Personalized Language Model Selection Through Gamified Elicitation of Contrastive Concept Preferences

Area of Science:

• Natural Language Processing
• Human-Computer Interaction
• Gamification
• Visual Analytics

Background:

• Personalization of language models is crucial but challenging due to difficulties in acquiring user-specific training data.
• Existing methods often require extensive user data, which can be difficult and time-consuming to obtain.
• Adapting model parameters is not the only solution; matching models to users' mental models of language concepts is a viable alternative.

Purpose of the Study:

• To capture individual users' mental models of language concepts for personalized language model suggestions.
• To address challenges of user disengagement and the contextual nature of language understanding.
• To integrate gamification and visual analytics for effective user knowledge externalization.

Main Methods:

• Developed Concept Universe, a gamified visual analytics playground.
• Users explain concept names with keywords in a four-phased game.
• Implemented constant visual, verbal, and auditory feedback mechanisms.
• Integrated a virtual opponent to enhance user engagement and input specificity.

Main Results:

• User study with six participants demonstrated high user engagement.
• Gamified approach led to more specific user input compared to baseline.
• Generated personalized concept descriptions effectively.
• Personalized concepts were used for language model suggestions.

Conclusions:

• Gamified visual analytics can successfully engage users in externalizing their mental models of language.
• Concept Universe provides a novel approach to personalized language model selection.
• The method offers a promising alternative to traditional fine-tuning for personalization.