Semi-supervised CCG Lexicon Extension

Proceedings of the 2011 Conference on Empirical Methods in Natural Language Processing, pages 1246–1256, Edinburgh, Scotland, UK, July 27–31, 2011. c⃝2011 Association for Computational Linguistics Semi-supervised CCG Lexicon Extension Emily Thomforde University of Edinburgh e.j.thomforde@sms.ed.ac.uk Mark Steedman University of Edinburgh steedman@inf.ed.ac.uk Abstract This paper introduces Chart Inference (CI), an algorithm for deriving a CCG category for an unknown word from a partial parse chart. It is shown to be faster and more pre- cise than a baseline brute-force method, and to achieve wider coverage than a rule-based system. In addition, we show the application of CI to a domain adaptation task for ques- tion words, which are largely missing in the Penn Treebank. When used in combination with self-training, CI increases the precision of the baseline StatCCG parser over subject- extraction questions by 50%. An error analy- sis shows that CI contributes to the increase by expanding the number of category types avail- able to the parser, while self-training adjusts the counts. 1 Introduction Unseen lexical items are a major cause of error in strongly lexicalised parsers such as those based on CCG (Clark and Curran, 2003; Hockenmaier, 2003). The problem is especially acute for less privileged languages, but even in the case of English, we are aware of many category types entirely missing from the Penn Treebank (Clark et al., 2004). In the case of totally unseen words, the standard method used by StatCCG (Hockenmaier, 2003) and many other treebank parsers is part-of-speech back- off, which is quite effective, affording an F-score of 93% over dependencies in §00 in the optimal config- uration. It is difficult to say how backing off affects dependency errors, but when we examine category match accuracy of the CCGBank-trained parser, we find that POS backoff has been used on 19.6% of to- kens, which means that those tokens are unseen, or too infrequent in the training data to be included in the lexicon. Of the 3320 items the parser labelled incorrectly, 675 (20.3%) are words that are miss- ing from the lexicon entirely.1 In the best case, if we were able to learn lexical entries for those 675, we could transfer them to lexical treatment, which is 93.5% accurate, rather than POS backoff, which is 89.3% accurate. Under these conditions, we pre- dict a further 631 word/category pairs to be tagged correctly by the parser, reducing the error rate from 7.4% to 6% on §00. Further to reducing parsing er- ror, a robust method for learning words from un- labelled data would result in the recovery of inter- esting and important category types that are missing from our standard lexical resources. This paper introduces Chart Inference (CI) as a strategy for deducing a ranked set of possible categories for an unknown word using the partial chart formed from the known words that surround it. CCG (Steedman, 2000) is particularly suited to this problem, because category types can be inferred from the types of the surrounding constituents. CI is designed to take advantage of this …

Content truncated to 500 words