A Multi-word Expression Dataset for Swedish

Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pages 4402–4409 Marseille, 11–16 May 2020 c⃝ European Language Resources Association (ELRA), licensed under CC-BY-NC 4402 A Multi-word Expression Dataset for Swedish Murathan Kurfalı, Robert ¨Ostling, Johan Sjons, Mats Wir´en† Department of Linguistics, Stockholm University Universitetsv¨agen 10C, SE-106 91 Stockholm {murathan.kurfali, robert, johan.sjons, mats.wiren}@ling.su.se Abstract We present a new set of 96 Swedish multi-word expressions annotated with degree of (non-)compositionality. In contrast to most previous compositionality datasets we also consider syntactically complex constructions and publish a formal specification of each expression. This allows evaluation of computational models beyond word bigrams, which have so far been the norm. Finally, we use the annotations to evaluate a system for automatic compositionality estimation based on distributional semantics. Our analysis of the disagreements between human annotators and the distributional model reveal interesting questions related to the perception of compositionality, and should be informative to future work in the area. Keywords: multi-word expressions, compositionality, distributional semantics 1. Introduction A major challenge to the interpretation of natural language for humans and computers alike is non-compositionality. For instance, the compositional meaning of the verb-object pair break the ice could be paraphrased as “clearing the ice from the sea so that ships could pass”. However, the same expression also has an established sense, “to overcome ini- tial stiffness or reserve in a social setting” (Ammer, 2013). In this example, the literal (compositional) sense is very distant from the idiomatic (non-compositional) sense. In other cases, such as “search engine,” only relatively minor semantic extensions of the inherent senses of ‘search’ and ‘engine’ are required to match the established meaning of “search engine.” Being able to estimate the level of com- positionality is important for several applications, includ- ing second language learning where idiomaticity is a major obstacle. However, the existing practical data is very lim- ited in size and only available for several high resource lan- guages. In the current study, we present the first resource for multi-word expressions (MWEs) in Swedish annotated for compositionality by humans1. 2. Background Several previous studies have performed computational in- vestigations of compositionality, and in some cases cre- ated extensive annotated resources of multi-word expres- sions. The main focus in these studies has been on sim- ple word bi-gram constructions (adjective–noun and noun– noun compounds as well as verb–particle pairs). Early computational work on compositionality relied on binary classifications by the authors, as in the semi-manual method of Korkontzelos and Manandhar (2009) to classify MWEs from WordNet (Fellbaum, 1998) as either compositional or † Author order is alphabetical. Author contributions are as follows. Conception of study: MK, JS; survey design and data collection: MK, R ¨O, JS; implementation of distributional model: MK; formal specification of expressions: R ¨O; drafting manuscript: MK, R ¨O, JS, MW; final approval: MK, R ¨O, JS, MW. 1The dataset can be accessed at https://github.com/MurathanKurfali/swedish-mwe-dataset non-compositional. Reddy et al. (2011) obtained composi- tionality scores for 90 English two-word noun compounds and their constituent words, using the Amazon …

Content truncated to 500 words