WHUNlp at SemEval-2016 Task DiMSUM: A Pilot Study in Detecting Minimal Semantic Units and their Meanings using Supervised Models

Proceedings of SemEval-2016, pages 918–924, San Diego, California, June 16-17, 2016. c⃝2016 Association for Computational Linguistics WHUNlp at SemEval-2016 Task DiMSUM: A Pilot Study in Detecting Minimal Semantic Units and their Meanings using Supervised Models Xin Tang and Fei Li and Donghong Ji Computer School, Wuhan University, Wuhan, China 1648706227@qq.com, lifei csnlp and dhji@whu.edu.cn Abstract This paper describes our approach towards the SemEval-2016 Task 10: Detecting Mini- mal Semantic Units and their Meanings (DiM- SUM). We consider that the two problems are similar to multiword expression detection and supersense tagging, respectively. The former problem is formalized as a sequence label- ing problem solved by first-order CRFs, and the latter one is formalized as a classification problem solved by Maximum Entropy Algo- rithm. To carry out our pilot study quickly, we extract some simple features such as words or part-of-speech tags from the training set, and avoid using external resources such as Word- Net or Brown clusters which are allowed in the supervised closed condition. Experimental re- sults show that much further work on feature engineering and model optimization needs to be explored. 1 Introduction In the community of natural language processing, multiword expressions (MWEs) detection (Schnei- der et al., 2014b; Schneider et al., 2014a) and super- sense tagging (Ciaramita and Johnson, 2003; Cia- ramita and Altun, 2006) have received much re- search attention due to their various application- s such as syntactic parsing (Candito and Constan- t, 2014; Bengoetxea et al., 2014), semantic pars- ing (Banarescu et al., 2013), and machine transla- tion (Carpuat and Diab, 2010). However, not much attention has been paid to the relationship between MWEs and supersenses (Piao et al., 2005; Schnei- der and Smith, 2015). Input: SecurityNOUN increasedV ERB inADP MumbaiP ROP N amidADP terrorNOUN threatsNOUN aheadADP ofADP GaneshotsavP ROP N Output: Securityn.state increasedv.change in Mumbain.location amid terror threatsn.communication ahead of Ganeshotsavn.event Figure 1: An DiMSUM Example. Given a tok- enized and POS-tagged sentence, outputs will be a representation annotated with MWEs and supersens- es. Noun and verb supersenses start with “n.” and “v.”, respectively. “ ” joins tokens within a MWE. The DiMSUM shared task (Schneider et al., 2016) at SemEval 2016 aims to predict a broad-coverage representation of lexical semantics giving an English sentence. This representation consists of two facets: a segmentation into minimal semantic units, and a labeling of some of those units with semantic class- es. Based on the task descriptions, we consider the concepts of minimal semantic units and semantic classes are identical to those of MWEs and super- senses, respectively. Figure 1 shows an input exam- ple and its corresponding outputs of DiMSUM task. Prior work on MWE detection using unsuper- vised methods includes lexicon lookup (Bejˇcek et al., 2013), statistical association measures (Ramisch et al., 2012), parallel corpora (Tsvetkov and Wint- ner, 2010), or hybrid methods (Tsvetkov and Wint- ner, 2011). More sophisticated methods use super- vised techniques such as conditional random field- s (CRFs) (Shigeto et al., 2013; Constant et al., 2012; Vincze et al., 2013) or …

Content truncated to 500 words