Learning to Rank Lexical Substitutions

Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 1926–1932, Seattle, Washington, USA, 18-21 October 2013. c⃝2013 Association for Computational Linguistics Learning to rank lexical substitutions Gy¨orgy Szarvas1 Amazon Inc. szarvasg@amazon.com R´obert Busa-Fekete2 Eyke H¨ullermeier University of Marburg Hans-Meerwein-Str., 35032 Marburg, Germany busarobi@mathematik.uni-marburg.de eyke@mathematik.uni-marburg.de Abstract The problem to replace a word with a syn- onym that fits well in its sentential context is known as the lexical substitution task. In this paper, we tackle this task as a supervised ranking problem. Given a dataset of target words, their sentential contexts and the poten- tial substitutions for the target words, the goal is to train a model that accurately ranks the candidate substitutions based on their contex- tual fitness. As a key contribution, we cus- tomize and evaluate several learning-to-rank models to the lexical substitution task, includ- ing classification-based and regression-based approaches. On two datasets widely used for lexical substitution, our best models signifi- cantly advance the state-of-the-art. 1 Introduction The task to generate lexical substitutions in context (McCarthy and Navigli, 2007), i.e., to replace words in a sentence without changing its meaning, has be- come an increasingly popular research topic. This task is used, e.g. to evaluate semantic models with regard to their accuracy in modeling word meaning in context (Erk and Pad´o, 2010). Moreover, it pro- vides a basis of NLP applications in many fields, including linguistic steganography (Topkara et al., 2006; Chang and Clark, 2010), semantic text simi- larity (Agirre et al., 2012) and plagiarism detection (Gipp et al., 2011). While closely related to WSD, 1Work was done while working at RGAI of the Hungarian Acad. Sci. and University of Szeged. 2 R. Busa-Fekete is on leave from the Research Group on Artificial Intelligence of the Hungarian Academy of Sciences and University of Szeged. lexical substitution does not rely on explicitly de- fined sense inventories (Dagan et al., 2006): the pos- sible substitutions reflect all conceivable senses of the word, and the correct sense has to be ascertained to provide an accurate substitution. While a few lexical sample datasets (McCarthy and Navigli, 2007; Biemann, 2012) with human- provided substitutions exist and can be used to evaluate different lexical paraphrasing approaches, a practically useful system must also be able to rephrase unseen words, i.e., any word for which a list of synonyms is provided. Correspondingly, unsupervised and knowledge-based approaches that are not directly dependent on any training material, prevailed in the SemEval 2007 shared task on En- glish Lexical Substitution and dominated follow-up work. The only supervised approach is limited to the combination of several knowledge-based lexi- cal substitution models based on different underly- ing lexicons (Sinha and Mihalcea, 2009).3 A recent work by Szarvas et al. (2013) de- scribes a tailor-made supervised system based on delexicalized features that – unlike earlier super- vised approaches, and similar to unsupervised and knowledge-based methods proposed for this task – is able to generalize to an open vocabulary. For each target word to paraphrase, they first compute a …

Content truncated to 500 words