Verb Noun Construction MWE Token Classification

Mona T. Diab, Pravin BhutadaMWE@IJCNLP 2009 · OpenReview

Introduction

Multiword expressions (MWEs), such as "kick the bucket" or "make a decision," present a significant challenge in Natural Language Processing (NLP) due to their non-compositional nature. While traditional research often focuses on type-level idiomaticity, this study shifts the focus to token-level classification. The goal is to determine whether a specific instance of a Verb-Noun Construction (VNC) in a sentence is used literally or idiomatically. For example, "spilling the beans" can refer to revealing a secret (idiomatic) or an actual kitchen mishap (literal). Accurate token classification is essential for downstream tasks like Machine Translation to avoid semantic errors.

Methodology

The authors approach the problem as a sequence labeling task using the Inside-Outside-Beginning (IOB) tagging framework. The classification model is built using YamCha, a Support Vector Machine (SVM) toolkit utilizing polynomial kernels.

Features and Context

The model incorporates several linguistic features:

  • Linguistic Context: Part-of-Speech (POS) tags, lemmas, and character n-grams (suffix-based morphology).
  • Named Entity Recognition (NER): Two strategies were tested: Named Entity Separate (NES), where NE tags are added as features, and Named Entity InText (NEI), where specific entities are replaced by their class labels (e.g., "John" becomes "PER").
  • Context Windows: The authors experimented with various window sizes to determine the optimal amount of surrounding text needed for classification.

Dataset

The study utilizes a dataset from Cook et al. (2008) derived from the British National Corpus (BNC). It consists of 2,432 sentences covering 53 VNC types. The dataset is imbalanced, with 2,020 idiomatic tokens and 551 literal tokens. Evaluation was performed using 5-fold cross-validation.

Results and Analysis

The inclusion of NER features, particularly the NEI approach, proved highly effective. NEI acts as a form of dimensionality reduction, allowing the model to generalize across different contexts rather than overfitting to specific named entities.

Performance metrics for various feature combinations, showing the superiority of the NEI+L+N+P model.

As shown in the results table, the best-performing configuration (NEI+L+N+P) achieved an overall F-measure of 84.58%. A notable finding was the discrepancy in performance between classes: the model was significantly more accurate at identifying idioms (89.86% F-measure) than literal usages (62.03% F-measure). This suggests that literal usage is more linguistically diverse or that the data imbalance hinders the learning of literal patterns.

The authors also investigated the impact of context window size on classification performance.

Impact of context window size on classification metrics.

The experimental results indicate that a context window of +/- 3 tokens provides the optimal balance for classification accuracy, achieving the highest F-measures for both idiomatic and literal classes.

Conclusion

This research demonstrates that supervised token classification for VNCs is highly effective when leveraging context-aware linguistic and NER features. While the model achieves state-of-the-art results, the difficulty in classifying literal tokens remains a challenge, likely due to dataset imbalance. Future work may involve incorporating Word Sense Disambiguation (WSD) features and deeper syntactic analysis to further refine the distinction between literal and idiomatic usage.

Summary updated

BibTeX
@inproceedings{diab2009verb,
  author = {Mona T. Diab and Pravin Bhutada},
  title = {{Verb Noun Construction MWE Token Classification}},
  booktitle = {MWE@IJCNLP},
  year = {2009},
  url = {https://openreview.net/forum?id=hjFELNamSZ},
}
Figure