Can Recognising Multiword Expressions Improve Shallow Parsing?

Human Language Technologies: The 2010 Annual Conference of the North American Chapter of the ACL, pages 636–644, Los Angeles, California, June 2010. c⃝2010 Association for Computational Linguistics Can Recognising Multiword Expressions Improve Shallow Parsing? Ioannis Korkontzelos, Suresh Manandhar Department of Computer Science The University of York Heslington, York, YO10 5NG, UK {johnkork, suresh}@cs.york.ac.uk Abstract There is significant evidence in the literature that integrating knowledge about multiword expressions can improve shallow parsing ac- curacy. We present an experimental study to quantify this improvement, focusing on com- pound nominals, proper names and adjective- noun constructions. The evaluation set of multiword expressions is derived from Word- Net and the textual data are downloaded from the web. We use a classification method to aid human annotation of output parses. This method allows us to conduct experiments on a large dataset of unannotated data. Experi- ments show that knowledge about multiword expressions leads to an increase of between 7.5% and 9.5% in accuracy of shallow pars- ing in sentences containing these multiword expressions. 1 Introduction Multiword expressions are sequences of words that tend to co-occur more frequently than chance and are characterised by various levels of idiosyncracy (Baldwin et al., 2003; Baldwin, 2006). There is ex- tended literature on various issues relevant to mul- tiword expression; recognition, classification, lexi- cography, etc. (see Section 6). The vast majority of these publications identifies as motivation for mul- tiword expression research its potential contribution to deep or shallow parsing. On the other side of this issue, the state-of-the-art parsing systems seem to ignore the fact that treating multiword expressions as syntactic units would potentially increase parser’s accuracy. In this paper, we present an experimental study attempting to estimate the contribution of integrat- ing multiword expressions into shallow parsing. We focus on multiword expressions that consist of two successive tokens; in particular, compound nominals proper names and adjective-noun constructions. We also present a detailed classification method to aid human annotation during the procedure of deciding if a parse is correct or wrong. We present experi- mental results about the different classes of changes that occur in the parser output while unifying multi- word expression components. We conclude that treating known multiwords ex- pressions as singletons leads to an increase of be- tween 7.5% and 9.5% in accuracy of shallow pars- ing of sentences containing these multiword expres- sions. Increase percentages are higher for multiword expressions that consist of an adjective followed by a noun (12% to 15%); and even higher for non- compositional multiword expressions1 that consist of an adjective and a noun (15.5% to 19.5%). The rest of the paper is structured as follows: In Section 2 we present how multiword expressions can be annotated in text and used by a shallow parser. In Section 3 we present an overview of our experimen- tal process. Section 4 explains how the set of target multiword expressions and textual corpora were cre- ated. In Section 5 we present and discuss the results of the experimental process. In Section 6 we …

Content truncated to 500 words