Utility-based generation of referring expressions.

Area of Science:

• Cognitive Science
• Computational Linguistics
• Artificial Intelligence

Background:

• Referring expression generation is crucial for human-computer interaction in task-oriented dialogue systems.
• Existing models, like the incremental algorithm, have limitations in explaining human language production.
• The iMAP task presents unique challenges, such as unreliable color features for identifying referents.

Purpose of the Study:

• To develop and evaluate two cognitive models for generating referring expressions in the iMAP task.
• To compare a modified incremental algorithm with a feature-utility-based template model.
• To understand how cognitive models can capture human adaptation in language use.

Main Methods:

• Development of a general cognitive model based on Dale and Reiter's incremental algorithm, enhanced with feedback adaptation.
• Development of a specialized template model prioritizing feature utility within the iMAP task context.
• Comparison of model predictions against observed human behavior in the iMAP task.

Main Results:

• The template model showed a higher correlation with empirical data compared to the adapted incremental model.
• The adapted incremental model overpredicted distinguishing expressions and underpredicted overspecified ones.
• Both models successfully predicted the observed decrease in color term usage and increase in useful feature term usage.

Conclusions:

• A feature-utility-driven approach, as implemented in the template model, better simulates human referring expression production in the iMAP task.
• Cognitive models need to account for task-specific properties and adaptive learning to accurately reflect human language behavior.
• The findings have implications for designing more natural and effective human-computer dialogue systems.