Count regression models for keyness analysis

Abstract

A wide variety of measures have been used in previous work to assess the keyness of items in a particular domain of language use. The present paper explores an approach to keyword analysis based on regression modeling. Specifically, we use a form of negative binomial regression, which offers a number of advantages compared to existing techniques for identifying typical items in a target corpus. Thus, it is responsive to the multidimensional nature of keyness and can address multiple aspects of typicalness simultaneously, using a single statistical model. Further, metrics of interest can be enriched with confidence intervals, which allows us to isolate descriptive and inferential indicators of keyness. Finally, all quantities are based on a text-level analysis, which accounts for the fact that the target and reference corpus consist of text files and adjusts uncertainty estimates accordingly. As an illustrative case study, we rely on COCA to identify key verbs in academic writing and demonstrate how negative binomial regression may be used to this end. Our checks on the coverage rate of the 95% confidence intervals indicate that this model seems to be adequate for purposes of statistical inference. Due consideration will also be given to the limitations of this procedure, and we conclude by outlining the kinds of keyness analyses for which count regression models may be a worthwhile approach. The online supplementary material for this paper provides data and R code for the implementation of keyness regression.

Cite this work

No public copy is stored for this record, so only the metadata is cited.

Stored only when both papers are in the corpus — 1 citations exist upstream.

Papers that cite the same references — the relational fallback while embeddings are still backfilling.