Multilingual and Cross-Lingual Graded Lexical Entailment

Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4963–4974 Florence, Italy, July 28 - August 2, 2019. c⃝2019 Association for Computational Linguistics 4963 Multilingual and Cross-Lingual Graded Lexical Entailment Ivan Vuli´c1, Simone Paolo Ponzetto2, Goran Glavaš2 1 PolyAI Ltd., London, United Kingdom 2 Data and Web Science Group, University of Mannheim, Germany ivan@poly-ai.com {simone, goran}@informatik.uni-mannheim.de Abstract Grounded in cognitive linguistics, graded lexi- cal entailment (GR-LE) is concerned with fine- grained assertions regarding the directional hi- erarchical relationships between concepts on a continuous scale. In this paper, we present the first work on cross-lingual generalisation of GR-LE relation. Starting from Hyper- Lex, the only available GR-LE dataset in En- glish, we construct new monolingual GR-LE datasets for three other languages, and com- bine those to create a set of six cross-lingual GR-LE datasets termed CL-HYPERLEX. We next present a novel method dubbed CLEAR (Cross-Lingual Lexical Entailment Attract- Repel) for effectively capturing graded (and bi- nary) LE, both monolingually in different lan- guages as well as across languages (i.e., on CL- HYPERLEX). Coupled with a bilingual dictio- nary, CLEAR leverages taxonomic LE knowl- edge in a resource-rich language (e.g., En- glish) and propagates it to other languages. Supported by cross-lingual LE transfer, CLEAR sets competitive baseline performance on three new monolingual GR-LE datasets and six cross-lingual GR-LE datasets. In addition, we show that CLEAR outperforms current state-of- the-art on binary cross-lingual LE detection by a wide margin for diverse language pairs. 1 Introduction Word-level lexical entailment (LE), also known as the TYPE-OF or hyponymy-hypernymy relation, is a fundamental asymmetric lexical relation (Collins and Quillian, 1972; Beckwith et al., 1991). It is a key principle behind the hierarchical structure found in semantic networks such as WordNet (Fell- baum, 1998) or ConceptNet (Speer et al., 2017). As opposed to simpler discrete and binary LE de- tection (e.g., oregano is a TYPE-OF food), graded lexical entailment (GR-LE) measures the strength of the LE relation between two concepts on a con- tinuous scale (Vuli´c et al., 2017; Rei et al., 2018). GR-LE is concerned with fine-grained directional assertions of hierarchical arrangements between concepts. The notion of graded LE is rooted in theories of concept (proto)typicality and category vagueness from cognitive science (Rosch, 1973, 1975; Kamp and Partee, 1995). Instead of answer- ing the simpler (discrete) question “Is X a type of Y?”, as in standard LE detection tasks (Kotler- man et al., 2010; Turney and Mohammad, 2015), GR-LE aims at answering the following question: “To what degree is X a type of Y?” The concept of LE gradience is also empirically confirmed by hu- man judgements elicited for HyperLex (Vuli´c et al., 2017), a GR-LE resource in English.1 Furthermore, while simpler binary LE detection has been predominantly studied in monolingual settings only (Geffet and Dagan, 2005; Weeds et al., 2014; Santus et al., 2014; Kiela et al., 2015; Shwartz et al., 2016, 2017; Glavaš and Ponzetto, 2017; Roller et al., 2018, inter alia), more general reasoning over cross-lingual and multilingual LE …

Content truncated to 500 words