Abstract: We reformulate metonymy resolution as a classification task. This is motivated by the regularity of metonymic readings and makes general classification and word sense disambiguation methods available for metonymy resolution. We then present a case study for location names, presenting both a corpus of location names annotated for metonymy as well as experiments with a supervised classification algorithm on this corpus. We especially explore the contribution of features used in word sense disambiguation to metonymy resolution.
Introduction
Metonymy is a figure of speech where one entity is used to refer to another related entity—for example, saying "The White House issued a statement" when referring to the administration. Traditionally, Natural Language Processing (NLP) viewed metonymy through the lens of "selectional restriction violations," where figurative language was only flagged when a literal interpretation was physically impossible (e.g., a "ham sandwich" cannot "wait for a check").
However, this traditional approach is often brittle and assumes metonymy is an unpredictable exception. Many metonymic expressions do not violate literal logic; for instance, "Vietnam" can refer to either a country or a war within the same grammatical context. This paper argues that metonymy is a regular linguistic pattern that can be addressed using supervised machine learning, moving away from complex, hand-coded rules.
Methodology
The authors treat metonymy resolution as a supervised classification task, similar to Word Sense Disambiguation (WSD). They employ Decision Lists, an algorithm that ranks features based on their predictive power (log-likelihood) and classifies test cases based on the single highest-ranking feature present. Three main feature types were explored:
- Co-occurrences: Words appearing within a specific window (e.g., 1 to 25 words) of the target location name.
- Collocations: Specific words immediately adjacent to the target, or short sequences of words (n-grams).
- Grammatical Features: The syntactic role of the location name (subject, direct object, etc.) and the specific verb it is associated with (e.g., "[Location] expressed readiness").
Data and Annotation
The researchers developed a benchmark corpus consisting of 1,000 occurrences of country names randomly sampled from the British National Corpus (BNC). After removing homonyms and other noise, 925 samples remained. Two linguists annotated these samples into categories such as literal, place-for-people, and place-for-event. The high inter-annotator agreement (Kappa = 0.88) demonstrates that these categories are well-defined and distinguishable.
Key Results and Insights
The study compared the supervised classifier against a baseline that assumes every usage is literal, which is correct 79.7% of the time for this dataset. The results indicate that syntactic context is the most powerful predictor for metonymy resolution.
As shown in the performance table, grammatical features outperformed simple bag-of-words approaches. The combination of grammatical role and the associated verb yielded the highest accuracy at 84.3%. While the grammatical model achieved high precision (75%), its recall was lower (33.9%) because many sentences lacked clear subject/object relationships.
Crucially, the supervised classifier proved superior to traditional rule-based systems by identifying metonymy in "neutral" sentences where no logical violation occurred. This confirms that metonymy resolution can be treated as a class-based WSD task, applying patterns across entire groups of words (like "locations") rather than individual lexemes.
Conclusion and Future Work
While the results are promising, data sparsity remains a challenge; many specific verb-location pairings appeared only once in the corpus. Future work could address this by using semantic resources like WordNet to group verbs into classes (e.g., "Communication Verbs"), allowing the model to generalize patterns like "[Location] [Communication Verb]" more effectively. Additionally, exploring more sophisticated algorithms and expanding the scope to other metonymic classes, such as "Author-for-Works," remains a key area for further research.
Summary updated