The Use of Dynamically Extracted Context for Anaphoric Reference Resolution

TL;DR

This paper proposes a novel approach to anaphoric reference resolution by dynamically extracting local context based on the structural relationships between sentences in a text. Instead of relying on pre-defined frames or scripts, it models the text as a graph of expansion/temporal continuation relationships, allowing for more robust resolution of ambiguous references.

What Problem Does This Paper Address?

Resolving anaphora – understanding what pronouns, noun phrases, or even entire sentences refer to – is a fundamental challenge in natural language understanding. Traditional approaches often struggle because they either assume a fixed text structure or rely on external knowledge bases like frames and scripts, which can be brittle and difficult to adapt. This paper tackles the problem of dynamically determining the relevant context for anaphora resolution by analyzing the inherent structural connections within the text itself, rather than imposing external structures. They argue that the way writers present information guides the reader's understanding, and this guidance can be computationally leveraged.

Methods

The core of their approach is to model the text as a directed graph where nodes are sentences and edges represent "expansion" or "temporal continuation" relationships. They propose that each sentence either expands upon (provides more detail, cause, effect, etc.) or temporally continues a previous sentence or set of sentences. The default assumption is that a sentence relates to the immediately preceding one, but this can be overridden by semantic and world knowledge.

The process involves:

  1. Building a Context Graph: As each new sentence is processed, it's fitted into the existing graph of relationships.
  2. Hypothesizing Relations: The algorithm hypothesizes expansion/temporal continuation relations between the new sentence and previous ones.
  3. Testing Hypotheses: These sentence-level hypotheses are then used to derive and validate item-to-item reference hypotheses. This means that if a sentence is determined to be an expansion of a previous one, then anaphoric references within that sentence are more likely to relate to items in the previous sentence.
  4. Dynamic Context Extraction: This graph-based approach allows for the dynamic extraction of local context at each point in the text, adapting to the evolving narrative.

Datasets

Not discussed in the paper. The paper uses examples (like the "circus" and "late to work" scenarios) to illustrate their concepts, but does not detail formal training or evaluation datasets.

Key Insights and Ideas

  • Text Structure as a Primary Guide: The paper strongly advocates for using the inherent structural relationships between sentences (expansion/continuation) as the primary driver for anaphora resolution, rather than relying solely on external knowledge.
  • Dynamic Context is Crucial: Local context isn't static; it needs to be dynamically extracted and updated as the text unfolds. Their graph-based model achieves this.
  • Writer's Choices Matter: The way a writer structures their sentences and groups semantic predicates significantly influences how readers will resolve later references. This highlights the importance of analyzing surface-level syntactic choices.
  • Broader Definition of Anaphora: They adopt a broad definition of anaphora, including not just pronouns but also definite noun phrases, pro-verbs, and even sentence-level references, making their resolution task more comprehensive.

Key Results

Not discussed in the paper. The paper presents a theoretical framework and an algorithmic sketch, but does not include quantitative results from an implemented system or evaluation on specific datasets.

Limitations and Future Work

  • "Guess and Test" Algorithm: The authors acknowledge that an anaphora resolver must be a "guess and test" algorithm, implying that their approach involves hypothesis generation and validation, which can be computationally intensive.
  • Need for Directed Inferencing: The resolution process requires directed inferencing that leverages semantics and world knowledge, in addition to text structure. The paper doesn't detail the specifics of this inferencing mechanism.
  • Future Work: The paper mentions that their ideas can form the "rudiments of a theory of how frames/scripts can be dynamically shifted in and out of focus under control of a text's structure." They also refer to a technical report (Columbia University AI Project Technical Report 77-2-1) for a sketch of an algorithm, suggesting that implementation and empirical evaluation are the next logical steps.

Summary updated

BibTeX
@inproceedings{1977use,
  title = {{The Use of Dynamically Extracted Context for Anaphoric Reference Resolution}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1977},
}
Figure