Title:Casting a Wide Net: Robust Extraction of Potentially Idiomatic Expressions

Abstract:Idiomatic expressions like `out of the woods' and `up the ante' present a range of difficulties for natural language processing applications. We present work on the annotation and extraction of what we term potentially idiomatic expressions (PIEs), a subclass of multiword expressions covering both literal and non-literal uses of idiomatic expressions. Existing corpora of PIEs are small and have limited coverage of different PIE types, which hampers research. To further progress on the extraction and disambiguation of potentially idiomatic expressions, larger corpora of PIEs are required. In addition, larger corpora are a potential source for valuable linguistic insights into idiomatic expressions and their variability. We propose automatic tools to facilitate the building of larger PIE corpora, by investigating the feasibility of using dictionary-based extraction of PIEs as a pre-extraction tool for English. We do this by assessing the reliability and coverage of idiom dictionaries, the annotation of a PIE corpus, and the automatic extraction of PIEs from a large corpus. Results show that combinations of dictionaries are a reliable source of idiomatic expressions, that PIEs can be annotated with a high reliability (0.74-0.91 Fleiss' Kappa), and that parse-based PIE extraction yields highly accurate performance (88% F1-score). Combining complementary PIE extraction methods increases reliability further, to over 92% F1-score. Moreover, the extraction method presented here could be extended to other types of multiword expressions and to other languages, given that sufficient NLP tools are available.