Coffee With a Hint of Data: Towards Using Data-Driven Approaches in Personalised Long-Term Interactions

Area of Science:

• Human-Robot Interaction
• Artificial Intelligence
• Machine Learning

Background:

• Traditional rule-based systems lack flexibility for real-world human-robot interaction.
• Data-driven approaches offer flexibility but typically lack long-term memory and personalization capabilities.
• Lifelong learning and few-shot learning are crucial for adaptive, personalized long-term human-robot interactions.

Purpose of the Study:

• To introduce the Barista Datasets for evaluating data-driven approaches in long-term human-robot interaction.
• To assess the performance of state-of-the-art data-driven dialogue models in personalized, long-term interactions.
• To identify limitations of current data-driven models in handling real-world challenges like recognition errors and preference changes.

Main Methods:

• Development of the text-based Barista Datasets simulating real-world interaction challenges.
• Evaluation of several state-of-the-art data-driven dialogue models: Supervised Embeddings, Sequence-to-Sequence, End-to-End Memory Network, Key-Value Memory Network, and Generative Profile Memory Network.
• Analysis of model performance in generic and personalized long-term human-robot interactions.

Main Results:

• Data-driven approaches are effective for generic task-oriented dialogue and real-time interactions.
• No evaluated model demonstrated sufficient performance for personalized long-term human-robot interaction deployment.
• Key limitations identified include the inability to learn new identities and poor recall of user-specific data.

Conclusions:

• Current data-driven dialogue models are not yet suitable for personalized lifelong learning in human-robot interaction.
• Significant advancements are needed in memory retention and identity management for robust, adaptive AI systems.
• The Barista Datasets provide a valuable resource for future research in this challenging domain.