Punctuation Normalisation for Cleaner Treebanks and Parsers

Punctuation normalisation for cleaner treebanks and parsers Daniel Tse and James R. Curran School of Information Technologies University of Sydney NSW 2006, Australia {dtse6695,james}@it.usyd.edu.au Abstract Although punctuation is pervasive in written text, their treatment in parsers and corpora is often second-class. We examine the treatment of commas in CCGbank, a wide-coverage corpus for Com- binatory Categorial Grammar (CCG), re- analysing its comma structures in order to eliminate a class of redundant rules, obtaining a more consistent treebank. We then eliminate these rules from C&C, a wide-coverage statistical CCG parser, obtain- ing a 37% increase in parsing speed on the standard CCGbank test set and a considerable reduction in memory consumed, without af- fecting parser accuracy. 1 Introduction Although parsers must all at least accept text con- taining punctuation, there is much variation in how punctuation is treated during training, parsing and evaluation. The work of Nunberg (1990) brought attention to the fact that parsers can often reject oth- erwise ambiguous parses on the basis of punctua- tion, by considering a formal grammar of text units: chunks of text delimited by punctuation tokens. In this work, we explore how the treatment of commas, the most common class of punctuation in the Penn Treebank (Marcus et al., 1994), affects their analysis in CCGbank (Hockenmaier and Steed- man, 2007), a 1.2 million word corpus for Combina- tory Categorial Grammar (CCG), generated from a subset of the Penn Treebank. We uncover a system- atic inconsistency in the way CCGbank represents sentences involving commas, and perform a trans- formation of the original corpus which eliminates this source of uninformative variation, resulting in a version of CCGbank which assigns a uniform struc- ture to each of the syntactic roles played by commas in the corpus. By removing the superfluous rules induced by this inconsistency from the corpus, we obtain cleaner analyses of comma structures which require fewer CCG rules to explain. These changes to the cor- pus, in turn, allow us to simplify the implementa- tion of the C&C parser (Clark and Curran, 2007), a wide-coverage statistical CCG parser for English, by barring the parser from considering the CCG rules which we have made redundant. Training the C&C parser on our modified CCGbank yields average speed gains of 37% and 21% on the standard CCGbank test set and develop- ment set respectively, and a 47% reduction in mem- ory consumed by the chart parsing process as evalu- ated on the standard CCGbank test set. Parser cover- age increases slightly while accuracy is not affected. Consistency, a powerful and general guideline in corpus design, can improve the usefulness and qual- ity of corpora and the applications built with them. We eliminate a source of systematic inconsistency in comma representation from a wide-coverage cor- pus and parser, resulting in cleaner versions of both resources. The combination of a cleaner treebank and parser leads to gains in speed and a reduction in memory consumption without affecting parser accu- racy. 2 Background Nunberg (1990) characterises the traditional view of written …

Content truncated to 500 words