The use of focus as a tool for disambiguation of definite noun phrases

Candace L. Sidner1978 · DOI

This classic 1979 paper from the MIT Artificial Intelligence Laboratory, authored by Candace L. Sidner, tackles one of the thorniest problems in natural language processing (NLP): how to program a computer to understand what "the" refers to. By introducing the concept of focus as a computational tool, Sidner provides a framework for disambiguating definite noun phrases (defnps) that moves beyond simple keyword matching or exhaustive inference.

TL;DR

Sidner proposes that "focus"—the current center of attention in a discourse—acts as an index function to resolve ambiguous definite noun phrases. By combining focus with hierarchical semantic networks and a set of co-reference rules, the paper offers a structured way for machines to identify whether a phrase refers to a previously mentioned entity, a generic class, or an associated concept.

What Problem Does This Paper Address?

In human conversation, we use definite noun phrases (like "the car," "the meeting," or "the murderer") constantly. However, for a computer, these are minefields of ambiguity. Does "the car" refer to a vehicle mentioned two sentences ago? Is it a generic statement about all cars? Or is it a car that was never mentioned but is implied because the speaker just talked about a "garage"?

Sidner identifies six open problems in computational linguistics, ranging from defining what "reference" actually means to determining what data structures are needed to detect co-reference (when two different phrases refer to the same thing). The core challenge is that purely inference-driven models (popular at the time) were computationally expensive and lacked the structural constraints provided by the flow of actual conversation.

Methods

Sidner’s approach relies on three primary components:

  1. Focus as an Index: The "focus" is the entity the speaker is currently talking about. It serves as the primary starting point for searching for a referent.
  2. Hierarchical Semantic Networks: The system uses a network of associations to understand relationships between concepts. For example, the concept of a "meeting" is linked to "place," "time," and "participants." Hierarchical net for a 'meeting' concept
  3. The Inferencing Mechanism: This is used to validate potential matches. If the system thinks "it" refers to "the meeting," the inference engine checks if a meeting is the kind of thing that can "occur at 3 p.m."
Hierarchical net for a 'meeting' concept

The paper introduces the Explicit Backwards Co-referencing (EBC) Rule. In its simplest form, if a defnp has the same head noun as the current focus and doesn't provide conflicting new information, the system assumes they co-refer.

Datasets

Not discussed in the paper. As a foundational theoretical work in computational linguistics, the paper uses constructed discourse examples (labeled D1 through D18) to demonstrate how the proposed rules handle various linguistic scenarios rather than testing against a large-scale corpus.

Key Insights and Ideas

  • Internal vs. Co-reference: Sidner distinguishes between an "internal reference" (a link between a noun phrase and a database object representing a real-world entity) and "co-reference" (a link between two noun phrases). Reference links between noun phrases
  • The Perspicuity Maxim: Borrowing from philosopher H.P. Grice, Sidner assumes speakers are "perspicuous"—they aren't trying to be confusing. This allows the system to favor the most "focused" interpretation rather than searching a global database of every object it knows.
  • Implicit and Inferred Reference: The paper explains how we resolve phrases for things not previously mentioned. In "I went to a restaurant. The waitress was nasty," the "waitress" is an implicit referent found via the semantic network's association with "restaurant." In "The heiress died. The murderer was never found," the "murderer" is an inferred referent based on the context of a mysterious death.
  • Focus is Dynamic: The focus can shift. The paper notes that detecting these shifts (or "sub-topic shifts") is crucial for maintaining an accurate index of referents as a conversation progresses.
Reference links between noun phrases

Key Results

The paper provides a formal set of rules that successfully categorize and resolve different types of definite noun phrases:

  • Specific: Refers to a unique, identifiable entity.
  • Generic: Refers to a whole class (e.g., "The elephant is a large mammal").
  • Attributive: Describes an entity without necessarily pointing to a specific individual (e.g., "Smith's murderer," whoever that may be).

By using focus, the system can disambiguate phrases that are identical on the surface but different in context. For instance, it explains why "The elephant" in one context refers to a specific pet, while in another, it refers to the entire species, based on whether the "focus" of the preceding sentence was an individual or a category.

Limitations and Future Work

Sidner acknowledges several areas for further development:

  • Focus Shift Detection: While the paper uses focus, it refers to other works (like Bullwinkle 1977) for the complex mechanics of how the focus actually moves from one entity to another during a long discourse.
  • Hearer Beliefs: The model assumes a relatively simple shared knowledge base. It does not fully account for situations where the hearer’s beliefs differ from the speaker’s, which can lead to "incomplete" or "incorrect" comprehension.
  • Modal Contexts: Phrases inside "expected," "wanted," or "believed" (e.g., "Mary expected a present") create complex semantic layers that the EBC rules struggle to resolve without more sophisticated modal logic.

Summary updated

BibTeX
@inproceedings{sidner1978use,
  author = {Candace L. Sidner},
  title = {{The use of focus as a tool for disambiguation of definite noun phrases}},
  booktitle = {Proceedings of the 1978 workshop on Theoretical issues in natural language processing -},
  year = {1978},
  doi = {10.3115/980262.980276},
}
Figure