Vancouver Welcomes You! Minimalist Location Metonymy Resolution

Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics, pages 1248–1259 Vancouver, Canada, July 30 - August 4, 2017. c⃝2017 Association for Computational Linguistics https://doi.org/10.18653/v1/P17-1115 Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics, pages 1248–1259 Vancouver, Canada, July 30 - August 4, 2017. c⃝2017 Association for Computational Linguistics https://doi.org/10.18653/v1/P17-1115 Vancouver Welcomes You! Minimalist Location Metonymy Resolution Milan Gritta, Mohammad Taher Pilehvar, Nut Limsopatham and Nigel Collier Language Technology Lab Department of Theoretical and Applied Linguistics University of Cambridge {mg711,mp792,nl347,nhc30}@cam.ac.uk Abstract Named entities are frequently used in a metonymic manner. They serve as ref- erences to related entities such as peo- ple and organisations. Accurate identifi- cation and interpretation of metonymy can be directly beneficial to various NLP ap- plications, such as Named Entity Recog- nition and Geographical Parsing. Until now, metonymy resolution (MR) methods mainly relied on parsers, taggers, dictio- naries, external word lists and other hand- crafted lexical resources. We show how a minimalist neural approach combined with a novel predicate window method can achieve competitive results on the Se- mEval 2007 task on Metonymy Resolu- tion. Additionally, we contribute with a new Wikipedia-based MR dataset called RelocaR, which is tailored towards loca- tions as well as improving previous defi- ciencies in annotation guidelines. 1 Introduction In everyday language, we come across many types of figurative speech. These irregular expressions are understood with little difficulty by humans but require special attention in NLP. One of these is metonymy, a type of common figurative language, which stands for the substitution of the concept, phrase or word being meant with a semantically related one. For example, in “Moscow traded gas and aluminium with Beijing.”, both location names were substituted in place of governments. Named Entity Recognition (NER) taggers have no provision for handling metonymy, meaning that this frequent linguistic phenomenon goes largely undetected within current NLP. Classi- fication decisions presently focus on the entity using features such as orthography to infer its word sense, largely ignoring the context, which provides the strongest clue about whether a word is used metonymically. A common classifica- tion approach is choosing the N words to the immediate left and right of the entity or the whole paragraph as input to the model. However, this “greedy” approach also processes input that should in practice be ignored. Metonymy is problematic for applications such as Geographical Parsing (Monteiro et al., 2016; Gritta et al., 2017, GP) and other information extraction tasks in NLP. In order to accurately identify and ground location entities, for example, we must recognise that metonymic entities consti- tute false positives and should not be treated the same way as regular locations. For example, in “London voted for the change.”, London refers to the concept of “people” and should not be classified as a location. There are many types of metonymy (Shutova et al., 2013), however, in this paper, we primarily address metonymic location mentions with reference to GP and NER. Contributions: (1) We investigate how to improve …

Content truncated to 500 words