Death and Lightness: Using a Demographic Model to Find Support Verbs

Mark Dras, Mike JohnsonarXiv 1996 · arXiv

Abstract

Identifying support verbs (SVs)—verbs that lose their independent semantic weight to function as a "prop" for a noun—is a persistent challenge in natural language processing. This paper introduces a statistical model that treats linguistic "lightness" as analogous to demographic mortality. By weighting local verb-noun frequencies against a "global lightness" score, the proposed Global Information Model significantly outperforms frequency-based baselines, achieving a 93% success rate on valid test cases.

Introduction

In natural language, verbs often exhibit binary ambiguity. In the sentence "Kim took a photograph," the verb take acts as a support verb, carrying little meaning of its own. Conversely, in "Kim took a painting," take retains its full semantic weight. Accurately identifying these support verbs is critical for machine translation, information retrieval, and automated style checking.

Previous approaches to SV identification have typically relied on complex manual semantic rules or local frequency counts. However, local methods often suffer from data sparsity, failing to distinguish between a verb that is genuinely "light" and one that simply appears frequently with a specific noun by chance.

The Demographic Analogy

The core innovation of this work is the application of demographic mortality models to linguistic data. In demography, a "Crude Death Rate" can be misleading if population age structures differ. To account for this, demographers use an Indirect Standardised Death Rate (ISDR), which compares observed deaths in a target population to expected deaths based on a larger standard population.

The authors map these demographic concepts to the problem of identifying support verbs, establishing a framework where verb productivity across a corpus serves as a proxy for its semantic "lightness."

The Global Information Model

The proposed model moves beyond local verb-noun frequencies by incorporating "global productivity." The process involves three main components:

  1. Local Model: Measures how frequently a verb governs a specific noun.
  2. Global Model: Calculates a "global lightness" probability ($p_i$) for every verb across the entire corpus, representing its general tendency to act as a support verb.
  3. Integration: The most likely support verb for a given noun is determined by multiplying the local frequency by the verb's global productivity.

This weighting allows the model to recognize that highly productive verbs like make or have are likely support verbs even when their local frequency with a specific noun is relatively low. This approach effectively "smooths" the data, making the model robust against the "long tail" of rare noun occurrences.

Experimental Results

The model was trained on the 1992 Grolier’s Encyclopedia (approximately 8 million words) and tested against 18 verb-noun constructions drawn from linguistic literature.

As shown in the results table, the Global Information Model correctly identified the support verb in 13 out of 18 cases. When excluding informal phrases absent from the encyclopedic training data (such as "have a snooze"), the accuracy rose to 14 out of 15 (93%). This performance significantly exceeded contemporary state-of-the-art methods, demonstrating that a sophisticated statistical model can compensate for smaller training corpora.

Conclusion

By leveraging the concept of global productivity through a demographic lens, the authors provide a robust method for identifying support verbs. While the model is sensitive to the genre of the training corpus and currently does not distinguish between abstract and concrete noun usages, its success in handling sparse data marks a significant advancement in statistical linguistics. Future work may involve iterative refinement of global lightness scores to further enhance accuracy.

Summary updated

Abstract

Some verbs have a particular kind of binary ambiguity: they can carry their normal, full meaning, or they can be merely acting as a prop for the nominal object. It has been suggested that there is a detectable pattern in the relationship between a verb acting as a prop (a \term{support verb}) and the noun it supports. The task this paper undertakes is to develop a model which identifies the support verb for a particular noun, and by extension, when nouns are enumerated, a model which disambiguates a verb with respect to its support status. The paper sets up a basic model as a standard for comparison; it then proposes a more complex model, and gives some results to support the model's validity, comparing it with other similar approaches.
BibTeX
@misc{dras1996death,
  author = {Mark Dras and Mike Johnson},
  title = {{Death and Lightness: Using a Demographic Model to Find Support Verbs}},
  year = {1996},
  eprint = {cmp-lg/9610001},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/cmp-lg/9610001},
}
Figure