Lemmaval

Lemmaval, short for lemma value, is a scalar metric used in computational linguistics to quantify a lemma's usefulness for natural language processing tasks. The metric is designed to capture how representative a lemma is of its word family, how stable it is across inflected forms, and how well it covers its semantic senses within a corpus.

Computationally, lemmaval is produced by combining multiple signals. Typical components include: normalized lemma frequency across a reference corpus; a sense-disambiguation reliability score indicating how confidently occurrences can be assigned to senses; and a measure of lexical diversity or polysemy (the number of distinct senses). The resulting score is scaled from 0 to 1 or 0 to 100, and weights can be adjusted to fit language, domain, or task.

Uses and limitations: Lemmaval is used to prioritize lemmas when constructing lexicons, guiding annotation budgets, selecting features for machine learning models, and informing dictionary or database indexing. It helps identify high-value lemmas for transfer learning, lexicon consolidation, or domain adaptation. However, the score depends on the corpora, sense inventories, and annotation quality deployed, so cross-domain comparisons require careful normalization, and the measure may undervalue domain-specific terms with low general frequency.

Origin: The term is a neologism in computational linguistics, reflecting a practical measurement approach rather than a universally standardized concept. Its exact definition and calculation can vary by study.