Automatic Identification of Support Verbs: A Step Towards a Definition of Semantic Weight

Mark Dras

Abstract This paper explores a foundational challenge in computational linguistics: the mathematical definition of a word’s "semantic weight." While it is intuitively understood that verbs like make in "make a decision" carry less semantic content than the verb decide, formalizing this "lightness" is essential for improving readability metrics and information retrieval. We propose a computationally tractable definition of semantic weight based on a verb's productivity in "light" syntactic constructions. By combining local co-occurrence data with a global measure of productivity, we demonstrate an automated method for identifying support verbs across various nominalizations.

Introduction

Traditional readability measures often rely on surface-level metrics like word length or syllable counts. However, these fail to account for lexical density—the distinction between words that carry significant propositional content and those that primarily serve grammatical functions. While linguists have long discussed "content" versus "non-content" words, the boundary remains fluid. Verbs such as make or take are traditionally lexical items, yet they function as support verbs in constructions like make an attempt. In these contexts, they become semantically "light." This research addresses the challenge of creating an automated, statistical method to identify these light instances without relying on static dictionaries.

Methodology

Building on previous work in lexical statistics, this approach introduces a "global" weighting component to identify support verbs. The method utilizes two primary types of information:

  1. Local Information: This measures the frequency of specific verb-object pairs (e.g., the co-occurrence of make and proposal).
  2. Global Information: This measures the "productivity" of a verb—how many different nominalizations it supports across a corpus. A verb like make is highly productive, supporting a wide array of nouns (decision, attempt, statement), whereas a "heavy" verb like bear is restricted to fewer constructions (resemblance, witness).

Unlike standard Information Retrieval techniques like TF-IDF that downweight common words, our algorithm upweights verbs with high global frequency in support-verb constructions. Candidates are ranked by calculating the product of their local and global relative frequencies.

Results and Evaluation

The system was evaluated using the 1992 Grolier’s Encyclopedia (approx. 8 million words) and a lexicon of 1,671 deverbal nominalizations derived from the Longman’s Dictionary of Contemporary English (LDOCE). The system's performance was tested against 18 established light verb constructions.

Table 1: Comparison of system choices against reference support verbs across various nominalizations.

As shown in Table 1, the system's first choice matched the source text in 13 out of 18 cases (72%). A significant factor in the failures was data scarcity related to the corpus register; informal constructions like have a snooze or give a shove did not appear in the formal encyclopedia text. When excluding cases with insufficient data and allowing for valid alternative support verbs (e.g., cause harm vs. do harm), the success rate rises to 93% (14 out of 15).

Key Insights

  • Productivity as a Proxy for Lightness: The core finding is that as a verb's meaning becomes "bleached," it becomes more versatile and productive in supporting various nouns.
  • Contextual Lightness: High frequency alone does not indicate semantic lightness. For instance, include is more frequent than make in the corpus but remains a "heavy" content verb. Lightness is specifically tied to behavior within Verb + NP complement constructions.
  • The Concretization Problem: Nominalizations can sometimes refer to physical objects (e.g., a physical "proposal" document). Global productivity weighting helps the system filter out these "heavy" usages by focusing on the verb's broader functional patterns.

Conclusion

This research demonstrates that support verbs can be identified with high accuracy through statistical measures of productivity. While the current results are influenced by the register of the source corpus, the methodology provides a robust framework for defining semantic weight. Future work will extend this analysis to light verbs with adjectival complements (e.g., get ready) to further refine measures of lexical density for stylistic analysis.

Summary updated

Abstract

Current measures of the readability of texts are very simplistic, typically based on counts of words or syllables per sentence. A more sophisticated analysis needs to take account of the fact that the particular distributions of meanings across wordings chosen by the writer, and the consequent variations in syntactic structure, have a significant effect on readability. A step towards the required sophistication is provided by the notion of lexical density (Halliday, 1985), which suggests that different words carry different amounts of semantic weight; this idea of semantic weight is also used implicitly in areas such as information retrieval and authorship attribution. Current definitions of these notions of lexical density and semantic weight are based on the division of words into closed and open classes, and on intuition. This paper develops a computationally tractable definition of semantic weight, concentrating on what it means for a word to be semantically light; the definition involves looking at the frequency of a word in particular syntactic constructions which are indicative of lightness. Verbs such as make and take, when they function as support verbs, are often considered to be semantically light. To test our definition, we carried out an experiment based on that of Grefenstette and Teufel (1995), where we automatically identify light instances of these words in a corpus; this was done by incorporating our frequency-related definition of semantic weight into a statistical approach similar to that of Grefenstette and Teufel. The results show that this is a plausible definition of semantic lightness for verbs, which can possibly be extended to defining semantic lightness for other classes of words.
BibTeX
@misc{dras1995automatic,
  author = {Mark Dras},
  title = {{Automatic Identification of Support Verbs: A Step Towards a Definition of Semantic Weight}},
  year = {1995},
  eprint = {cmp-lg/9510007},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/cmp-lg/9510007},
}
Figure