Focusing and Reference Resolution in PUNDIT

Deborah A. DahlAAAI 1986

Abstract

This paper, originally presented at AAAI-86, introduces the reference resolution component of PUNDIT (Prolog UNDerstander of Integrated Text). While many natural language processing systems of the era treated "focusing"—the tracking of the central entity in a discourse—primarily as a tool for resolving pronouns, this research argues for a more versatile mechanism. By shifting from a semantic "thematic role" approach to a "surface syntactic" approach, PUNDIT achieves more robust handling of complex linguistic phenomena like elided phrases and "one-anaphora" across diverse technical domains.

Introduction

In the mid-1980s, a major challenge in Computational Linguistics was reference resolution: the process by which a computer identifies the referents of terms like "it," "the motor," or "the blue one" in a stream of text. Existing models, most notably Sidner’s (1979), relied heavily on thematic roles—identifying the "Agent," "Theme," or "Instrument" of a verb—to rank potential candidates for the focus of a sentence.

However, this approach faced two significant hurdles. First, thematic roles are difficult to generalize; every verb requires specific mapping, making systems hard to port to new domains. Second, evidence suggested that surface syntax (word order and grammatical position) often exerts a stronger influence on discourse focus than underlying semantic meaning.

The PUNDIT System Architecture

Implemented in Quintus Prolog, PUNDIT processes text through a multi-stage pipeline:

  1. Restriction Grammar: A syntactic component that performs a top-down parse.
  2. Semantic Interpreter: Based on Palmer’s "Inference Driven Semantic Analysis," it decomposes verbs and fills thematic roles.
  3. Discourse Component: The site of the novel focusing algorithm.

The Syntactic Focusing Algorithm

The core innovation of PUNDIT is its departure from thematic-role-based focusing. Instead, the system ranks potential foci based on their syntactic position. This shift is illustrated by comparing PUNDIT's hierarchy to the then-standard model proposed by Sidner.

Comparison of focusing hierarchies: Sidner's thematic roles vs. PUNDIT's syntactic positions.

The system maintains a FocusList, a ranked list of entities mentioned in the discourse. When encountering a pronoun or a "zero-NP" (an elided phrase common in technical manuals), PUNDIT searches the top of the FocusList to identify the most likely referent. By prioritizing positions like the Direct Object and Subject over semantic roles like "Theme," the system better captures how speakers naturally shift focus through word order.

Handling Complex Linguistic Phenomena

PUNDIT's syntactic approach allows it to address several challenging discourse problems:

  • Syntax over Semantics: Dahl demonstrates that surface word order is a superior predictor of focus. For instance, in the active sentence "The bookcases are blocking the outlets," the focus is on the outlets. In the passive "The outlets are blocked by the bookcases," the focus shifts to the bookcases, despite the "Theme" (the outlets) remaining constant.
  • One-Anaphora: The paper treats "one-anaphora" (e.g., "I have two disk drives; one failed") as a discourse-level focusing problem. PUNDIT resolves "one" by identifying the most recent set of entities in the FocusList.
  • Zero-NPs (Elision): In technical sublanguages, subjects are frequently omitted (e.g., "[The unit] Won't power up"). PUNDIT treats these "zeroes" as implicit pronouns referring to the current discourse focus.
  • Implicit Associates: The system can resolve references like "the motor" even if only "the disk drive" was previously mentioned, provided a "part-whole" relationship exists in the knowledge base.

Evaluation and Portability

The system was primarily developed using maintenance reports for computer equipment failures and subsequently ported to a domain involving air compressor failures. Because PUNDIT’s focusing rules are based on general English syntax (Subject, Object) rather than verb-specific semantics, it demonstrated high domain portability. The syntactic algorithm correctly predicted focus in cases where thematic-role-based models failed, particularly during shifts between active and passive voice.

Conclusion and Future Work

While PUNDIT demonstrates that surface syntax is a powerful tool for reference resolution, the author acknowledges that focusing provides candidates that must still be validated by world knowledge. Future improvements suggested include more sophisticated treatment of reflexive pronouns based on Binding Theory and the inclusion of "implicit sets" (sets inferred from context) within the FocusList.

Summary updated

BibTeX
@inproceedings{dahl1986focusing,
  author = {Deborah A. Dahl},
  title = {{Focusing and Reference Resolution in PUNDIT}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1986},
}
Figure