Exploring Metaphorical Senses and Word Representations for Identifying Metonyms

A metonym is a word with a figurative meaning, similar to a metaphor. Because metonyms are closely related to metaphors, we apply features that are used successfully for metaphor recognition to the task of detecting metonyms. On the ACL SemEval 2007 Task 8 data with gold standard metonym annotations, our system achieved 86.45% accuracy on the location metonyms. Our code can be found on GitHub.

💡 Research Summary

The paper proposes a novel framework for metonym detection that leverages features originally designed for metaphor recognition. Recognizing that both metaphors and metonyms involve a shift of meaning from a literal to a figurative sense, the authors adopt a set of lexical and semantic cues that have proven effective in metaphor identification and adapt them to the task of spotting metonyms, specifically location‑based metonyms such as “Washington” referring to the U.S. government.

The feature set consists of three main components. First, abstractness–concreteness scores are derived from established psycholinguistic norms (e.g., Brysbaert et al.) to quantify how abstract or concrete a target word and its surrounding context are. Second, semantic‑field information is extracted from WordNet, allowing the system to detect when a word’s typical domain (geography, politics, culture, etc.) is being used in an atypical domain, a hallmark of metonymic usage. Third, modern contextual word embeddings (BERT, RoBERTa, or similar transformer‑based models) are employed to capture fine‑grained, context‑dependent meaning shifts that static embeddings cannot represent. The authors combine the contextual vectors (either the CLS token or a mean‑pooled representation) with the lexical features into a single feature vector for each candidate instance.

For classification, a suite of supervised learners is evaluated, including linear Support Vector Machines, logistic regression, and a shallow multilayer perceptron. Hyper‑parameters are tuned via ten‑fold cross‑validation on the SemEval 2007 Task 8 dataset, which provides gold‑standard annotations for location metonyms. The baseline systems consist of traditional frequency‑based heuristics and rule‑based approaches previously reported in the literature.

Experimental results demonstrate that the proposed model achieves an accuracy of 86.45 % on the location‑metonym test set, surpassing the prior state‑of‑the‑art (approximately 77 %) by a substantial margin. Precision, recall, and F1 scores also improve across the board, with the most notable gains attributed to the synergy between abstractness/concreteness cues and contextual embeddings. Error analysis reveals that the remaining mistakes are largely concentrated on cases where a geographic name denotes an institution (e.g., “Paris” for the French fashion industry) or where cultural symbols are used to refer to physical places; these errors suggest that richer sense‑level information or deeper discourse modeling could further close the gap.

All code, preprocessing scripts, and trained models are released on GitHub, enabling reproducibility and facilitating future extensions. The authors argue that their work illustrates the broader potential of transferring insights between related figurative‑language tasks, and they outline future directions such as expanding to other metonym categories (organization, person, etc.), applying the approach to multilingual corpora, and integrating sense‑disambiguation modules or graph‑based semantic representations to capture more nuanced meaning shifts.