Construction of Large-scale English Verbal Multiword Expression Annotated Corpus

Construction of Large-scale English Verbal Multiword Expression Annotated Corpus Akihiko Kato, Hiroyuki Shindo, Yuji Matsumoto Graduate School of Information and Science Nara Institute of Science and Technology 8916-5, Takayama, Ikoma, Nara, 630-0192, Japan {kato.akihiko.ju6, shindo, matsu} @is.naist.jp Abstract Multiword expressions (MWEs) consist of groups of tokens, which should be treated as a single syntactic or semantic unit. In this work, we focus on verbal MWEs (VMWEs), whose accurate recognition is challenging because they could be discontinuous (e.g., take .. off). Since previous English VMWE annotations are relatively small-scale in terms of VMWE occurrences and types, we conduct large-scale annotations of VMWEs on the Wall Street Journal portion of English Ontonotes by a combination of automatic annotations and crowdsourcing. Concretely, we first construct a VMWE dictionary based on the English-language Wiktionary. After that, we collect possible VMWE occurrences in Ontonotes and filter candidates with the help of gold dependency trees, then we formalize VMWE annotations as a multiword sense disambiguation problem to exploit crowdsourcing. As a result, we annotate 7,833 VMWE instances belonging to various categories, such as phrasal verbs, light verb constructions, and semi-fixed VMWEs. We hope this large-scale VMWE-annotated resource helps to develop models for MWE recognition and dependency parsing that are aware of English MWEs. Our resource is publicly available. Keywords: Multiword expressions, Phrasal verbs, Crowdsourcing 1. Introduction Multiword expressions (MWEs) consist of groups of to- kens, which should be treated as a single syntactic or se- mantic unit. MWEs are also known as “idiosyncratic inter- pretations that cross word boundaries” (Sag et al., 2002). In this paper, we focus on verbal MWEs (VMWEs) among various types of MWEs, such as compound nouns and com- pound function words. An accurate recognition of VMWEs is challenging because VMWEs could be discontinuous (e.g., take .. off). We show the main categories of VMWEs in Table 1. While dependency parsing and MWE recognition could be solved independently, dependency struc- tures in that each MWE is a syntactic unit are preferable to word-based dependency structures for downstream NLP tasks, such as semantic parsing. Because MWE recognition could help syntactic pars- ing (Nivre and Nilsson, 2004; Eryi˘git et al., 2011), sev- eral works tackle MWE-aware dependency parsing in French (Candito and Constant, 2014; Nasr et al., 2015). They use French Treebank (Abeill´e et al., 2003) because of its explicit MWE annotations. Regarding English MWEs, Schneider et al. (2014) con- structs an MWE-annotated corpus based on English Web Treebank (Bies et al., 2012). However, the number of VMWE occurrences (1,444) and types (1,155) in their cor- pus is relatively small-scale. In this work, we conduct full-scale VMWE annotations on the Wall Street Journal (WSJ) portion of English Ontonotes (Pradhan et al., 2007), which results in 7,833 VMWE occurrences and 1,608 types. Concretely, we con- struct a VMWE dictionary based on the English-language Wiktionary 1. Based on this dictionary, we collect possible 1https://en.wiktionary.org VMWE occurrences from Ontonotes and filter candidates with the help of gold dependency trees. To exploit crowd- sourcing, we formalize VMWE annotations …

Content truncated to 500 words