underspecification

Underspecification is a situation in which a specification, model, or theory does not determine a unique interpretation, outcome, or value given the available information. It can arise in language, formal systems, and computational models, and it highlights the gap between constraints provided by data or rules and the range of possible consequences that satisfy them.

In linguistics and semantics, underspecification describes an intentional leaving of certain semantic or pragmatic elements underspecified in the explicit representation. This can include referential identity, scope of quantifiers, tense or aspect, or attachment of syntactic constituents. The language user or processing system resolves underspecification using context, discourse structure, salience, or pragmatic inference, producing different interpretations as needed.

In logic, philosophy of science, and formal semantics, underspecification occurs when a theory does not fix all truth conditions or when a theory allows multiple compatible models. This is related to the problem of underdetermination, where data do not determine a unique theory, and to model- or process-agnostic representations that admit several compatible assignments.

In computer science and machine learning, underspecification refers to underconstrained problems or models where multiple hypotheses fit the data or satisfy the specification. This includes algorithm choice, parameter values, or feature representations that all yield similar performance. Consequences include fragility under distribution shift and difficulties in reproducibility and causal inference.

Approaches to address underspecification include making choices explicit, collecting additional evidence, using priors or invariants, designing experiments to discriminate among alternatives, and evaluating robustness across plausible interpretations.