Paving the Way to a Large-scale Pseudosense-annotated Dataset

Proceedings of NAACL-HLT 2013, pages 1100–1109, Atlanta, Georgia, 9–14 June 2013. c⃝2013 Association for Computational Linguistics Paving the Way to a Large-scale Pseudosense-annotated Dataset Mohammad Taher Pilehvar and Roberto Navigli Department of Computer Science Sapienza University of Rome {pilehvar,navigli}@di.uniroma1.it Abstract In this paper we propose a new approach to the generation of pseudowords, i.e., artificial words which model real polysemous words. Our approach simultaneously addresses the two important issues that hamper the gener- ation of large pseudosense-annotated datasets: semantic awareness and coverage. We eval- uate these pseudowords from three different perspectives showing that they can be used as reliable substitutes for their real counterparts. 1 Introduction A fundamental problem in computational linguis- tics is the paucity of manually annotated data, such as part-of-speech tagged sentences, treebanks, and logical forms, which exist only for few languages (Ide et al., 2010). A case in point is the lack of abundant sense annotated data, which hampers the performance and coverage of lexical semantic tasks such as Word Sense Disambiguation (Navigli, 2009; Navigli, 2012, WSD) and semantic role labeling (Gildea and Jurafsky, 2002). A possible way to break this bottleneck is to use pseudowords, i.e., arti- ficial words constructed by conflating a set of unam- biguous words, with the aim of modeling polysemy in real ambiguous words. The idea of pseudowords was originally proposed by Gale et al. (1992) and Sch¨utze (1992) for WSD evaluation, but later found application in other tasks such as selectional prefer- ences (Erk, 2007; Bergsma et al., 2008; Chambers and Jurafsky, 2010), Word Sense Induction (Bor- dag, 2006; Di Marco and Navigli, 2013) or studies concerning the effects of the amount of data on ma- chine learning for natural language disambiguation (Banko and Brill, 2001). Being made up of monose- mous words, pseudowords can potentially be used to create large amounts of pseudosense-annotated data at virtually no cost, hence enabling large-scale stud- ies in lexical semantics. Unfortunately, though, the extent of their usability for such a purpose is ham- pered by two main issues: semantic awareness and wide coverage. Semantic awareness corresponds to the constraint that pseudowords, in order to be realistic, are ex- pected to have senses which are in a semantic rela- tionship (thus modeling systematic polysemy). Re- cent work has focused on this issue and, by exploit- ing either specific lexical hierarchies (Nakov and Hearst, 2003; Lu et al., 2006), or the WordNet struc- ture (Otrusina and Smrz, 2010), have succeeded in generating pseudowords which are comparable to real words in terms of disambiguation difficulty. The second challenge is coverage, which corresponds to the number of distinct pseudowords an algorithm can generate. When coupled with the semantic awareness issue, wide coverage is hampered by the difficulty in generating thousands of pseudowords which mimic existing polysemous words. Unfortunately, none of the existing approaches to the generation of pseudowords can meet both these challenges simultaneously, and this has hindered the generation of a large pseudosense-annotated dataset. For instance, approaches which exploit the monosemous neighbors of a target sense …

Content truncated to 500 words