Single Malt or Blended? A Study in Multilingual Parser Optimization

Proceedings of the CoNLL Shared Task Session of EMNLP-CoNLL 2007, pp. 933–939, Prague, June 2007. c⃝2007 Association for Computational Linguistics Single Malt or Blended? A Study in Multilingual Parser Optimization Johan Hall∗ Jens Nilsson∗ Joakim Nivre∗† G¨uls¸en Eryiˇgit‡ Be´ata Megyesi† Mattias Nilsson† Markus Saers† ∗V¨axj¨o University, School of Mathematics and Systems Engineering E-mail: firstname.lastname@vxu.se †Uppsala University, Dept. of Linguistics and Philology E-mail: firstname.lastname@lingfil.uu.se ‡Istanbul Technical University, Computer Engineering Dept. E-mail: gulsen.cebiroglu@itu.edu.tr Abstract We describe a two-stage optimization of the MaltParser system for the ten languages in the multilingual track of the CoNLL 2007 shared task on dependency parsing. The first stage consists in tuning a single-parser system for each language by optimizing pa- rameters of the parsing algorithm, the fea- ture model, and the learning algorithm. The second stage consists in building an ensem- ble system that combines six different pars- ing strategies, extrapolating from the opti- mal parameters settings for each language. When evaluated on the official test sets, the ensemble system significantly outperforms the single-parser system and achieves the highest average labeled attachment score. 1 Introduction In the multilingual track of the CoNLL 2007 shared task on dependency parsing, a single parser must be trained to handle data from ten different languages: Arabic (Hajiˇc et al., 2004), Basque (Aduriz et al., 2003), Catalan, (Mart´ı et al., 2007), Chinese (Chen et al., 2003), Czech (B¨ohmov´a et al., 2003), English (Marcus et al., 1993; Johansson and Nugues, 2007), Greek (Prokopidis et al., 2005), Hungarian (Csendes et al., 2005), Italian (Montemagni et al., 2003), and Turkish (Oflazer et al., 2003).1 Our contribution is a study in multilingual parser optimization using the freely available MaltParser system, which performs 1For more information about the task and the data sets, see Nivre et al. (2007). deterministic, classifier-based parsing with history- based feature models and discriminative learning, and which was one of the top performing systems in the CoNLL 2006 shared task (Nivre et al., 2006). In order to maximize parsing accuracy, optimiza- tion has been carried out in two stages, leading to two different, but related parsers. The first of these is a single-parser system, similar to the one described in Nivre et al. (2006), which parses a sentence deter- ministically in a single left-to-right pass, with post- processing to recover non-projective dependencies, and where the parameters of the MaltParser system have been tuned for each language separately. We call this system Single Malt, to emphasize the fact that it consists of a single instance of MaltParser. The second parser is an ensemble system, which combines the output of six deterministic parsers, each of which is a variation of the Single Malt parser with parameter settings extrapolated from the first stage of optimization. It seems very natural to call this system Blended. Section 2 summarizes the work done to optimize the Single Malt parser, while section 3 explains how the Blended parser was constructed from the Single Malt parser. Section 4 gives a brief analysis of the experimental results, and section 5 concludes. 2 …

Content truncated to 500 words