A Compositional Semantics for Focusing Subjuncts

Daniel Lyons, Graeme HirstACL 1990 · DOI

Introduction

Focusing subjuncts, such as only, even, and also, present a significant challenge for natural language processing due to their "promiscuous" syntactic and semantic behavior. These words can appear in various positions within a sentence and can focus on almost any constituent, with the resulting meaning heavily dependent on intonational stress. For instance, the sentence "John only introduced Bill to SUE" carries a different truth condition than "John only introduced BILL to Sue." Traditional slot-filler semantic models often struggle with these words because they do not simply fill a slot but rather transform the entire proposition's meaning. This paper presents a computationally efficient, frame-based semantic system that incorporates intonation to resolve these ambiguities.

Computational Framework

The proposed system builds upon Rooth’s (1985) domain selection theory, which utilizes higher-order logic to define sets of alternative propositions (p-sets). To adapt this for practical AI applications, the authors translate these concepts into a frame-based semantic formalism based on Hirst’s Absity system.

A central component of this framework is the Focus Feature, a binary attribute (+/- focus) assigned to stressed words. This feature "percolates" up the syntax tree; the level at which it stops determines the scope of the focus. The mapping between syntactic structures and their semantic counterparts is defined through compositional rules.

Table of syntax and semantic rules for sentence (S), verb phrase (VP), and prepositional phrase (PP) constituents.

As shown in the table above, each syntactic rule is paired with a semantic rule that constructs a frame-based representation. This allows the system to maintain a compositional approach while handling the complexities of focus.

Semantic Composition and Pragmatics

The system employs a two-part semantic representation for every constituent: one for the asserted meaning and another for the non-asserted (presupposed) meaning. This distinction is crucial for words like even, which carry pragmatic implications (e.g., unlikelihood) that are separate from the core truth-conditional assertion.

Rather than applying complex transformations at every node, the system allows a latent operator (e.g., only) to percolate to the sentence level. At this point, the operator acts upon the fully developed proposition and its associated p-set. This separation of concerns simplifies the parsing process while maintaining semantic richness.

Implementation: The IDEO System

The authors implemented these theories in a system called IDEO (Interpreter Designed for Even and Only), written in C-Prolog. IDEO interacts with a toy knowledge base to demonstrate the validity of the rules. The system can:

  1. Process sentences with intonational markers.
  2. Generate multiple semantic interpretations based on focus percolation.
  3. Update the knowledge base with "frame implications"—logical constraints derived from the semantics of focusing subjuncts.

Conclusion and Limitations

While the IDEO system successfully demonstrates a frame-based approach to focusing subjuncts, it has limitations. The implementation was restricted to a small, hand-crafted environment and typically handled only one subjunct per sentence. Furthermore, the system uses a simplified model of presupposition. Despite these constraints, the work provides a vital bridge between abstract linguistic theory and computational semantics by treating intonation as a primary data source for interpretation.

Summary updated

BibTeX
@inproceedings{lyons1990compositional,
  author = {Daniel Lyons and Graeme Hirst},
  title = {{A Compositional Semantics for Focusing Subjuncts}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {1990},
  doi = {10.3115/981823.981831},
}
Figure