Re-estimation of Lexical Parameters for Treebank PCFGs

Proceedings of the 22nd International Conference on Computational Linguistics (Coling 2008), pages 193–200 Manchester, August 2008 Re-estimation of Lexical Parameters for Treebank PCFGs Tejaswini Deoskar Department of Linguistics Cornell University td72@cornell.edu Abstract We present procedures which pool lexical information estimated from unlabeled data via the Inside-Outside algorithm, with lex- ical information from a treebank PCFG. The procedures produce substantial im- provements (up to 31.6% error reduction) on the task of determining subcategoriza- tion frames of novel verbs, relative to a smoothed Penn Treebank-trained PCFG. Even with relatively small quantities of unlabeled training data, the re-estimated models show promising improvements in labeled bracketing f-scores on Wall Street Journal parsing, and substantial benefit in acquiring the subcategorization prefer- ences of low-frequency verbs. 1 Introduction In order to obtain the meaning of a sentence au- tomatically, it is necessary to have access to its syntactic analysis at some level of complexity. Many NLP applications like translation, question- answering, etc. might benefit from the avail- ability of syntactic parses. Probabilistic parsers trained over labeled data have high accuracy on in- domain data: lexicalized parsers get an f-score of up to 90.0% on Wall Street Journal data (Charniak and Johnson (2005)’s re-ranking parser), while re- cently, unlexicalized PCFGs have also been shown to perform much better than previously believed (Klein and Manning, 2003). However, the limited size of annotated training data results in many pa- rameters of a PCFG being badly estimated when c⃝ 2008. Licensed under the Creative Commons Attribution-Noncommercial-Share Alike 3.0 Unported li- cense (http://creativecommons.org/licenses/by-nc-sa/3.0/). Some rights reserved. trained on annotated data. The Zipfian nature of a text corpus results in PCFG parameters related to the properties of specific words being espe- cially badly estimated. For instance, about 38% of verbs in the training sections of the Penn Treebank (PTB) (Marcus et al., 1993) occur only once – the lexical properties of these verbs (such as their most common subcategorization frames ) cannot be rep- resented accurately in a model trained exclusively on the Penn Treebank. The research reported here addresses this issue. We start with an unlexicalized PCFG trained on the PTB. We then re-estimate the parameters of this PCFG from raw text using an unsupervised estimation method based on the Inside-Outside al- gorithm (Lari and Young, 1990), an instance of the Expectation Maximization algorithm (Demp- ster et al., 1977) for PCFG induction. The re- estimation improves f-score on the standard test section of the PTB significantly. Our focus is on learning lexical parameters i.e. those parameters related to the lexico-syntactic properties of open- class words. Examples of such properties are: sub- categorization frames of verbs and nouns, attach- ment preference of adverbs to sentential, verbal or nominal nodes, attachment preference of PPs to a verbal or nominal node, etc. The current research is related to semi- supervised training paradigms like self-training – these methods are currently being explored to im- prove the performance of existing PCFG models by utilizing unlabeled data. For example, Mc- Closkey et al. (2006) achieve a 1.1% …

Content truncated to 500 words