Perceptron Reranking for CCG Realization

Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 410–419, Singapore, 6-7 August 2009. c⃝2009 ACL and AFNLP Perceptron Reranking for CCG Realization Michael White and Rajakrishnan Rajkumar Department of Linguistics The Ohio State University Columbus, OH, USA {mwhite,raja}@ling.osu.edu Abstract This paper shows that discriminative reranking with an averaged perceptron model yields substantial improvements in realization quality with CCG. The paper confirms the utility of including language model log probabilities as features in the model, which prior work on discrimina- tive training with log linear models for HPSG realization had called into question. The perceptron model allows the combina- tion of multiple n-gram models to be opti- mized and then augmented with both syn- tactic features and discriminative n-gram features. The full model yields a state- of-the-art BLEU score of 0.8506 on Sec- tion 23 of the CCGbank, to our knowledge the best score reported to date using a re- versible, corpus-engineered grammar. 1 Introduction In this paper, we show how discriminative train- ing with averaged perceptron models (Collins, 2002) can be used to substantially improve surface realization with Combinatory Categorial Gram- mar (Steedman, 2000, CCG). Velldal and Oepen (2005) and Nakanishi et al. (2005) have shown that discriminative training with log-linear (maximum entropy) models is effective in realization rank- ing with Head-Driven Phrase Structure Grammar (Pollard and Sag, 1994, HPSG). Here we show that averaged perceptron models also perform well for realization ranking with CCG. Averaged per- ceptron models are very simple, just requiring a decoder and a simple update function, yet despite their simplicity they have been shown to achieve state-of-the-art results in Treebank and CCG pars- ing (Huang, 2008; Clark and Curran, 2007a) as well as on other NLP tasks. Along the way, we address the question of whether it is beneficial to incorporate n-gram log probabilities as baseline features in a discrimina- tively trained realization ranking model. On a lim- ited domain corpus, Velldal & Oepen found that including the n-gram log probability of each can- didate realization as a feature in their log-linear model yielded a substantial boost in ranking per- formance; on the Penn Treebank (PTB), however, Nakanishi et al. found that including an n-gram log prob feature in their model was of no benefit (with the use of bigrams instead of 4-grams suggested as a possible explanation). With these mixed results, the utility of n-gram baseline features for PTB- scale discriminative realization ranking has been unclear. In our particular setting, the question is: Do n-gram log prob features improve performance in broad coverage realization ranking with CCG, where factored language models over words, part- of-speech tags and supertags have previously been employed (White et al., 2007; Espinosa et al., 2008)? We answer this question in the affirmative, con- firming the results of Velldal & Oepen, despite the differences in corpus size and kind of lan- guage model. We show that including n-gram log prob features in the perceptron model is highly beneficial, as the discriminative models we tested without …

Content truncated to 500 words