Extracting the Unextractable: A Case Study on Verb-particles

Extracting the Unextractable: A Case Study on Verb-particles Timothy Baldwin∗ and Aline Villavicencio† ∗ CSLI, Ventura Hall, Stanford University Stanford, CA 94305-4115 USA tbaldwin@csli.stanford.edu † University of Cambridge, Computer Laboratory, William Gates Building JJ Thomson Avenue, Cambridge CB3 OFD, UK Aline.Villavicencio@cl.cam.ac.uk Abstract This paper proposes a series of techniques for ex- tracting English verb–particle constructions from raw text corpora. We initially propose three basic methods, based on tagger output, chunker output and a chunk grammar, respectively, with the chunk grammar method optionally combining with an at- tachment resolution module to determine the syn- tactic structure of verb–preposition pairs in ambigu- ous constructs. We then combine the three methods together into a single classifier, and add in a number of extra lexical and frequentistic features, producing a final F-score of 0.865 over the WSJ. 1 Introduction There is growing awareness of the pervasiveness and idiosyncrasy of multiword expressions (MWEs), and the need for a robust, structured handling thereof (Sag et al., 2002; Calzolari et al., 2002; Copestake et al., 2002). Examples of MWEs are lexically fixed expressions (e.g. ad hoc), idioms (e.g. see double), light verb constructions (e.g. make a mistake) and institutionalised phrases (e.g. kindle excitement). MWEs pose a challenge to NLP due to their syn- tactic and semantic idiosyncrasies, which are often unpredictable from their component parts. Large- scale manual annotation of MWEs is infeasible due to their sheer volume (at least equivalent to the num- ber of simplex words (Jackendoff, 1997)), produc- tivity and domain-specificity. Ideally, therefore, we would like to have some means of automatically ex- tracting MWEs from a given domain or corpus, al- lowing us to pre-tune our grammar prior to deploy- ment. It is this task of extraction that we target in this paper. This research represents a component of the LinGO multiword expression project,1 which is targeted at extracting, adequately handling and rep- resenting MWEs of all types. As a research testbed and target resource to expand/domain-tune, we use the LinGO English Resource Grammar (LinGO- ERG), a linguistically-precise HPSG-based gram- mar under development at CSLI (Copestake and Flickinger, 2000; Flickinger, 2000). The particular MWE type we target for extrac- tion is the English verb-particle construction. Verb-particle constructions (“VPCs”) consist of a 1http://lingo.stanford.edu/mwe head verb and one or more obligatory particles, in the form of intransitive prepositions (e.g. hand in), adjectives (e.g. cut short) or verbs (e.g. let go) (Villavicencio and Copestake, 2002a; Villavicen- cio and Copestake, 2002b; Huddleston and Pullum, 2002); for the purposes of this paper, we will fo- cus exclusively on prepositional particles—by far the most common and productive of the three types— and further restrict our attention to single-particle VPCs (i.e. we ignore VPCs such as get along to- gether). We define VPCs to optionally select for an NP complement, i.e. to occur both transitively (e.g. hand in the paper) and intransitively (e.g. battle on). One aspect of VPCs that makes them difficult to extract (cited in, e.g., Smadja (1993)) is that the verb and particle can be non-contiguous, e.g. …

Content truncated to 500 words