Predicting the Evocation Relation between Lexicalized Concepts

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 1657–1668, Osaka, Japan, December 11-17 2016. Predicting the Evocation Relation between Lexicalized Concepts Yoshihiko Hayashi Faculty of Science and Engineering, Waseda University 2-4-12 Ohkubo, Shinjuku, Tokyo 169-0072, Japan yshk.hayashi@aoni.waseda.jp Abstract Evocation is a directed yet weighted semantic relationship between lexicalized concepts. Al- though evocation relations are considered potentially useful in several semantic NLP tasks, the prediction of the evocation relation between an arbitrary pair of concepts remains difficult, since evocation relationships cover a broader range of semantic relations rooted in human perception and experience. This paper presents a supervised learning approach to predict the strength (by regression) and to determine the directionality (by classification) of the evocation relation that might hold between a pair of lexicalized concepts. Empirical results that were obtained by in- vestigating useful features are shown, indicating that a combination of the proposed features largely outperformed individual baselines, and also suggesting that semantic relational vectors computed from existing semantic vectors for lexicalized concepts were indeed effective for both the prediction of strength and the determination of directionality. 1 Introduction Evocation, defined as the extent to which one concept (the source concept, s) brings to mind another (the target concept, t), is a directed yet weighted semantic relationship between semantic units (Boyd-Graber et al., 2006). As in the previous work (Boyd-Graber et al., 2006; Ma, 2013), the present work also considers evocation to be a semantic relationship between lexicalized concepts, rather than a relation between words. The weight of an evocation relation instance should measure the strength of the directed association from s to t, which we cannot directly observe nor compute from corpora. Although evocation relations are potentially useful in several semantic NLP tasks, such as the mea- surement of textual similarity/relatedness and the lexical chaining in discourse, the prediction of the evo- cation relation between an arbitrary pair of concepts remains more difficult than measuring conventional similarities (synonymy, as well as hyponymy/hypernymy) or relatednesses (further including antonymy, meronymy/holonymy, as well as predicate-argument relations and maybe more), since evocation rela- tionships cover a far broader range of semantic relationships (Cramer, 2008). Besides, as Ma (2013) argues, some types of evocation relations might be rooted in human perception and experience, implying that the acquisition of evocation relations solely from textual corpora is rather difficult, as they are the outcome of already accomplished activities of language production. To the best of our knowledge, this paper is the first to present a supervised learning approach to predict the strength of an evocation as a regression task, and to determine the directionality as a classification task. We utilize Princeton WordNet (PWN) (Fellbaum, 1998) as the inventory of lexicalized concepts (synsets). Our empirical results show that combining a range of features is effective as intended, and that the semantic relational vectors (detailed in section 3.2.3) computed from existing semantic vectors for word senses and lexicalized concepts are indeed effective for both the prediction of strength and the de- …

Content truncated to 500 words