Combining Collocations, Lexical and Encyclopedic Knowledge for Metonymy Resolution

Vivi Nastase, Michael StrubeEMNLP 2009

Metonymy is a linguistic phenomenon where one entity is used to refer to another related entity—for instance, saying "The White House issued a statement" when referring to the administration within the building. For Natural Language Processing (NLP) systems, resolving these references is a significant hurdle for tasks like question answering and machine translation.

This paper presents a hybrid approach to metonymy resolution, determining when a word refers to something other than its literal meaning. By enriching a standard supervised model with unsupervised features derived from the British National Corpus (BNC), WordNet, and Wikipedia, the authors achieved state-of-the-art results on location-based metonymy.

The Challenge of Context

The core difficulty in metonymy resolution is its high dependence on context. While supervised machine learning is a standard tool, manually annotated data is scarce, leading to "data sparseness." If a model only encounters "Kawasaki" in its training set, it may struggle to generalize to "Yamaha." Previous solutions often relied on rigid identity constraints. This research seeks to bridge that gap by using large, unannotated corpora and encyclopedic knowledge to generalize what a word's "neighbors" imply about its intended meaning.

Methodology: Hybrid Supervised Learning

The authors utilize a supervised learning framework based on Support Vector Machines (SVMs), enhanced by unsupervised feature enrichment. The approach consists of three main components:

  1. M&N Features: The system starts with the standard feature set from Markert and Nissim (2005), including grammatical roles (subject, object), lemmatized heads, and determiners.
  2. Selectional Preference Features: This is the core novelty. The authors extract grammatical collocations from the BNC. For a verb like excel, they analyze its typical subjects (e.g., player, car, computer). Using WordNet and Wikipedia, they generalize these into "supersenses" (e.g., person, artifact). This allows the system to calculate the probability that a specific grammatical context prefers a certain interpretation.
  3. Encyclopedic Knowledge (Wikipedia): Wikipedia is mined to create binary features such as has-product (suggesting a manufacturer) and has-event (associating a location with specific historical occurrences).
Feature enrichment example showing how lemmas are mapped to supersense probabilities like 'person' or 'artifact' to aid resolution.

Key Insights

  • Generalization is Key: Instead of looking for exact word matches, the model asks if a verb usually takes a person or an object as a subject. This abstraction helps the model handle unseen entities.
  • Wikipedia as a Gap-Filler: While WordNet is excellent for general nouns, it lacks specific named entities. Wikipedia provides the necessary "encyclopedic" layer to identify that a specific company is a manufacturer or a city is a frequent site of political events.
  • Local vs. Global Context: Local grammatical relations (like subject-verb) are highly effective for locations but less definitive for companies, which often require broader discourse analysis.

Evaluation and Results

The system, labeled SMO<sub>wiki</sub>, was evaluated using the SemEval 2007 Task 8 benchmark, which covers both Locations and Organizations.

Detailed performance comparison of the SMO_wiki model against baselines across various granularities of metonymy resolution.

On the Location task, the model outperformed previous state-of-the-art systems across all levels of granularity. For the "coarse" task (literal vs. non-literal), it achieved 86.1% accuracy, surpassing the previous SemEval best of 85.2%. The model showed significant gains in recognizing "mixed" and "metonymic" readings, with the F-score for the non-literal class rising by 1.5 percentage points.

For Organizations, results were competitive but slightly below the state-of-the-art (74.9% vs 76.7%). The authors hypothesize that company metonymy relies more on global discourse links than local grammatical ones, suggesting that future work should integrate coreference resolution to track how metonymic interpretations evolve across a full text.

Summary updated

BibTeX
@inproceedings{nastase2009combining,
  author = {Vivi Nastase and Michael Strube},
  title = {{Combining Collocations, Lexical and Encyclopedic Knowledge for Metonymy Resolution}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2009},
}
Figure