Fully Unsupervised Core-Adjunct Argument Classification

Omri Abend, Ari RappoportACL 2010 · OpenReview

Abstract

The core-adjunct argument distinction is a basic one in the theory of argument structure. The task of distinguishing between the two has strong relations to various basic NLP tasks such as syntactic parsing, semantic role labeling and subcategorization acquisition. This paper presents a novel unsupervised algorithm for the task that uses no supervised models, utilizing instead state-of-the-art syntactic induction algorithms. This is the first work to tackle this task in a fully unsupervised scenario.

Introduction

In linguistics, verbs possess "argument structures" that define the participants in an event. These participants are categorized as either core arguments—essential, predicate-specific participants (e.g., the location in "He put the book on the table")—or adjuncts—optional, general modifiers providing context like time or place (e.g., "He ate lunch in the park").

Distinguishing between cores and adjuncts is vital for Natural Language Processing (NLP) tasks such as semantic role labeling and syntactic parsing. While supervised models trained on human-annotated datasets like PropBank are accurate, they are expensive to develop and often lack domain portability. This research presents the first fully unsupervised algorithm for this task, learning the distinction entirely from raw, unannotated text.

Methodology

The proposed pipeline requires zero human intervention and consists of three primary stages:

  1. Unsupervised Pre-processing: The system utilizes the Seginer parser for inducing sentence structures and Clark’s POS tagger to cluster words into categories (e.g., verbs, prepositions) without relying on standard linguistic labels.
  2. Statistical Collocation Measures: Three "weak" classifiers are proposed based on distinct linguistic properties:
  3. Selectional Preference (SP): Measures the verb's preference for a specific argument head (e.g., "operate" preferring "patient").
  4. Predicate-Slot Collocation (PS): Measures how often a specific verb appears with a specific prepositional slot. Cores are typically more obligatory in these pairings.
  5. Argument-Slot Collocation (AS): Measures if an argument head (e.g., "Friday") tends to appear with a specific preposition (e.g., "on") regardless of the verb, indicating a predictable adjunct.
  6. Ensemble & Self-Training: These measures are combined using a majority-vote ensemble. To improve coverage—the ability to label unseen arguments—a self-training step is employed where high-confidence labels from the ensemble train a broader classifier.

Experimental Setup

The model was trained on the North American News Text Corpus (NANC), using a subset of approximately 5 million short sentences (up to 10 words). Evaluation was conducted against PropBank (sections 2–21 of the Wall Street Journal corpus), which served as the gold standard. Additionally, a web-crawled corpus of 100 million arguments was used in one variant to test the impact of data scale on Selectional Preference.

Results and Discussion

The experimental results demonstrate that unsupervised methods can effectively capture the core-adjunct distinction.

Performance of individual collocation measures and baselines.

A key insight is the efficacy of the Argument-Slot Collocation (AS) measure, which emerged as the most accurate single predictor. This suggests that the head word of an argument often provides more information regarding its adjunct status than the verb itself. Furthermore, focusing on prepositional arguments proved effective, as non-prepositional arguments (like direct objects) are almost always cores.

Performance of ensemble methods and self-training across different scenarios.

In the fully unsupervised scenario, the system achieved 70.6% accuracy on prepositional arguments. When including non-prepositional arguments, accuracy exceeded 80%. The model significantly outperformed the "Window Baseline," indicating that structural and collocation measures capture deeper linguistic truths than simple proximity. After self-training, the model reached an effective accuracy of 61.4% while achieving 100% coverage on the test set.

Conclusion

This work demonstrates the feasibility of fully unsupervised core-adjunct classification. While performance is currently limited by data sparsity in Selectional Preference measures and the quality of unsupervised parsing, the results are promising. Future work may involve integrating more sophisticated models of prepositional semantics and extending the approach to morphologically rich languages where case markings replace prepositions.

Summary updated

Abstract

The core-adjunct argument distinction is a basic one in the theory of argument structure. The task of distinguishing between the two has strong relations to various basic NLP tasks such as syntactic parsing, semantic role labeling and subcategorization acquisition. This paper presents a novel unsupervised algorithm for the task that uses no supervised models, utilizing instead state-of-the-art syntactic induction algorithms. This is the first work to tackle this task in a fully unsupervised scenario.
BibTeX
@inproceedings{abend2010fully,
  author = {Omri Abend and Ari Rappoport},
  title = {{Fully Unsupervised Core-Adjunct Argument Classification}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2010},
  url = {https://openreview.net/forum?id=By-A6ieuZr},
}
Figure