lextrapolation

Lextrapolation is a term used in linguistics and computational linguistics to describe the process of extending a lexicon by predicting new lexical items, senses, or forms from existing patterns. It involves extrapolating beyond attested data to generate plausible additions while respecting the morphology, phonology, and semantics of the language.

Approaches to lextrapolation can be grouped into rule-based, analogy-based, and data-driven methods. Rule-based extrapolation applies productive morphology and phonotactic constraints to create new forms from existing words, such as attaching common affixes or forming plausible compounds. Analogy-based methods exploit systematic relationships among word forms or senses, leveraging patterns observed within a language or across related languages. Data-driven methods use large corpora or neural language models to propose candidates that are plausible given context, followed by filtering with language-model scores, semantic similarity measures, or human judgment. Hybrid approaches combine these elements, using a seed lexicon to generate candidates and then validating them computationally or with lexicographic expertise.

Applications of lextrapolation include expanding lexicons for under-resourced languages, generating neologism candidates for dictionaries, supporting historical linguistics research on word formation, and improving language technologies by augmenting training data with synthetic but plausible lexemes.

Challenges include judging actual usage and acceptance, avoiding creation of non-words or misleading coinages, managing semantic drift, and mitigating biases present in training data. Evaluation often relies on expert assessment, corpus validation, or user studies to determine whether generated items are credible and useful within a given linguistic community.

Example: from the verb cloud, a plausible extrapolated form is cloudify, a verb meaning to make something cloud-based, illustrating how productive morphology and semantic drift can guide lextrapolation.