Jointly or Separately: Which is Better for Parsing Heterogeneous Dependencies?

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 530–540, Dublin, Ireland, August 23-29 2014. Jointly or Separately: Which is Better for Parsing Heterogeneous Dependencies? Meishan Zhang†, Wanxiang Che†, Yanqiu Shao‡ , Ting Liu†∗ †Research Center for Social Computing and Information Retrieval Harbin Institute of Technology, China {mszhang, car, tliu}@ir.hit.edu.cn ‡Beijing Language and Culture University yqshao163@163.com Abstract For languages such as English, several constituent-to-dependency conversion schemes are pro- posed to construct corpora for dependency parsing. It is hard to determine which scheme is better because they reflect different views of dependency analysis. We usually obtain dependen- cy parsers of different schemes by training with the specific corpus separately. It neglects the correlations between these schemes, which can potentially benefit the parsers. In this paper, we study how these correlations influence final dependency parsing performances, by proposing a joint model which can make full use of the correlations between heterogeneous dependencies, and finally we can answer the following question: parsing heterogeneous dependencies jointly or separately, which is better? We conduct experiments with two different schemes on the Penn Treebank and the Chinese Penn Treebank respectively, arriving at the same conclusion that joint- ly parsing heterogeneous dependencies can give improved performances for both schemes over the individual models. 1 Introduction Dependency parsing has been intensively studied in recent years (McDonald et al., 2005; Nivre, 2008; Zhang and Clark, 2008; Huang et al., 2009; Koo and Collins, 2010; Zhang and Nivre, 2011; Sartorio et al., 2013; Choi and McCallum, 2013; Martins et al., 2013). Widely-used corpus for training a dependen- cy parser is usually constructed according to a specific constituent-to-dependency conversion scheme. Several conversion schemes for certain languages have been available. For example, the English lan- guage has at least four schemes based on the Penn Treebank (PTB), including the Yamada scheme (Ya- mada and Matsumoto, 2003), the CoNLL 2007 scheme (Nilsson et al., 2007), the Stanford scheme (de Marneffe and Manning, 2008) and the LTH scheme (Johansson and Nugues, 2007). There are dif- ferent conversion schemes for the Chinese Penn Treebank (CTB) as well, including the Zhang scheme (Zhang and Clark, 2008) and the Stanford scheme (de Marneffe and Manning, 2008). It is hard to judge which scheme is more superior, because each scheme reflects a specific view of dependency analy- sis, and also there is another fact that different natural language processing (NLP) applications can prefer different conversion schemes (Elming et al., 2013). Traditionally, we get dependency parsers of different schemes by training with the specific corpus separately. The method neglects the correlations between these schemes, which can potentially help different dependency parsers. On the one hand, there are many consistent dependencies across heteroge- neous dependency trees. Some dependency structures remain constant in different conversion schemes. Taking the Yamada and the Stanford schemes as an example, overall 70.27% of the dependencies are identical (ignoring the dependency labels), according to our experimental analysis. We show a concrete example for the two heterogeneous dependency trees in Figure 1, …

Content truncated to 500 words