Development of a Multilingual CCG Treebank via Universal Dependencies Conversion

Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022), pages 5220–5233 Marseille, 20-25 June 2022 © European Language Resources Association (ELRA), licensed under CC-BY-NC-4.0 5220 Development of a Multilingual CCG Treebank via Universal Dependencies Conversion Tu-Anh Tran, Yusuke Miyao The University of Tokyo 7-3-1 Hongo, Bunkyo, Tokyo, Japan {anh, yusuke}@is.s.u-tokyo.ac.jp Abstract This paper introduces an algorithm to convert Universal Dependencies (UD) treebanks to Combinatory Categorial Grammar (CCG) treebanks. As CCG encodes almost all grammatical information into the lexicon, obtaining a high-quality CCG derivation from a dependency tree is a challenging task. Our algorithm relies on hand-crafted rules to assign categories to constituents, and a non-statistical parser to derive full CCG parses given the assigned categories. To evaluate our converted treebanks, we perform lexical, sentential, and syntactic rule coverage analysis, as well as CCG parsing experiments. Finally, we discuss how our method handles complex constructions, and propose possible future extensions. Keywords: treebank, combinatory categorial grammar, universal dependencies 1. Introduction Combinatory Categorial Grammar (Steedman, 2000) is a lexicalized grammar formalism that can capture both syntactic and semantic information, while allowing fast and efficient parsing. Derived syntactic structures and semantic representations can be used for various downstream tasks without task-specific training data, such as question answering (Clark et al., 2004), re- lation extraction (Krishnamurthy and Mitchell, 2012), and recognizing textual entailment (Mart´ınez-G´omez et al., 2017). The English CCGbank (Hockenmaier and Steedman, 2007), one of the first available treebanks for CCG, plays an important role in the development of many wide-coverage CCG parsers for English. Having a similar resource for other languages and domains ac- celerates NLP research, in particular on resource-scarce languages/domains where one cannot rely on massive training data needed for training large neural network models (Peters et al., 2018; Devlin et al., 2019). Multi- lingual CCG resources also contribute to cross-linguistic research on syntactic/semantic theories and multilingual CCG parsing. Since manual annotation is expensive, conversion from a source treebank is a preferable approach. Besides En- glish, independent works have been done in the past to create CCG treebanks for several languages from source treebanks of different grammar formalisms, such as for German (Hockenmaier, 2006), Italian (Bos et al., 2009), Chinese (Tse and Curran, 2010), Japanese (Uematsu et al., 2013), and Hindi (Ambati et al., 2018). Such works often involve conversion rules that are specific to the languages and treebanks being converted, making the process difficult to adapt and generalize to others. In this paper, we propose a method to create a mul- tilingual collection of CCG treebanks by converting from dependency treebanks. To minimize the need for language-specific conversion rules, we select the Uni- versal Dependencies (Nivre et al., 2016) as our source treebanks. UD, as of v2.9, contains over 200 treebanks in 122 languages that follow cross-linguistically consis- tent annotation guidelines.1 Our goal is to develop a universal set of hand-crafted rules that can be applied to a wide range of languages in UD, while sacrificing as little as possible the conversion quality and coverage of each …

Content truncated to 500 words