Child-Centric Robot Dialogue Systems: Fine-Tuning Large Language Models for Better Utterance Understanding and...

Area of Science:

• Artificial Intelligence
• Human-Computer Interaction
• Computational Linguistics

Background:

• Current dialogue systems, including large language models (LLMs), struggle to accurately interpret children's unique linguistic features like incomplete syntax and mispronunciations.
• Effective child-robot interaction necessitates dialogue systems that can comprehend children's utterance intentions, similar to human understanding.

Purpose of the Study:

• To develop a fine-tuning methodology for LLM-based dialogue systems to improve their ability to interpret children's utterance intentions.
• To enable natural and adaptive verbal interactions between children and robots, even with non-standard speech.

Main Methods:

• Proposed a fine-tuning methodology using two types of data: LLM-human judgment discrepancies and interactive response data.
• LLM-human judgment discrepancy data captured cases where LLM and human interpretations of children's responses differed.
• Interactive response data consisted of robot responses tailored to children's utterance intentions, generated by the LLM.

Main Results:

• Developed a fine-tuned dialogue system capable of human-like interpretation of children's utterances.
• The system demonstrated adaptive responses, effectively handling syntactic incompleteness and mispronunciations.
• Human assessments using Robotic Social Attributes Scale (RoSAS) and Sensibleness and Specificity Average (SSA) metrics validated the system's performance.

Conclusions:

• The proposed fine-tuning methodology significantly enhances LLM-based dialogue systems' performance in interpreting children's utterance intentions.
• The system facilitates more natural verbal interactions in child-robot scenarios, accommodating linguistic variations.
• This approach bridges the gap between LLM capabilities and human-level understanding of child-directed speech.