Modelling Polysemy in Adjective Classes by Multi-Label Classification

Gemma Boleda, Sabine Schulte im Walde, Toni BadiaEMNLP 2007

TL;DR

This paper tackles the challenge of modeling polysemy (words having multiple meanings) in adjectives by framing it as a multi-label classification problem. They demonstrate that combining various linguistic features (morphology, syntax, semantics) through ensemble methods significantly boosts classification accuracy, outperforming single-feature approaches.

What Problem Does This Paper Address?

Polysemy is a pervasive phenomenon in natural language, meaning many words, including adjectives, have multiple distinct meanings. However, most previous work on automatic semantic class acquisition has largely ignored this issue, focusing on single-meaning assignments. This paper addresses the gap in modeling polysemy for adjectives, specifically for Catalan, by proposing and evaluating a multi-label classification architecture. This is crucial because adjectives play a vital role in defining entities and their properties, impacting tasks like information extraction and reference resolution.

Methods

The core of the methodology is a two-level multi-label classification architecture. First, binary decisions are made for each potential semantic class (basic, event-related, object-related). Then, these individual decisions are combined to form a final, potentially multi-label classification that accounts for polysemy.

Key techniques include:

  • Multi-label Classification Architecture: This is the central framework, allowing an adjective to be assigned to more than one class simultaneously.
  • Decision Trees (J48 algorithm): Used as the base classifier for the binary decisions, chosen for their transparency and ease of error analysis.
  • Feature Engineering: Features were extracted from five levels of linguistic description:
  • Morphology: Derivational type (denominal, deverbal, etc.) and suffix.
  • Syntax: Syntactic function (modifier, predicative) and part-of-speech unigrams/bigrams around the adjective.
  • Semantics: Heuristic cues related to properties like gradability, non-restrictivity, etc.
  • All: A combination of features from all levels.
  • Feature Selection: A WrapperSubsetEval method was used to select the most informative features for each classifier.
  • Ensemble Classifier: A majority voting ensemble classifier was employed to combine the predictions from different feature sets, aiming to leverage the strengths of each linguistic level.

Datasets

  • Gold Standard: A manually classified dataset of 210 Catalan adjectives was created by three experts. The adjectives were sampled to balance frequency, morphological type, and suffix. Each adjective was tagged as belonging to one of six categories: basic (B), event (E), object (O), or polysemous combinations (BE, BO, EO).
  • Corpus: The paper mentions using a research corpus from the Institut d'Estudis Catalans, likely used for feature extraction.

Key Insights and Ideas

  • Polysemy Naturally Fits Multi-Label: The paper strongly advocates that multi-label classification is a suitable and effective framework for modeling adjective polysemy, as it allows for simultaneous assignment to multiple semantic categories.
  • Ensemble Power: Combining different types of linguistic evidence (morphology, syntax, semantics) through ensemble methods significantly outperforms any single type of information. This suggests that no single linguistic level holds all the answers for adjective classification.
  • Morphology is Important, but Not Sufficient: While morphological features alone provided strong results, especially for distinguishing basic and object adjectives, they struggled with finer distinctions and couldn't fully capture all nuances of polysemy.
  • Distributional Cues are Sensitive but Not Always Robust: Syntactic and semantic features, which capture distributional properties, are sensitive to deviations from expected mappings but often lack robustness in distinguishing between basic and event classes.
  • Majority Voting Ensemble is Highly Effective: A simple majority voting ensemble classifier achieved a dramatic improvement in accuracy, demonstrating that combining diverse linguistic insights is more critical than the specific features used.

Key Results

  • Binary Classifiers: The best single binary classifier achieved an accuracy of 75.3% (level 'all') for the basic class, 89.4% for the event class, and 85.4% for the object class. The 'morph' level also performed very well, particularly for the event class (89.1%).
  • Combined Classifier (Decision Tree): The best single decision tree classifier (using the 'all' feature set) achieved a full accuracy of 62.3% and a partial accuracy of 90.7%. This was a significant improvement over the baseline.
  • Ensemble Classifier: The majority voting ensemble classifier, combining all levels of description, achieved a full accuracy of 84% and a partial accuracy of 95.7%. This represents a raw improvement of 33% over the baseline and 21.7% over the best single classifier.

Limitations and Future Work

  • Limited Number of Classes: The paper acknowledges that the classification into three broad semantic classes might be an oversimplification of adjective semantics.
  • Dataset Bias: The sampling procedure for the Gold Standard explicitly balanced for morphological factors, potentially underrepresenting denominal and participial adjectives and overrepresenting non-derived and deverbal ones. This might have influenced the perceived importance of morphology.
  • Robustness of Distinctions: The paper notes that distributional differences between basic and event classes are not always robust, impacting classification accuracy.
  • Future Work:
    • Refinement of Classification: Further refining the classification itself based on experimental results.
  • Additional Linguistic Evidence: Incorporating other linguistic cues like selectional restrictions.
  • Broader NLP Applications: Applying the acquired semantic class information to tasks like POS-tagging disambiguation and terminology extraction.
  • Finer-Grained Classification: Extending the multi-label approach to a larger number of semantic classes.

Summary updated

BibTeX
@inproceedings{boleda2007modelling,
  author = {Gemma Boleda and Sabine Schulte im Walde and Toni Badia},
  title = {{Modelling Polysemy in Adjective Classes by Multi-Label Classification}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2007},
}

Loading related content...

Figure