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 context from the text itself, rather than relying on pre-defined frames or scripts. The core idea is that the way a writer structures sentences provides crucial clues for resolving references, and this structure can be represented as a graph to guide the resolution process.

What Problem Does This Paper Address?

Anaphoric reference resolution is the task of identifying what a pronoun, noun phrase, or other linguistic element refers to in a text. For example, in "John went to the store. He bought milk," "He" refers to "John." Traditional approaches often struggle with the dynamic nature of context and rely on external knowledge bases like frames or scripts, which can be rigid and difficult to adapt. This paper addresses the limitation of existing methods by proposing a system that dynamically extracts contextual information directly from the text's structure, allowing for more flexible and accurate resolution. They argue that the way sentences are connected (expansion or temporal continuation) forms a "context graph" that is key to understanding these references.

Methods

The authors propose an algorithm that treats the relationships between sentences as a graph. Each sentence is seen as either expanding upon or temporally continuing previous sentences. This "context graph" is built dynamically as the text is processed. When a new sentence is encountered, the algorithm attempts to fit it into the existing graph by hypothesizing and testing sentence-level relationships (expansion/temporal continuation). Item-to-item reference hypotheses are then derived from these validated sentence relationships. This approach emphasizes using the text's inherent structure to guide the resolution process, moving away from reliance on external knowledge.

Datasets

Not discussed in the paper. The paper provides two examples (1 and 2) to illustrate their concepts but does not mention specific datasets used for training or evaluation.

Key Insights and Ideas

  • Text Structure as Context: The most significant insight is that the way writers structure their sentences and connect them (e.g., through expansion or temporal continuation) is a rich source of information for resolving anaphoric references. This structure can be modeled as a graph.
  • Dynamic Context Extraction: Instead of relying on static knowledge bases, the algorithm dynamically extracts context by analyzing the relationships between consecutive sentences. This allows for adaptation to the specific flow of any given text.
  • "Guess and Test" Algorithm: Because there's no definitive test for anaphoric reference, the proposed resolver uses a "guess and test" approach. It hypothesizes potential referents and then uses directed inferencing, incorporating semantics, world knowledge, and text structure, to validate these hypotheses.
  • Rejection of Frames/Scripts: The authors explicitly reject the notion that pre-defined frames or scripts are the primary solution to anaphoric resolution, arguing that they ignore the dynamic structural information present in the text itself.

Key Results

Not discussed in the paper. The paper focuses on presenting the theoretical framework and algorithmic approach rather than empirical results from a specific implementation or evaluation.

Limitations and Future Work

  • Limitations: The paper acknowledges that their approach is a "sketch of an algorithm" and that a full implementation and evaluation are needed. The reliance on "guess and test" implies potential computational complexity and the need for robust inferencing mechanisms.
  • Future Work: The authors suggest that their ideas can form the basis of a theory for how frames/scripts can be dynamically shifted in and out of focus under the control of a text's structure. They also refer to a technical report (Columbia University AI Project Technical Report 77-2-1) for a more detailed algorithmic sketch.

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