Disambiguate First Parse Later: Generating Interpretations for Ambiguity Resolution in Semantic Parsing

  • 2025-02-25 18:42:26
  • Irina Saparina, Mirella Lapata
  • 0

Abstract

Handling ambiguity and underspecification is an important challenge innatural language interfaces, particularly for tasks like text-to-SQL semanticparsing. We propose a modular approach that resolves ambiguity using naturallanguage interpretations before mapping these to logical forms (e.g., SQLqueries). Although LLMs excel at parsing unambiguous utterances, they showstrong biases for ambiguous ones, typically predicting only preferredinterpretations. We constructively exploit this bias to generate an initial setof preferred disambiguations and then apply a specialized infilling model toidentify and generate missing interpretations. To train the infilling model, weintroduce an annotation method that uses SQL execution to validate differentmeanings. Our approach improves interpretation coverage and generalizes acrossdatasets with different annotation styles, database structures, and ambiguitytypes.

 

Quick Read (beta)

loading the full paper ...