Using Large Monolingual and Bilingual Corpora to Improve Coordination Disambiguation

Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics, pages 1346–1355, Portland, Oregon, June 19-24, 2011. c⃝2011 Association for Computational Linguistics Using Large Monolingual and Bilingual Corpora to Improve Coordination Disambiguation Shane Bergsma, David Yarowsky, Kenneth Church Deptartment of Computer Science and Human Language Technology Center of Excellence Johns Hopkins University sbergsma@jhu.edu, yarowsky@cs.jhu.edu, kenneth.church@jhu.edu Abstract Resolving coordination ambiguity is a clas- sic hard problem. This paper looks at co- ordination disambiguation in complex noun phrases (NPs). Parsers trained on the Penn Treebank are reporting impressive numbers these days, but they don’t do very well on this problem (79%). We explore systems trained using three types of corpora: (1) annotated (e.g. the Penn Treebank), (2) bitexts (e.g. Eu- roparl), and (3) unannotated monolingual (e.g. Google N-grams). Size matters: (1) is a mil- lion words, (2) is potentially billions of words and (3) is potentially trillions of words. The unannotated monolingual data is helpful when the ambiguity can be resolved through associ- ations among the lexical items. The bilingual data is helpful when the ambiguity can be re- solved by the order of words in the translation. We train separate classifiers with monolingual and bilingual features and iteratively improve them via co-training. The co-trained classifier achieves close to 96% accuracy on Treebank data and makes 20% fewer errors than a su- pervised system trained with Treebank anno- tations. 1 Introduction Determining which words are being linked by a co- ordinating conjunction is a classic hard problem. Consider the pair: +ellipsis rocket\w1 and mortar\w2 attacks\h −ellipsis asbestos\w1 and polyvinyl\w2 chloride\h +ellipsis is about both rocket attacks and mortar at- tacks, unlike −ellipsis which is not about asbestos chloride. We use h to refer to the head of the phrase, and w1 and w2 to refer to the other two lexical items. Natural Language Processing applications need to recognize NP ellipsis in order to make sense of new sentences. For example, if an Internet search en- gine is given the phrase rocket attacks as a query, it should rank documents containing rocket and mor- tar attacks highly, even though rocket and attacks are not contiguous in the document. Furthermore, NPs with ellipsis often require a distinct type of re- ordering when translated into a foreign language. Since coordination is both complex and produc- tive, parsers and machine translation (MT) systems cannot simply memorize the analysis of coordinate phrases from training text. We propose an approach to recognizing ellipsis that could benefit both MT and other NLP technology that relies on shallow or deep syntactic analysis. While the general case of coordination is quite complicated, we focus on the special case of com- plex NPs. Errors in NP coordination typically ac- count for the majority of parser coordination errors (Hogan, 2007). The information needed to resolve coordinate NP ambiguity cannot be derived from hand-annotated data, and we follow previous work in looking for new information sources to apply to this problem (Resnik, 1999; Nakov and Hearst, 2005; Rus et al., 2007; Pitler …

Content truncated to 500 words