Alignment in interactive reference production: content planning, modifier ordering, and referential...

Area of Science:

• Psycholinguistics
• Computational Linguistics

Background:

• Referring expression generation (REG) algorithms often lack sensitivity to conversational context.
• Existing REG models rely on domain-specific preferences for content and linguistic realization.
• Human language production in interactive settings is complex and context-dependent.

Purpose of the Study:

• To investigate if humans adapt their referring expression production based on prior interaction.
• To determine if speakers utilize dispreferred attributes and modifier orders after priming.
• To examine if overspecified references are more likely following priming.

Main Methods:

• Conducted three experiments to test human language production in interactive settings.
• Primed participants with specific attributes and modifier orders in preceding interactions.
• Analyzed the choices of attributes, modifier orderings, and level of description in referring expressions.

Main Results:

• Humans opted for dispreferred attributes and modifier orders when primed in prior interactions, often unconsciously.
• Speakers were more prone to producing overspecified references, including dispreferred attributes, after priming.
• Priming influenced language production beyond minimal, preferred descriptions.

Conclusions:

• Human referring expression generation is sensitive to recent conversational history.
• Unconscious priming significantly impacts the selection of linguistic features in referring expressions.
• Future REG algorithms should incorporate context-sensitivity and priming effects for more naturalistic output.