sensations. This definition is complemented by Beijer’s [12] definition of emotive utterances, which he describes as every utterance in which the speaker in question is emotionally involved, and this involvement is expressed linguistically. 1 In this paper we use italic for Japanese expressions. － 825 －