Combining PCFG-LA Models with Dual Decomposition: A Case Study with Function Labels and Binarization

Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 1158–1169, Seattle, Washington, USA, 18-21 October 2013. c⃝2013 Association for Computational Linguistics Combining PCFG-LA Models with Dual Decomposition: A Case Study with Function Labels and Binarization Joseph Le Roux†, Antoine Rozenknop†, Jennifer Foster∗ † Universit´e Paris 13, Sorbonne Paris Cit´e, LIPN, F-93430, Villetaneuse, France ∗ NCLT/CNGL, School of Computing, Dublin City University, Dublin 9, Ireland joseph.leroux@lipn.fr antoine.rozenknop@lipn.fr jfoster@computing.dcu.ie Abstract It has recently been shown that different NLP models can be effectively combined using dual decomposition. In this paper we demon- strate that PCFG-LA parsing models are suit- able for combination in this way. We exper- iment with the different models which result from alternative methods of extracting a gram- mar from a treebank (retaining or discarding function labels, left binarization versus right binarization) and achieve a labeled Parseval F-score of 92.4 on Wall Street Journal Sec- tion 23 – this represents an absolute improve- ment of 0.7 and an error reduction rate of 7% over a strong PCFG-LA product-model base- line. Although we experiment only with bina- rization and function labels in this study, there is much scope for applying this approach to other grammar extraction strategies. 1 Introduction Because of the large amount of possibly contra- dictory information contained in a treebank, learn- ing a phrase-structure-based parser implies making several choices regarding the prevalent annotations which have to be kept – or discarded – in order to guide the learning algorithm. These choices, which include whether to keep function labels and empty nodes, how to binarize the trees and whether to alter the granularity of the tagset, are often motivated em- pirically by parsing performance rather than by the different aspects of the language they may be able to capture. Recently Rush et al. (2010), Martins et al. (2011) and Koo et al. (2010) have shown that Dual De- composition or Lagrangian Relaxation is an elegant S f e d c b (a) Original example S ⟨S⟩ ⟨S⟩ ⟨S⟩ f e d c b (b) Left Binarized example S f ⟨S⟩ e ⟨S⟩ d ⟨S⟩ c b (c) Right Binarized example Figure 1: Binarization with markovization framework for combining different types of NLP tasks or for building parsers from simple slave pro- cesses that only check partial well-formedness. Here we propose to follow this idea, but with a different objective. We want to mix different parsers trained on different versions of a treebank each of which makes some annotation choices in order to learn more specific or richer information. We will use state-of-the-art unlexicalized probabilistic context- free grammars with latent annotations (PCFG-LA) in order to compare our approach with a strong base- line of high-quality parses. Dual Decomposition is used to mix several systems (between two and four) that may in turn be combinations of grammars, here products of PCFG-LAs (Petrov, 2010). The systems being combined make different choices with regard to i) function labels and ii) grammar binarization. Common sense would suggest that information in …

Content truncated to 500 words