Using LTAG Based Features in Parse Reranking

Using LTAG Based Features in Parse Reranking∗ Libin Shen Dept. of Computer & Info. Sci. University of Pennsylvania libin@cis.upenn.edu Anoop Sarkar School of Computing Science Simon Fraser University anoop@cs.sfu.ca Aravind K. Joshi Dept. of Computer & Info. Sci. University of Pennsylvania joshi@cis.upenn.edu Abstract We propose the use of Lexicalized Tree Adjoining Grammar (LTAG) as a source of features that are useful for reranking the output of a statistical parser. In this paper, we extend the notion of a tree ker- nel over arbitrary sub-trees of the parse to the derivation trees and derived trees pro- vided by the LTAG formalism, and in ad- dition, we extend the original definition of the tree kernel, making it more lexi- calized and more compact. We use LTAG based features for the parse reranking task and obtain labeled recall and precision of 89.7%/90.0% on WSJ section 23 of Penn Treebank for sentences of length ≤ 100 words. Our results show that the use of LTAG based tree kernel gives rise to a 17% relative difference in f-score im- provement over the use of a linear kernel without LTAG based features. 1 Introduction Recent work in statistical parsing has explored al- ternatives to the use of (smoothed) maximum likeli- hood estimation for parameters of the model. These alternatives are distribution-free (Collins, 2001), providing a discriminative method for resolving parse ambiguity. Discriminative methods provide a ranking between multiple choices for the most plau- sible parse tree for a sentence, without assuming that a particular distribution or stochastic process gener- ated the alternative parses. ∗We would like to thank Michael Collins for providing the original n-best parsed data on which we ran our experiments and the anonymous reviewers for their comments. The sec- ond author is partially supported by NSERC, Canada (RGPIN: 264905). Discriminative methods permit the use of feature functions that can be used to condition on arbitrary aspects of the input. This flexibility makes it possi- ble to incorporate features of various of kinds. Fea- tures can be defined on characters, words, part of speech (POS) tags and context-free grammar (CFG) rules, depending on the application to which the model is applied. Features defined on n-grams from the input are the most commonly used for NLP applications. Such n-grams can either be defined explicitly us- ing some linguistic insight into the problem, or the model can be used to search the entire space of n- gram features using a kernel representation. One example is the use of a polynomial kernel over se- quences. However, to use all possible n-gram fea- tures typically introduces too many noisy features, which can result in lower accuracy. One way to solve this problem is to use a kernel function that is tailored for particular NLP applications, such as the tree kernel (Collins and Duffy, 2001) for statistical parsing. In addition to n-gram features, more complex high-level features are often exploited to obtain higher accuracy, especially when discriminative models are used for statistical parsing. For ex- ample, all possible sub-trees …

Content truncated to 500 words