Anaphora for Limited Domain Systems

Philip J. Hayes

Philip J. Hayes (1981)

Introduction

In the early 1980s, Human-Computer Interaction (HCI) faced a significant hurdle: natural language processing. Specifically, the resolution of anaphora—the use of pronouns like "it" or "them" to refer back to previously mentioned entities—presented a computational challenge that seemed to require vast amounts of real-world knowledge. This paper from Carnegie Mellon University proposes a pragmatic alternative: a rule-based mechanism designed for limited-domain systems that prioritizes predictability over cognitive simulation.

The Problem: Context and Knowledge

Human-like anaphora resolution is notoriously difficult because it often relies on social context, intent, and subtle linguistic cues. For instance, in the command "Display the messages from Campbell, then delete them," a system must correctly identify that "them" refers to the specific messages just displayed. Hayes argues that for restricted tasks like database management or email systems, a full simulation of human cognition is unnecessary. Instead, a system should provide "equivalent functionality" through a set of clear, learnable rules.

Methodology: Context Structures and Selection Records

The proposed algorithm manages three distinct context structures: CURRENT (the most recent exchange), NEW (the context being built), and BACKUP (older items). To track these entities, the system utilizes specific properties to define referents and their relationships.

Definitions of properties used in anaphoric descriptions

As shown in Figure 1, the system tracks domain types, grammatical number, and whether a description is a pronoun. It also handles "set selection" logic, which allows the system to manage subsets of data. For example, if a user requests "the first two" items from a list, the system maintains a record of the original superset, the current subset, and the cumulative subset of items already addressed.

Heuristics for Resolution

A key innovation is the "Noun vs. Pronoun" heuristic:

  1. Noun-headed phrases (e.g., "the last message") trigger a global search of the database for the absolute last entry.
  2. Pronoun-headed phrases (e.g., "the last one") are treated as anaphoric, triggering a local search within the immediate context of what was just discussed.

Furthermore, the system employs selectional restrictions. If a user says "Delete it," the system filters for entities that are "deletable" (e.g., messages) and ignores incompatible types (e.g., dates or people), even if they were mentioned more recently.

Illustrative Examples

The paper demonstrates these concepts using a hypothetical electronic mail system. The system operates on a database of messages, such as those shown in the following example list:

Sample message database used for illustrative examples

Using this data, the algorithm can resolve complex follow-ups like "the others" by calculating the complement of the cumulative subset within the current superset. This logic ensures that "the others" correctly identifies remaining items without the list of "already selected" items growing indefinitely.

Conclusion and Limitations

The strength of this approach lies in its predictability. By following strict rules rather than guessing at human intent, the system allows users to adapt to its "mental model." However, limitations remain, particularly with nested selections (subsets of subsets) and instances where strict logic clashes with human pragmatics—such as whether a "deleted" item should still be considered part of "the others." Hayes concludes that this functional approach to linguistic problems like ellipsis and ungrammatical input is essential for creating "habitable" natural language interfaces.

Summary updated

BibTeX
@inproceedings{hayes1981anaphora,
  author = {Philip J. Hayes},
  title = {{Anaphora for Limited Domain Systems}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1981},
}
Figure