The Use of Dynamically Extracted Context for Anaphoric Reference Resolution

David Klappholz, Abe Lockman

Abstract

This paper, originally published in 1977, challenges the traditional, narrow view of anaphora resolution in Natural Language Processing (NLP). Rather than treating anaphora as a simple mapping between pronouns and nouns, the authors propose a framework where resolution is driven by a dynamic "context graph." By modeling discourse through sentence-level relationships—specifically expansion and temporal continuation—the authors suggest a "guess and test" algorithm that utilizes discourse structure to resolve complex references that static knowledge frames cannot handle in isolation.

Introduction

In the late 1970s, the field of Natural Language Processing was primarily concerned with how machines could parse syntax and understand isolated semantic units. A significant hurdle was context: how could a system understand what a word referred to based on previous sentences? At the time, most systems focused on simple pronominal anaphora (e.g., linking "he" to "John").

Klappholz and Lockman argued that this approach was insufficient. They posited that anaphora is a much broader phenomenon. For instance, in the sequence "I went to the circus. I saw a tattooed lady," the phrase "the tattooed lady" functions anaphorically because her existence is implied by the "circus" context. To resolve such references, a system cannot rely solely on static look-up tables; it must understand the dynamic flow of the discourse.

The Problem with Static Knowledge Structures

Existing theories of the era, such as Marvin Minsky’s "frames" or Roger Schank’s "scripts," attempted to provide machines with pre-packaged clusters of world knowledge. While useful, Klappholz and Lockman argued these structures were too rigid. They identified a gap in how these frames were applied: there was no clear mechanism to determine which frame was relevant or how a new sentence modified the current context.

Because the "flow of context" is dynamic, the authors argued that text structure itself—not just world knowledge—must guide the resolution process. They suggested that a system should actively hypothesize how each new sentence fits into the preceding discourse rather than simply searching for a matching antecedent in a static database.

The Context Graph and "Guess and Test" Algorithm

The core of the proposed method is a theoretical framework known as the Context Graph. This approach relies on several key components:

1. Broadened Definition of Anaphora

The authors include pronouns, pro-verbs, definite/indefinite references, and even entire sentences as potential anaphoric elements. For example, a sentence describing a car failing to start can be seen as an anaphoric reference to the "cause" of a character being late mentioned in a prior sentence.

2. Categorization of Sentence Relations

The framework organizes the relationship between sentences into two primary categories:

  • Expansion: The current sentence provides further detail, cause, effect, or an explanation of "how" regarding a previous statement.
  • Temporal Continuation: The current sentence moves the narrative timeline forward.

3. Graph Construction and Resolution

As a text is processed, the system constructs a graph where sentences serve as nodes and the relations (expansion or continuation) serve as edges. When a new sentence is encountered, the resolver attempts to "fit" it into the graph by searching upward through previous nodes.

This follows a "top-down" validation logic: the system first hypothesizes a sentence-level relationship (e.g., "Sentence B expands Sentence A"). If this relationship is validated through semantics and world knowledge, the specific item-to-item references within those sentences are resolved as a natural byproduct of that structural link.

Key Insights

  • The Default Assumption: The authors suggest that readers (and thus resolvers) naturally assume a sentence expands or continues the immediately preceding sentence. A resolver should only look further back in the discourse if the link to the most recent node is found to be weak.
  • Structure Over Semantics: While world knowledge is necessary to validate a link, the structure of the text—how the writer groups ideas into sentences—provides the primary clues for resolution.
  • Dynamic Frame Management: Rather than frames "solving" anaphora, the authors view frames as tools that must be dynamically shifted in and out of focus based on the evolving structure of the context graph.

Limitations and Conclusion

As a theoretical and programmatic paper from 1977, the work lacks empirical testing on large-scale datasets, which were not yet standard in the field. The authors used "toy" examples to illustrate their logic and referred to technical reports for algorithmic sketches.

The primary limitations were computational. The "guess and test" method requires "directed inferencing" using vast amounts of world knowledge—a requirement that far exceeded the computational power and digital knowledge bases available at the time. Furthermore, the broad categories of "expansion" and "continuation" left many nuances of meaning unresolved.

Despite these limitations, the paper was forward-thinking in its insistence that anaphora resolution is a search problem within a structural graph rather than a simple table-lookup. It shifted the focus from isolated words to the holistic relationship between sentences, prefiguring modern discourse analysis in computational linguistics.

Summary updated

BibTeX
@inproceedings{klappholz1977use,
  author = {David Klappholz and Abe Lockman},
  title = {{The Use of Dynamically Extracted Context for Anaphoric Reference Resolution}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1977},
}
Figure