Search and Inference Strategies in Pronoun Resolution: an Experimental Study

Kate EhrlichACL 1981 · DOI

Abstract

This study investigates the cognitive mechanisms underlying pronoun resolution, specifically distinguishing between serial search strategies and lexical inference processes. Through behavioral experiments, the research examines how the distance of an antecedent and the requirement for bridging inferences affect response times and accuracy. The findings suggest that humans employ a serial search starting with the most recent antecedent and perform lexical inferences spontaneously during the reading process rather than post-hoc.

Introduction

Pronoun resolution—the process of determining the referent of an anaphoric expression—is a cornerstone of language comprehension. This paper addresses a fundamental tension in linguistics: is pronoun resolution driven by simple structural rules like recency, or by complex semantic reasoning? Specifically, the study seeks to understand the interplay between search strategies (the machinery dictating the order of candidate evaluation) and inference strategies (the evaluator checking semantic fit).

Methods

The author conducted two behavioral experiments using a "read-and-respond" paradigm. Subjects read three-sentence passages where the first two sentences introduced potential antecedents (e.g., Fred and John) and the third contained a pronoun.

Two primary variables were manipulated:

  1. Distance: The target antecedent was either "Near" (the most recent noun) or "Far" (the noun from the preceding sentence).
  2. Inference: The third sentence either matched the context directly or required a "bridging" lexical inference (e.g., linking "in his room" to a previous mention of being "inside").

Study 1 measured response time (RT) and accuracy (PC) immediately after reading. Study 2 introduced a 1.5-second delay to determine if these inferences occur "online" (during reading) or "offline" (prompted by a question).

Results

The quantitative data reveals a clear "cost" associated with both distance and the need for inference. As shown in the experimental results, resolving a "Near" antecedent was significantly faster and more accurate than resolving a "Far" one.

Table 1: Response times (R.T.) and percentage correct (P.C.) across distance and inference conditions.

Key observations from the data include:

  • Recency Advantage: Near antecedents were resolved in approximately 1.32s compared to 1.56s for far antecedents.
  • Accuracy Gap: Accuracy dropped from 95% for near antecedents to roughly 70-72% for far antecedents, suggesting that serial search may fail when the most recent candidate is not the target.
  • Inference Latency: Requiring a lexical inference added approximately 100ms to response times for near antecedents (increasing from 1.32s to 1.42s).

Discussion and Conclusion

The results support a Serial Search Hypothesis, where the brain checks the most recent noun first. However, search alone is insufficient; the "inference engine" must validate the candidate. A critical finding was that for unambiguous text, these inferences are performed spontaneously during the act of reading. Study 2 demonstrated that when a delay is introduced, the inference penalty shifts from response time to reading time, proving the cognitive work is integrated into the initial comprehension phase. This suggests that pronoun resolution is a dynamic process where structural search and semantic evaluation work in tandem to achieve coherence.

Summary updated

BibTeX
@inproceedings{ehrlich1981search,
  author = {Kate Ehrlich},
  title = {{Search and Inference Strategies in Pronoun Resolution: an Experimental Study}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {1981},
  doi = {10.3115/981923.981946},
}
Figure