Predicting the Semantic Compositionality of Prefix Verbs

Proceedings of the 2010 Conference on Empirical Methods in Natural Language Processing, pages 293–303, MIT, Massachusetts, USA, 9-11 October 2010. c⃝2010 Association for Computational Linguistics Predicting the Semantic Compositionality of Prefix Verbs Shane Bergsma, Aditya Bhargava, Hua He, Grzegorz Kondrak Department of Computing Science University of Alberta {bergsma,abhargava,hhe,kondrak}@cs.ualberta.ca Abstract In many applications, replacing a complex word form by its stem can reduce sparsity, re- vealing connections in the data that would not otherwise be apparent. In this paper, we focus on prefix verbs: verbs formed by adding a pre- fix to an existing verb stem. A prefix verb is considered compositional if it can be decom- posed into a semantically equivalent expres- sion involving its stem. We develop a clas- sifier to predict compositionality via a range of lexical and distributional features, includ- ing novel features derived from web-scale N- gram data. Results on a new annotated cor- pus show that prefix verb compositionality can be predicted with high accuracy. Our system also performs well when trained and tested on conventional morphological segmentations of prefix verbs. 1 Introduction Many verbs are formed by adding prefixes to exist- ing verbs. For example, remarry is composed of a prefix, re-, and a stem, marry. We present an ap- proach to predicting the compositionality of prefix verbs. The verb remarry is compositional; it means to marry again. On the other hand, retire is gener- ally non-compositional; it rarely means to tire again. There is a continuum of compositionality in prefix verbs, as in other complex word forms and multi- word expressions (Bannard et al., 2003; Creutz and Lagus, 2005; Fazly et al., 2009; Xu et al., 2009). We adopt a definition of compositionality specifi- cally designed to support downstream applications that might benefit from knowledge of verb stems. For example, suppose our corpus contains the fol- lowing sentence: “Pope Clement VII denied Henry VIII permission to marry again before a decision was given in Rome.” A user might submit the ques- tion, “Which pope refused Henry VIII permission to remarry?” If we can determine that the meaning of remarry could also be provided via the stem marry, we could add marry to our search terms. This is known as morphological query expansion (Bilotti et al., 2004). Here, such an expansion leads to a better match between question and answer. Previous work has shown that “full morpholog- ical analysis provides at most very modest bene- fits for retrieval” (Manning et al., 2008). Stem- ming, lemmatization, and compound-splitting often increase recall at the expense of precision, but the results depend on the morphological complexity of the text’s language (Hollink et al., 2004). The lack of success in applying morphological analysis in IR is unsurprising given that most pre- vious systems are not designed with applications in mind. For example, the objective of the influ- ential Linguistica program is “to produce an out- put that matches as closely as possible the analy- sis that would be given by a human morphologist” (Goldsmith, 2001). Unsupervised systems achieve …

Content truncated to 500 words