Semantic underspecificationoccurs when linguistic expressions carry partial meaning, requiring context for full understanding. It poses key challenges across philosophy, cognitive science, and NLP. This review identifies five developmental stages: (1)Classical theories by Frege, Russell, and Davidson created truth-conditional frameworks but encountered difficulties with indexicaland belief contexts due to assumptions of full specification; (2)Formal models like QLF, MRS, and Hole Semantics introduced computational underspecification to handle structural ambiguities such as quantifier scope; (3)Cognitive studies show humans use underspecification strategically for efficiency, relying on pragmatic inference and semantic memory; (4)Hybrid neuro-symbolic models like UMR and Glue Semantics combined structural ambiguity resolution with neural inference but lacked uncertainty modeling; (5)Modern NLP research highlights gaps: LLMs can detect underspecification but often overcommit to deterministic interpretations, and multimodal systems do not effectively utilize context. Cross-linguistic entropy models suggest grammatical underspecification as a strategy for cognitive efficiency. To bridge human semantic flexibilityenabled by incremental processing and pragmatic co-construction—with computational systems, we propose integrated neuro-symbolic architectures incorporating explicit uncertainty modeling, multimodal grounding, and entropy-aware design. This approach paves the way for AI to achieve human-like language understanding.
Qayyum et al. (Sun,) studied this question.