Typological Features for Multilingual Delexicalised Dependency Parsing

Proceedings of NAACL-HLT 2019, pages 3919–3930 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 3919 Typological Features for Multilingual Delexicalised Dependency Parsing Manon Scholivet Franck Dary Alexis Nasr Benoit Favre Carlos Ramisch Aix-Marseille Univ, Universit´e de Toulon, CNRS, LIS, Marseille, France firstname.lastname@lis-lab.fr Abstract The existence of universal models to describe the syntax of languages has been debated for decades. The availability of resources such as the Universal Dependencies treebanks and the World Atlas of Language Structures make it possible to study the plausibility of univer- sal grammar from the perspective of depen- dency parsing. Our work investigates the use of high-level language descriptions in the form of typological features for multilingual depen- dency parsing. Our experiments on multilin- gual parsing for 40 languages show that typo- logical information can indeed guide parsers to share information between similar languages beyond simple language identification. 1 Introduction Human languages may share some syntactic fea- tures, but differ on others. For example, some lan- guages tend to place the subject before the verb (e.g., English) whereas others favour the reverse order (e.g., Arabic), and some do not exhibit a clear preference (e.g., Polish). These features can be viewed as the parameters of a language’s syn- tax (Greenberg, 1963; Chomsky, 1995). When training a multilingual parser, it could be interesting to explicitly represent these parame- ters, and to integrate them into the parsing model. If a successful strategy to do so was found, then, a parser could be trained simultaneously on several languages whose syntactic parameters have been explicitly represented. Such parser could then use a single model to parse texts in any language with known syntactic parameters. In theory, if we had at our disposal a set of pa- rameters that completely describes the syntax of languages as well as treebanks that explore the whole space of parameters and their values, then such a universal parser could be designed. To make such a program realistic, though, several is- sues have to be addressed. In this paper, we pro- pose to study the feasibility of learning such multi- lingual parser by addressing some of these issues. The first one is the choice of syntactic pa- rameters that will be used (Naseem et al., 2012; T¨ackstr¨om et al., 2013; Zhang and Barzilay, 2015). In our work, we approximate these pa- rameters by extracting syntactic information from the World Atlas of Language Structures (WALS) (Dryer and Haspelmath, 2013). 1 A language is represented by a vector containing the values it se- lects in the WALS. This vector plays the role of the parameters mentioned above. The second issue is the design of a unified scheme for representing syntax. Our natural choice is the Universal Dependencies (UD) initia- tive. 2 UD specifically proposes a set of univer- sal dependency relations, part-of-speech tags and morphological features (Nivre et al., 2016). The UD treebanks are available for many languages, annotated according to common guidelines. The third issue is the lexicon. UD proposes a common language …

Content truncated to 500 words