Anaphora Resolution: Short-Term Memory and Focusing

Raymonde GuindonACL 1985 · DOI

Raymonde Guindon’s 1984 paper represents a seminal intersection of computational linguistics and cognitive psychology. It investigates the cognitive mechanisms underlying "focusing"—the process by which we manage mental representations of entities during discourse to resolve anaphors like pronouns and definite descriptions.

Bridging the Computational-Cognitive Gap

In the early 1980s, researchers like Grosz and Sidner proposed "focusing" as an engineering solution for AI systems to track salient entities. Guindon sought to address the "validity gap": do these computational frameworks mirror human cognitive architecture? The paper posits that anaphora resolution is governed by the structure of human memory, specifically the interplay between a high-speed, limited-capacity Short-Term Memory (STM) "cache" and a larger, slower Operating Memory (OM).

Experimental Methodology

Guindon employed a novel "on-line activation" technique using Rapid Serial Visual Presentation (RSVP). Participants read sentences word-by-word and were periodically interrupted by a "test word" for an old-new recognition task. The experiment manipulated the Stimulus Onset Asynchrony (SOA)—the timing of the test word relative to the anaphor—to track the activation of referents over time.

Experimental design showing word sequences and SOAs

The study used 36 carefully controlled experimental texts. Each featured a referent (e.g., "thermometer") and an anaphor (e.g., "instrument"). Conditions were varied to test "In Focus" scenarios (immediate mention) versus "Not in Focus" scenarios (intervening topic shift).

Key Insights: The Cache Strategy

The study suggests the brain treats recent discourse as a high-speed cache. Anaphora resolution involves checking this cache first; if the referent is missing, the brain must "reinstate" it from Operating Memory.

One of the most intriguing findings is the "surprise effect" of over-specification. Using a full noun phrase (e.g., "the instrument") for an entity already in focus actually increases processing time. This violates the "cooperative principle" of communication; the brain expects a simple pronoun for items in focus and experiences a processing delay when presented with a more complex, redundant description.

Results and Discussion

The data confirmed several key hypotheses regarding memory management:

  • Focus Advantage: Recognition was significantly faster for items "In Focus" (~1100ms) than those "Not in Focus" (~1300ms), supporting the existence of a fast-access mental cache.
  • Anaphors as Triggers: Anaphors acted as effective memory triggers. Reading an anaphor facilitated the retrieval of an out-of-focus referent, pulling it back into the active cache.
  • Penalty of Redundancy: In "In Focus" conditions, using a definite noun phrase instead of a pronoun caused a statistically significant increase in reaction time (p < 0.001).
Recognition latencies for items in and out of focus across priming conditions

Limitations and Modern Context

While foundational, the study's scope is limited to simple pronouns and definite descriptions, leaving aside complex "bridging inferences" (e.g., "I looked at the car; the engine was smoking"). Furthermore, while the RSVP method provided "on-line" data, the exact millisecond-by-millisecond breakdown of linguistic analysis remains a subject for further research.

From a modern perspective, these findings predate transformer-based "attention" mechanisms. However, they provide a vital cognitive blueprint for how human memory constraints shape the way we process and produce language.

Summary updated

BibTeX
@inproceedings{guindon1985anaphora,
  author = {Raymonde Guindon},
  title = {{Anaphora Resolution: Short-Term Memory and Focusing}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {1985},
  doi = {10.3115/981210.981237},
}
Figure