contrastwhat

Contrastwhat is a term used in theoretical discussions of contrastive analysis to denote a systematic approach for identifying and interpreting salient contrasts between categories in data. The term blends contrast with what, emphasizing inquiry into what distinguishes items within a domain. There is no single canonical definition, and the term is used across disciplines to signal a general methodological idea rather than a fixed technique.

In practice, contrastwhat refers to a framework for discovering and evaluating contrastive features—features that separate, or diagnose, differences between groups. It can apply to phonology, semantics, visual data, or behavioral datasets. A typical workflow involves selecting a domain, collecting representative data for the groups, extracting candidate features, calculating a contrast score for each feature, and interpreting high-scoring features as salient contrasts. Common scoring methods include mutual information, chi-squared, information gain, or distributional delta.

Applications include linguistics, where researchers identify phonemic or morphosyntactic contrasts across dialects or languages, and computer science or data science, where the approach guides feature selection and model interpretability by highlighting what features most distinguish classes.

Examples might include distinguishing two language variants by pronunciation features, or separating object categories in images by color and texture features that show the greatest class separation.

Limitations include its informal status, varying definitions, and the risk of overemphasizing single features without context. Proponents emphasize that contrastwhat can complement other analyses by focusing attention on what matters for differentiation.

See also: contrastive analysis, feature selection, information theory, discriminant analysis.