Using Subcategorization to Resolve Verb Class Ambiguity

Mirella Lapata, Chris BrewEMNLP 1999 · OpenReview

Abstract

Levin's (1993) taxonomy of verbs and their classes is a widely used resource for lexical semantics. In her framework, some verbs, such as give exhibit no class ambiguity. But other verbs, such as write, can inhabit more than one class. In some of these ambiguous cases the appropriate class for a particular token of a verb is immediately obvious from inspection of the surrounding context. In others it is not, and an application which wants to recover this information will be forced to rely on some more or less elaborate process of inference. We present a simple statistical model of verb class ambiguity and show how it can be used to carry out such inference.

Introduction

A fundamental hypothesis in linguistics suggests that a verb's syntactic behavior—how it arranges its arguments—directly reflects its underlying meaning. This paper by Maria Lapata and Chris Brew explores this link by developing a probabilistic model that connects syntactic "frames" to semantic classes.

In lexical semantics, verbs are grouped into classes based on diathesis alternations, which are variations in argument expression that preserve core meaning (e.g., the causative/inchoative alternation: "Janet broke the cup" vs. "The cup broke"). However, many common verbs are polysemous, belonging to multiple classes. While a verb like give is unambiguous, write can belong to a "Message Transfer" class or a "Performance" class.

The scale of this ambiguity is significant. As shown in the table below, while the majority of verbs in Levin's taxonomy belong to a single class, hundreds of verbs inhabit two or more classes, creating a challenge for automated semantic analysis.

Distribution of verb class ambiguity in the Levin taxonomy and BNC.

The Probabilistic Model

The authors address this by quantifying the relationship between verbs ($v$), syntactic frames ($f$), and semantic classes ($c$). They propose a model to calculate the joint probability $P(v, f, c)$. Using Bayesian decomposition and a key independence assumption—that the semantic class determines the argument structure regardless of the specific verb—the model simplifies the inference process.

To handle cases where the syntactic frame is shared by multiple classes (Category B ambiguity), the model treats the class as a "hidden variable" and relies on statistical distributions to predict the most likely meaning.

Data and Parameter Estimation

The model relies on two primary resources:

  1. Levin (1993): A taxonomy of 3,024 English verbs organized into approximately 200 classes.
  2. British National Corpus (BNC): A 100-million-word corpus used to derive real-world frequency counts.

Since the BNC is not annotated with semantic classes, the authors used "ambiguity classes" (groups of verbs sharing the same potential Levin classes) to estimate probabilities. They extracted verb and frame frequencies using Gsearch. The resulting distributions, illustrated below, show the varying frequencies of semantic classes and syntactic frames across the corpus.

Frequencies of semantic classes and syntactic frames derived from the British National Corpus.

Results and Evaluation

The model was evaluated against a baseline that simply selects the most frequent class for a given verb.

  • Category A (Frame-disambiguated): The model achieved 91.8% accuracy, significantly outperforming the 65.7% baseline.
  • Category B (Genuinely ambiguous): The model achieved 83.9% accuracy, compared to a 61.3% baseline.

Qualitatively, the model's preferences aligned well with human judgment. For example, it correctly identified that save is more frequently used in the "GET" sense than the "BILL" sense.

Conclusion

This research demonstrates that syntax is a powerful proxy for semantics. By leveraging the distribution of syntactic frames, NLP systems can effectively resolve verb class ambiguity. Future work aims to refine the model by relaxing independence assumptions and incorporating selectional restrictions—the semantic content of the arguments themselves—to further improve disambiguation.

Summary updated

Abstract

Levin's (1993) taxonomy of verbs and their classes is a widely used resource for lexical semantics. In her framework, some verbs, such as give exhibit no class ambiguity. But other verbs, such as write, can inhabit more than one class. In some of these ambiguous cases the appropriate class for a particular token of a verb is immediately obvious from inspection of the surrounding context. In others it is not, and an application which wants to recover this information will be forced to rely on some more or less elaborate process of inference. We present a simple statistical model of verb class ambiguity and show how it can be used to carry out such inference.
BibTeX
@inproceedings{lapata1999using,
  author = {Mirella Lapata and Chris Brew},
  title = {{Using Subcategorization to Resolve Verb Class Ambiguity}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {1999},
  url = {https://openreview.net/forum?id=SkW4yffuWB},
}
Figure