An empirical study for Vietnamese dependency parsing

An empirical study for Vietnamese dependency parsing Dat Quoc Nguyen, Mark Dras and Mark Johnson Department of Computing Macquarie University, Australia dat.nguyen@students.mq.edu.au, {mark.dras, mark.johnson}@mq.edu.au Abstract This paper presents an empirical com- parison of different dependency parsers for Vietnamese, which has some unusual characteristics such as copula drop and verb serialization. Experimental results show that the neural network-based parsers perform significantly better than the tra- ditional parsers. We report the highest parsing scores published to date for Viet- namese with the labeled attachment score (LAS) at 73.53% and the unlabeled attach- ment score (UAS) at 80.66%. 1 Introduction Dependency parsing has become a key research topic in natural language processing in the last decade, boosted by the success of the CoNLL 2006 and 2007 shared tasks on multilingual depen- dency parsing (Buchholz and Marsi, 2006; Nivre et al., 2007a). McDonald and Nivre (2011) iden- tify two types of approaches for dependency pars- ing: graph-based approaches (McDonald et al., 2005) and transition-based approaches (Nivre et al., 2007b). Most traditional graph- or transition- based dependency parsers (McDonald et al., 2005; Nivre et al., 2007b; Bohnet, 2010; Zhang and Nivre, 2011; Martins et al., 2013; Choi and Mc- Callum, 2013) manually define a set of core and combined features associated with one-hot repre- sentations. Recent work shows that neural network-based parsers obtain the state-of-the-art parsing results across many languages. Chen and Manning (2014), Weiss et al. (2015), Pei et al. (2015), and Andor et al. (2016) represent the core fea- tures with dense vector embeddings and then feed them as inputs to neural network-based classi- fiers, while Dyer et al. (2015), Kiperwasser and Goldberg (2016a), and Kiperwasser and Goldberg (2016b) propose novel neural network architec- tures to solve the feature-engineering problem. Dependency parsing for Vietnamese has not been actively explored. One main reason is be- cause there is no manually labeled dependency treebank available. Thi et al. (2013) and Nguyen et al. (2014b) propose constituent-to-dependency conversion approaches to automatically translate the manually built constituent treebank for Viet- namese (Nguyen et al., 2009) to dependency tree- banks. The converted dependency treebanks are then used in later works on Vietnamese depen- dency parsing, including Vu-Manh et al. (2015), Le-Hong et al. (2015) and Nguyen and Nguyen (2015). All of the previous research works use ei- ther the MSTparser (McDonald et al., 2005) or the Maltparser (Nivre et al., 2007b) for their parsing experiments. Among them, Nguyen et al. (2014b) report the highest results with LAS at 71.66% and UAS at 79.08% obtained by MSTparser. However, MSTparser and Maltparser are no longer consid- ered state-of-the-art parsers. In this paper, we present an empirical study of Vietnamese dependency parsing. We make com- parisons between neural network-based parsers and traditional parsers, and also between graph- based parsers and transition-based parsers. We show that the neural network-based parsers ob- tain significantly higher scores than the traditional parsers. Specifically, we report the highest up-to- date scores for Vietnamese with LAS at 73.53% and UAS at 80.66%. We also examine poten- tial …

Content truncated to 500 words