Abstract: In previous research in automatic verb classification, syntactic features have proved the most useful features, although manual classifications rely heavily on semantic features. We show, in contrast with previous work, that considerable additional improvement can be obtained by using semantic features in automatic classification: verb selectional preferences acquired from corpus data using a fully unsupervised method. We report these promising results using a new framework for verb clustering which incorporates a recent subcategorization acquisition system, rich syntactic-semantic feature sets, and a variation of spectral clustering which performs particularly well in high dimensional feature space.
Introduction
Grouping verbs into semantic classes—such as "verbs of motion" or "verbs of communication"—is a fundamental task in Natural Language Processing (NLP), with applications ranging from machine translation to word sense disambiguation. Traditionally, automatic classification systems have relied on Subcategorization Frames (SCFs), which describe the syntactic structures a verb can inhabit. However, human-made taxonomies like Levin's classification rely heavily on semantic meaning.
The challenge has been effectively integrating semantic features, specifically Selectional Preferences (SPs)—the tendency of verbs to prefer certain types of arguments (e.g., "eat" prefers "food" as an object). Previous attempts often struggled with data sparsity and the difficulties of clustering in high-dimensional spaces. This paper presents a framework that successfully leverages automatically acquired SPs to achieve state-of-the-art results in unsupervised verb clustering.
Methodology
The proposed framework integrates three core components to bridge the gap between syntactic and semantic features:
- Subcategorization Acquisition: A robust system extracts 168 distinct verbal SCFs from large-scale corpora, providing a detailed syntactic profile for each verb.
- Unsupervised Selectional Preferences: Rather than using hand-coded resources like WordNet, the authors acquire SPs by clustering the head nouns found in subject and object positions. This data-driven approach identifies semantic categories based on actual usage patterns.
- Spectral Clustering (SPEC): To handle the high-dimensional feature space created by combining SCFs and SPs, the authors utilize a variation of spectral clustering based on the Multiway Normalized Cut (MNCut) algorithm. SPEC is particularly effective at identifying global structures in complex data distributions where traditional pairwise methods might fail.
As shown in the table above, the unsupervised method successfully identifies coherent semantic groups such as "Human," "Authority," and "Body-part" from raw text data, which serve as the basis for the selectional preference features.
Experimental Setup
The system was evaluated on two benchmark datasets:
- T1 (Joanis et al., 2008): 205 verbs across 14 Levin classes, characterized by significant noise and sparsity.
- T2 (Sun et al., 2008): 204 verbs across 17 fine-grained Levin classes, designed for balanced cross-class comparison.
Results and Analysis
The inclusion of selectional preferences, combined with the SPEC algorithm, yielded substantial performance gains. On the T1 dataset, the system achieved an F-measure of 57.75, more than doubling the previous unsupervised best of 31.0. On the T2 dataset, the performance reached 80.35, surpassing the previous unsupervised baseline (51.6) and even outperforming the best-known supervised system (67.3) at the time.
Key Insights
- Feature Synergy: SCFs provide the structural framework, while SPs provide the semantic "meat." Together, they allow the system to distinguish between verbs that are syntactically similar but semantically distinct.
- Subject vs. Object: SPs for subject positions were found to be more influential than those for objects, likely due to higher data density for subjects in the training corpora.
- Algorithmic Impact: The choice of spectral clustering was critical. Standard clustering methods often degrade when high-dimensional semantic features are added; SPEC's ability to find global structures mitigated this issue.
Conclusion
This research demonstrates that automatically acquired semantic features can significantly enhance verb clustering when paired with appropriate high-dimensional clustering techniques. By moving beyond purely syntactic features and utilizing unsupervised selectional preferences, the system achieves a new level of accuracy that rivals or exceeds supervised methods. Future work will explore the impact of polysemy and the application of these methods to languages other than English.
Summary updated