Multi-Word Expression Identification Using Sentence Surface Features

Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 468–477, Singapore, 6-7 August 2009. c⃝2009 ACL and AFNLP Multi-Word Expression Identification Using Sentence Surface Features Ram Boukobza School of Computer Science Hebrew University of Jerusalem ram.boukobza@mail.huji.ac.il Ari Rappoport School of Computer Science Hebrew University of Jerusalem arir@cs.huji.ac.il Abstract Much NLP research on Multi-Word Ex- pressions (MWEs) focuses on the discov- ery of new expressions, as opposed to the identification in texts of known expres- sions. However, MWE identification is not trivial because many expressions al- low variation in form and differ in the range of variations they allow. We show that simple rule-based baselines do not perform identification satisfactorily, and present a supervised learning method for identification that uses sentence surface features based on expressions’ canonical form. To evaluate the method, we have annotated 3350 sentences from the British National Corpus, containing potential uses of 24 verbal MWEs. The method achieves an F-score of 94.86%, compared with 80.70% for the leading rule-based base- line. Our method is easily applicable to any expression type. Experiments in pre- vious research have been limited to the compositional/non-compositional distinc- tion, while we also test on sentences in which the words comprising the MWE ap- pear but not as an expression. 1 Introduction Multi-Word Expressions (MWEs) such as ‘pull strings’, ‘make a face’ and ‘get on one’s nerves’ are very common in language. Such MWEs can be characterized as being non-compositional: the meaning of the expression does not transparently follow from the meaning of the words that com- prise it. Much of the work on MWEs in NLP has been in MWE extraction – the discovery of new MWEs from a corpus, using statistical and other methods. Identification of known MWEs in text has received less attention, but is necessary for many NLP applications, for example in machine translation. The current work deals with the MWE identification task: deciding if a sentence contains a use of a known expression. MWE identification is not as simple as may ini- tially appear, as will be shown by the performance of two rule-based baselines in our experiments. One source of difficulty is variations in expres- sions’ usage in text. Although MWEs generally show less variation than single words, they show enough that it cannot be ignored. In a study on V+NP idioms, Riehemann (2001) found that the idioms’ canonical form accounted for 75% of their appearances in a corpus. Additionally, expressions differ considerably in the types of variations they allow, which include passivization, nominalization and addition of modifying words (Moon, 1998). A second source of difficulty is that expressions consisting of very frequent words will often co- occur in sentences in a non-MWE usage and in similar but distinct expressions. MWE identification can be modeled as a two step process. Given a sentence and a known ex- pression, step (1) is to decide if the sentence con- tains a potential use of the expression. This is a relatively simple step based on the appearance …

Content truncated to 500 words