Verb subcategorization frequencies: American English corpus data, methodological studies, and cross-corpus comparisons

Susanne Gahl, Dan Jurafsky, Douglas RolandBehavior 2004 · DOI

Introduction

In psycholinguistics, "verb bias"—the statistical preference of a verb for a particular syntactic structure—is a critical predictor of sentence processing difficulty. For instance, a verb like remember can precede a direct object (DO) or a clausal complement (SC). Knowing which structure a verb "prefers" helps researchers predict reading times and processing difficulty. However, existing frequency resources often suffer from small sample sizes or inconsistent coding criteria. This paper presents a high-quality dataset of subcategorization frequencies for 281 English verbs based on manual analysis of large-scale American English corpora, while investigating how methodological choices impact these classifications.

Methodology

The authors sampled 200 sentences per verb from the TASA (educational materials) and Brown (balanced American English) corpora. Unlike automated parsers, which can be unreliable, linguistics graduate students were employed to manually label each instance using an 18-category coding scheme, achieving high inter-rater reliability (kappa = .84). The study specifically examined five methodological variations to determine their effect on "bias" labels (DO-biased, SC-biased, or equi-biased).

Impact of Coding Decisions

A primary finding is that coding decisions significantly alter results. The treatment of passive sentences is particularly impactful; when passives are excluded from transitive counts, a significant portion of verbs change their bias classification. Similarly, including adjectival passives (common in psychological state verbs like amuse or delight) shifts the bias for 25% of applicable verbs. The study also highlights the difference between absolute measures (e.g., is the verb transitive >66% of the time?) and relative measures (e.g., is it twice as likely to be a DO than an SC?), finding that relative measures better capture the nuances of human sentence processing.

Effect of transitivity definitions (including vs. excluding passives) on verb bias classification.

Cross-Corpus Comparison

The study compared its findings against 10 major resources, including corpus-based datasets (e.g., COMLEX, Wall Street Journal) and elicitation-based norming studies. Using relative bias measures, the study found a 97% agreement rate with other sources. This high level of consistency suggests that while genre sensitivity exists—for example, guess acts differently in lab studies versus natural speech—the underlying structural preferences remain relatively stable across American English registers.

Comparison of verb bias classifications across multiple corpus and elicitation studies.

Conclusion

This research provides a standardized, high-quality benchmark for the psycholinguistic community through the release of the Gahl2004norms.txt dataset. While limited by sample size for rare frames and the specific register of American English, the study demonstrates that rigorous manual coding and careful methodological definitions are essential for understanding the statistical properties of the mental lexicon.

Summary updated

BibTeX
@inproceedings{gahl2004verb,
  author = {Susanne Gahl and Dan Jurafsky and Douglas Roland},
  title = {{Verb subcategorization frequencies: American English corpus data, methodological studies, and cross-corpus comparisons}},
  booktitle = {Behavior},
  year = {2004},
  doi = {10.3758/bf03195591},
}
Figure