[Paper Review] ChaCha: Leveraging Large Language Models to Prompt Children to Share Their Emotions about Personal Events

Woosuk Seo, Chanmo Yang+1 more|arXiv (Cornell University)|Sep 21, 2023

Topic ModelingComputer Science3 citations

TL;DR

ChaCha is a child-friendly chatbot that leverages large language models (LLMs) and a state machine to guide children (8–12 years old) in sharing personal stories and expressing associated emotions through free-form, emotionally supportive conversations. In a lab study, children perceived ChaCha as a close friend and shared diverse personal experiences, demonstrating the potential of LLMs to support children’s emotional expression in developmentally appropriate ways.

ABSTRACT

Children typically learn to identify and express emotions through sharing their stories and feelings with others, particularly their family. However, it is challenging for parents or siblings to have emotional communication with children since children are still developing their communication skills. We present ChaCha, a chatbot that encourages and guides children to share personal events and associated emotions. ChaCha combines a state machine and large language models (LLMs) to keep the dialogue on track while carrying on free-form conversations. Through an exploratory study with 20 children (aged 8-12), we examine how ChaCha prompts children to share personal events and guides them to describe associated emotions. Participants perceived ChaCha as a close friend and shared their stories on various topics, such as family trips and personal achievements. Based on the findings, we discuss opportunities for leveraging LLMs to design child-friendly chatbots to support children in sharing emotions.

Motivation & Objective

• To address the challenge of helping children, especially those without siblings, express their emotions effectively due to underdeveloped communication skills.
• To design a conversational agent that supports children in identifying and articulating emotions linked to personal events in a natural, free-form dialogue.
• To explore how large language models can be leveraged to create emotionally supportive, child-friendly chatbots that encourage emotional sharing.
• To investigate children’s perceptions and behaviors when interacting with an LLM-driven chatbot focused on emotional storytelling.

Proposed method

• ChaCha combines a state machine with a large language model (GPT-4) to maintain dialogue structure while enabling free-form conversation.
• The state machine guides the conversation through distinct phases: Explore (initiating story sharing), Elaborate (encouraging detail), and Reflect (prompting emotion identification).
• When children struggle to name emotions, ChaCha provides a curated list of emotion words for guidance, reducing cognitive load.
• The system uses prompt engineering to align LLM responses with developmental appropriateness and emotional support goals.
• The chatbot was implemented in Korean, with model selection based on multilingual performance and token efficiency to handle Korean text’s higher tokenization cost.
• A prototype was evaluated in a lab-based user study with 20 children (8–12 years old), using qualitative analysis to examine dialogue patterns and emotional expression.

Experimental results

Research questions

• RQ1 How do children interact with an LLM-driven chatbot when prompted to share personal events and associated emotions?
• RQ2 To what extent does ChaCha successfully guide children to express emotions using natural, free-form dialogue?
• RQ3 How do children perceive ChaCha in terms of trust, relatability, and emotional support?
• RQ4 What challenges arise in maintaining dialogue coherence and emotional guidance over longer conversations with LLMs?
• RQ5 How do cultural and linguistic factors (e.g., Korean language use) affect the performance and usability of such a system?

Key findings

• Children perceived ChaCha as a close friend and engaged in emotionally meaningful conversations, sharing stories on topics such as family trips and personal achievements.
• Participants frequently shared personal events and were able to identify and describe associated emotions, especially when guided by ChaCha’s emotion word list.
• The state machine effectively structured the conversation, enabling a balance between free-form dialogue and emotional guidance, even as dialogue history lengthened.
• Despite the LLM’s strong performance, long dialogue histories occasionally caused the model to deviate from intended phases, suggesting a need for additional response validation mechanisms.
• The use of Korean language increased token usage significantly, limiting the effective context window and highlighting the need for model selection based on language-specific efficiency.
• Children’s emotional expression was not significantly influenced by gender in this study, but future work should explore potential gender differences in emotional communication with LLM chatbots.