Classification of Verb Particle Constructions with the Google Web1T Corpus

Classification of Verb-Particle Constructions with the Google Web1T Corpus Jonathan K. Kummerfeld and James R. Curran School of Information Technologies University of Sydney NSW 2006, Australia {jkum0593,james}@it.usyd.edu.au Abstract Manually maintaining comprehensive databases of multi-word expressions, for example Verb-Particle Constructions (VPCs), is infeasible. We describe a new type level classifier for potential VPCs, which uses information in the Google Web1T corpus to perform a simple lin- guistic constituency test. Specifically, we consider the fronting test, comparing the frequencies of the two possible orderings of the given verb and particle. Using only a small set of queries for each verb-particle pair, the system was able to achieve an F-score of 75.7% in our evaluation while processing thousands of queries a second. 1 Introduction Creating comprehensive linguistic resources manually is an expensive and slow process, making it infeasible for maintaining up-to-date resources of constantly evolving linguistic fea- tures, such as Multi-Word Expressions (MWEs). These resources are crucial for a range of Natural Language Processing (NLP) tasks, such as ac- curate parsing. Identification of MWEs does not prevent the production of syntactically accurate parses, but the extra information can improve results. Since manually creating these resources is a challenge, systems are needed that can automatically classify expressions accurately. Also, because these systems will be running during parsing, they need to be fast to prevent the creation of an additional bottleneck. Here we focus on Verb-Particle Constructions (VPCs), a type of MWE composed of a verb and a particle. Villavicencio (2003a) showed that VPCs are poorly covered by corpus data and constantly growing in number, making them a suitable can- didate for an automatic classification system. Pre- vious work on VPCs has mainly focused either on their compositionality (McCarthy et al., 2003), or on using sophisticated parsers to perform ex- traction from corpora (Baldwin and Villavicencio, 2002). While parser based methods have been very successful (Kim and Baldwin, 2006) they rely on contextual knowledge and complex pro- cessing. The Web has been used as a corpus previ- ously by Villavicencio (2003b), who used search engines as a source of statistics for classification, but her aim was to create new resources, rather than a tool that can be used for identification. We have constructed a high-throughput query system for the Google Web1T data, which we use to col- lect information to perform the fronting linguistic constituency test. The results of this process are used to train a classifier, which can then quickly and accurately classify a given verb-particle pair. The Google Web1T corpus is over twenty-four gigabytes of plain text in compressed form, mak- ing it infeasible to store the entire data-set in memory and perform queries directly. To han- dle the data we have aggregated the frequency counts for n-grams by using templates that con- tain a mixture of wild-cards and words, where the words are all possibilities taken from three sets. The new data this process produces is stored in a hash-table distributed across several computers, making fast queries possible. These queries pro- vide …

Content truncated to 500 words