Distributional Similarity of Multi-Word Expressions

Distributional Similarity of Multi-Word Expressions Laura Ingram and James R. Curran School of Information Technologies The University of Sydney NSW 2006, Australia {ling6188, james}@it.usyd.edu.au Abstract Most existing systems for automatically extracting lexical-semantic resources ne- glect multi-word expressions (MWEs), even though approximately 30% of gold-standard thesauri entries are MWEs. We present a distributional similarity system that identifies synonyms for MWEs. We ex- tend Grefenstette’s SEXTANT shallow parser to first identify bigram MWEs using colloca- tion statistics from the Google WEB1T cor- pus. We extract contexts from WEB1T to increase coverage on the sparser bigrams. 1 Introduction Lexical-semantic resources, such as WordNet (Fell- baum, 1998), are used in many applications in Natu- ral Language Processing (NLP). Unfortunately, they are expensive and time-consuming to produce and are prone to bias and limited coverage. Automat- ically extracting these resources is crucial to over- coming the knowledge bottleneck in NLP. Existing distributional approaches to semantic similarity focus on unigrams, with very little work on extracting synonyms for multi-word expressions (MWEs). In this work, we extend an existing system to support MWEs by identifying bigram MWEs using collocation statistics (Manning and Sch¨utze, 1999). These are calculated using n-gram counts from the Google WEB1T corpus (Brants and Franz, 2006). We evaluate against several gold-standard the- sauri and observe a slight decrease in overall per- formance when the bigram MWEs were included. This is unsurprising since the larger vocabulary and sparser contextual information for bigrams makes the task significantly harder. We also experimented with contexts extracted from WEB1T in an attempt to overcome the data sparseness problem. Inspec- tion of the results for individual headwords revealed many cases where the synonyms returned were sig- nificantly better when bigram data was included. 2 Background Distributional similarity relies on the distributional hypothesis that similar terms appear in similar con- texts (Harris, 1954). Here we extend the SEXTANT parser (Grefenstette, 1994) to include multi-word terms and syntactic contexts. Curran (2004) experiments with different parsers for extracting contextual information, including SEXTANT, MINIPAR (Lin, 1994), RASP (Briscoe and Carroll, 2002), and CASS (Abney, 1996). Lin (1998) used MINIPAR and Weeds (2003) used RASP for dis- tributional similarity calculations. MINIPAR is the only parser to identify a range of MWEs that has been used for distributional similarity. Weeds (2003) and Curran (2004) evaluate measures for calculating dis- tributional similarity. We follow (Curran, 2004) in using the weighted Jaccard measure with truncated t-test relation weighting for our experiments. 3 Detecting MWEs The initial step in creating a thesaurus for MWEs is to identify potential MWE headwords using colloca- tion statistics. We used various statistical tests, e.g. the t-test and the log-likelihood test (Manning and Sch¨utze, 1999), calculated over the Google WEB1T unigram and bigram counts. These counts, calcu- lated over 1 trillion words of web text, gave the most reliable counts. However, highly ranked terms, e.g. Proceedings of the Australasian Language Technology Workshop 2007, pages 146-148 146 Contact Us and Site Map, demonstrate bias towards web-related terminology. This list of selected bi- grams is used …

Content truncated to 500 words