Comparison of Pun Detection Methods Using Japanese Pun Corpus

Comparison of Pun Detection Methods Using Japanese Pun Corpus Motoki Yatsu* and Kenji Araki** *Aoyama Gakuin University, 5-10-1 Fuchinobe, Chuo-ku, Sagamihara, Kanagawa, 252-5258, Japan yatsu@it.aoyama.ac.jp **Hokkaido University Kita 14, Nishi 9, Kita-ku, Sapporo, Hokkaido, 060-0814, Japan araki@ist.hokudai.ac.jp Abstract A sampling survey of typology and component ratio analysis in Japanese puns revealed that the type of Japanese pun that had the largest proportion was a pun type with two sound sequences, whose consonants are phonetically close to each other in the same sentence which includes the pun. Based on this finding, we constructed rules to detect pairs of phonetically similar sequences as features for a supervised machine learning classifier. Using these features in addition to Bag-of-Words features, an evaluation experiment confirmed the effectiveness of adding the rule-based features to the baseline. Keywords: corpus, Japanese, humour, puns, detection 1. Introduction In recent years, there has been growing interest in machines expressing and understanding humour. There is an interest in computational humour improving the quality of life (QOL) of the user, such as alleviating the psychological stress of the person expressing himself (Yamada et al., 2012), and improving depressive symptoms (Tsukawaki et al., 2011). Research to obtain such effects in artificial agent and robot communication is under way. For example, Dybala et al. (2012) showed that the user feels unexpected to the system and the desire to continue dialogue with the system is increased under the condition that the dialog agent generates humour. Miyazawa et al. (2012) conducted a survey of factors that enhance the continuity of the dialogue between the dialogue system and the user. As a result, it was shown that a humourous dialogue would get enhanced its continuity. In this paper, first we take a look at what types of Japanese pun are prominent from a categorization survey. Subsequently, we describe the rule-based and machine learning-oriented features as training data for a supervised Figure 1: Component ratios of the types of Japanese puns (italic) found in the pun corpus noted in Section 2.2. machine learning method, reflecting the phonological characteristics that Japanese puns share. 2. Japanese Puns as Text Humour 2.1 Types of Japanese Puns According to (Takizawa, 1995), Japanese puns can be classified mainly into two classes: juxtaposed puns and superposed puns. In a juxtaposed pun, a seed expression refers to one or more independent (noun, verb, adjective, or emoticon) morpheme(s) or phrase(s) in the sentence. A transformed expression refers to a phoneme string in an arbitrary section in a sentence having phonological similarity with seed expression(s). For example, the first sentence (Stc. 1) shown in Figure 1 have a pair of phoneme sequences phonologically similar, one of which has an extra ’t’ character. As another example, Stc. 2 has more than one pair of similar phoneme sequences. These kinds of puns are classified into juxtaposed puns. Kawahara and Shinohara (2012) define perfect puns as juxtaposed puns whose seed and transformed expressions have no phonological difference (as seen in stc. 2). Similarly, those that do not match due to …

Content truncated to 500 words