Fleshing it out: A Supervised Approach to MWE-token and MWE-type Classification

Richard Fothergill, Timothy BaldwinIJCNLP 2011 · OpenReview

Abstract: Richard Fothergill, Timothy Baldwin. Proceedings of 5th International Joint Conference on Natural Language Processing.

2011.

Introduction

Multiword Expressions (MWEs), such as "kick the bucket" or "traffic light," present a significant challenge for Natural Language Processing (NLP) because their meanings often exceed the sum of their individual components. A critical task is MWE-token disambiguation: determining whether a specific instance of an MWE is used idiomatically (e.g., "he died") or literally (e.g., "he struck a pail").

Traditionally, this task suffers from a "knowledge acquisition bottleneck," requiring extensive labeled data for every unique idiom. This paper investigates crosstype classification, exploring whether a model trained on a set of known idioms can successfully classify a completely different, unseen idiom.

Methods and Feature Engineering

The authors employed Support Vector Machines (SVMs) to compare two strategies: type-specialized (trained and tested on the same MWEs) and crosstype (trained on one set and tested on a disjoint set of unseen MWEs).

The success of the system relies on three categories of features:

  1. WSD Features: Contextual features including bag-of-words, local sequences, and syntactic relations, extended with hypernyms (semantic categories) and domain information.
  2. Idiom Features: Indicators of linguistic "fixedness," such as voice (passive/active), negation, and adnominal modification.
  3. Type Features: Metadata about the MWE itself (e.g., the headword) to help the model identify similarities across different expressions.
Linguistic feature extraction for the Japanese MWE 'soccer no ude o ageru' (to improve one's soccer skills), showing semantic categories like "body part" and "sports".

Key Insights

  • Semantics Over Syntax: Semantic context features (WSD-style) proved far more effective for generalizing to unseen idioms than lexico-syntactic "fixedness" features. This suggests that the surrounding discourse is a more universal signal of idiomaticity than the internal structure of the idiom itself.
  • Categorical Generalization: Incorporating hypernyms and domain information significantly boosted performance. By learning that a specific context relates to a "sports domain," the model can apply that logic to entirely new MWEs within the same domain.
  • Breaking the Bottleneck: The crosstype classifier achieved performance levels comparable to baselines that possessed prior knowledge of the specific MWE's distribution, marking a significant step toward overcoming the data hurdle.

Results

The experimental results demonstrated that the crosstype classifier reached an accuracy of 74.5% on unseen MWEs, substantially outperforming the corpus baseline of 61.2%. When the model was allowed to see the target MWE during training (type-specialized), accuracy reached 89.0%.

Accuracy results for crosstype classification across different feature combinations, highlighting the strength of WSD features.

Conclusion

This research demonstrates that supervised models can generalize the concept of idiomaticity across different types of expressions. While a performance gap remains between crosstype and type-specialized models, the ability to achieve 74.5% accuracy on unseen idioms suggests that semantic signals are robust enough to facilitate "zero-shot" idiom detection in real-world NLP applications.

Summary updated

BibTeX
@inproceedings{fothergill2011fleshing,
  author = {Richard Fothergill and Timothy Baldwin},
  title = {{Fleshing it out: A Supervised Approach to MWE-token and MWE-type Classification}},
  booktitle = {IJCNLP},
  year = {2011},
  url = {https://openreview.net/forum?id=Ry-cHQpV7ymG},
}
Figure