Creating a Verb Synonym Lexicon Based on a Parallel Corpus

Creating a Verb Synonym Lexicon Based on a Parallel Corpus Zdeˇnka Urešová, Eva Fuˇcíková, Eva Hajiˇcová, Jan Hajiˇc Charles University, Faculty of Mathematics and Physics Institute of Formal and Applied Linguistics Malostranské nám. 25 11800 Prague 1, Czech Republic {uresova,fucikova,hajicova,hajic}@ufal.mff.cuni.cz Abstract This paper presents the first findings of our recently started project of building a new lexical resource called CzEngClass, which consists of bilingual verbal synonym groups. In order to create such a resource, we explore semantic ‘equivalence’ of verb senses (across different verb lexemes) in a bilingual (Czech-English) setting by using translational context of real-world texts in a parallel, richly annotated dependency corpus. When grouping semantically equivalent verb senses into classes of synonyms, we focus on valency (arguments as deep dependents with morphosyntactic features relevant for surface dependencies) and its mapping to a set of semantic “roles” for verb arguments, common within one class. We argue that the existence of core argument mappings and certain adjunct mappings to a common set of semantic roles is a suitable criterion for a reasonable verb synonymy definition, possibly accompanied with additional contextual restrictions. By mid-2018, the first version of the lexicon called CzEngClass will be publicly available. Keywords: Lexical Resource, Parallel Corpus, Semantics, Syntax, Synonymy, Valency 1. Introduction While synonym lexicons, such as the Roget’s thesaurus (Associates, 1988),1 or WordNet (Miller, 1995; Fellbaum, 1998) are well-known resources, used in both research and NLP applications, we believe that there is a significant gap in these resources (which concerns mainly verbs). We see two orthogonal problems: first, a granularity problem– current synonym classes (synsets, entries, frames, ...) as found in these resources are usually too broad (contain words which can be considered synonyms only in very specific contexts), or they use too fine-grained sense dis- tinctions, such as in WordNet, which often distinguish too many and too close verb senses (Palmer et al., 2007); sec- ond, a coverage problem - simply there are not enough verbs covered or some verb senses are missing.2 In our project, we attempt to introduce criteria, based mainly on valency or predicate-argument structure (taking into account both semantic and morphosyntactic proper- ties of verbs) to help define both the granularity of verb senses as well as the synonym classes themselves. Along with (Levin, 2013), we refer to predicate-argument struc- ture as to those structures that involve the realization of lexical item’s arguments, including morphosyntactic prop- erties that affect the morphosyntactic realization of argu- ments. This is similar to our valency approach, the FGD Valency Theory (FGDVT), first described in (Panevová, 1974). As (Levin, 2014) notes, the meaning of a verb may be characterized via the relations that its arguments bear to it and “Semantic Roles” (SRs) can be seen as labels for certain recurring predicate-argument relations. In addition, we have chosen to use multilingual, rather than monolingual-only evidence to help with both the sense dis- 1http://www.roget.org 2It is left for future investigation whether other parts of speech, especially nouns, would be subject to similar problems, but now we concentrate …

Content truncated to 500 words