UNAL: Discriminating between Literal and Figurative Phrasal Usage Using Distributional Statistics and POS tags

Second Joint Conference on Lexical and Computational Semantics (*SEM), Volume 2: Seventh International Workshop on Semantic Evaluation (SemEval 2013), pages 114–117, Atlanta, Georgia, June 14-15, 2013. c⃝2013 Association for Computational Linguistics UNAL: Discriminating between Literal and Figurative Phrasal Usage Using Distributional Statistics and POS tags Sergio Jimenez, Claudia Becerra Universidad Nacional de Colombia Ciudad Universitaria, edificio 453, oficina 114 Bogotá, Colombia sgjimenezv@unal.edu.co cjbecerrac@unal.edu.co Alexander Gelbukh CIC-IPN Av. Juan Dios Bátiz, Av. Mendizábal, Col. Nueva Industrial Vallejo CP 07738, DF, México gelbukh@gelbukh.com Abstract In this paper we describe the system used to participate in the sub task 5b in the Phrasal Se- mantics challenge (task 5) in SemEval 2013. This sub task consists in discriminating lit- eral and figurative usage of phrases with compositional and non-compositional mean- ings in context. The proposed approach is based on part-of-speech tags, stylistic features and distributional statistics gathered from the same development-training-test text collec- tion. The system obtained a relative improve- ment in accuracy against the most-frequent- class baseline of 49.8% in the “unseen con- texts” (LexSample) setting and 8.5% in “un- seen phrases” (AllWords). 1 Introduction The Phrasal Semantics task-5b in SemEval 2013 consisted in the discrimination of literal of figura- tive usage of phrases in context (Korkontzelos et al., 2013). For instance, the occurrence in a text of the phrase “a piece of cake” can be used whether to re- fer to something that is pretty easy or to an actual piece of cake. The motivation for this task is that such discrimination could improve the quality and performance of other tasks like machine translation and information retrieval. This problem has been studied in the past. Lin (1999) observed that the distributional characteris- tics of the literal and figurative usage are different. Katz and Giesbrecht (2006) showed that the similar- ities among contexts are correlated with their literal or figurative usage. Birke and Sarkar (2006) clus- tered literal and figurative contexts using a word- sense-disambiguation approach. Fazly et al. (2009) showed that literal and figurative usages are related to particular syntactical forms. Sporleder and Li (2009) showed that for a particular phrase the con- texts of its literal usages are more cohesive than those of its figurative usages. Inspired by these works and in a new observation, we proposed a set or features based on cohesiveness, syntax and stylom- etry (Section 2), which are used to train a machine learning classifier. The cohesiveness between a phrase an its context can be measured aggregating the relatedness of the context words against the target phrase. This cohe- siveness should be high for phrases used literally. Conversely, figurative usages can occur in a large variety of contexts implying low cohesiveness. For instance, the cohesiveness of the phrase “a piece of cake” against context words such as “coffee”, “birth- day” and “bakery” should be high. The distribu- tional measures used to obtain the needed related- ness scores and the proposed measures of cohesive- ness are presented in subsection 2.1. Moreover, we observed a stylistic trend in the …

Content truncated to 500 words