Customizing Emotional Support: How Do Individuals Construct and Interact With LLM-Powered Chatbots

Research Background and Issues

• Identified Problems or Challenges:

  • Diverse Psychological Support Needs: Many individuals suffer from poor mental health due to a lack of emotional support. Some need someone to talk to, others seek understanding, while some desire companionship or practical guidance. However, current psychological support systems have yet to fully explore how large language models (LLMs) can be customized to meet personalized psychological support needs.
  • Lack of Individualized Customization: Existing research focuses more on using LLMs to simulate mental health professionals or for general emotional support applications. However, there is limited research on how users can leverage the "customizability" of LLMs to design conversational agents tailored to their unique needs.
  • Limitations in Non-Linguistic Attributes: Current LLM research primarily focuses on adjusting textual outputs, with insufficient exploration of other attributes (e.g., voice, avatars) and their impact on emotional support effectiveness.

• Significance:

  • Tailored psychological support helps meet individual needs and promotes mental health.
  • Customization enhances user agency, enabling individuals to express their needs and receive personalized care.
  • The current research gap highlights the urgent need to develop LLM-powered conversational agents for personalized emotional support.

• Research Motivation and Related Work:

  • The research is driven by LLMs' advancements in large-scale natural language processing and generative reasoning, creating opportunities for users to customize chatbots.
  • Related studies include platforms like Character.AI, which allow users to customize predefined roles for chat agents. However, these rely on menu-based options and lack dynamic interaction and more flexible customization possibilities.

Solution

• Proposed Method or Solution:

  • Developed a prototype system called ChatLab, designed to enable users to create LLM-driven chatbots capable of providing emotional support.
  • In ChatLab, users can customize various attributes of the chatbot, including text prompts, voice, avatar settings, and interaction modes.

• Innovative Features:

  • Offers greater flexibility than traditional options through diverse customization features such as text editing, voice, and avatar selection, allowing users to build complex, personalized emotional support bots.
  • Encourages users to actively participate in customization rather than passively selecting preset options, enhancing their agency and creativity in design.
  • Integrates multiple social cues (e.g., non-verbal expressions) into the LLM customization process to strengthen emotional connections with users.

• Implementation Steps and Key Technologies:

  • Designed four functional modules: a chat customization interface (input descriptions and expectations), additional settings (voice, avatar), a conversation window, and a session history log.
  • Leveraged a prompt-based mechanism that allows users to guide chatbot outputs through natural language.
  • Introduced an "AI Refinement Feature" to help users create more effective prompts and support real-time adjustments to the language model.
  • Utilized technologies like Streamlit and LangChain for dynamic interaction and Firebase for data storage.

Research Outcomes

• Specific Results:

  • Participants created 118 unique chatbot roles within the ChatLab system, including roles such as friends, romantic partners, therapists, philosophers, and more.
  • Various voice and avatar settings were used to enhance the personalization of chatbot roles, creating richer and more diverse conversational experiences.
  • Surveys revealed that participants used the customization process to explore their inner thoughts, express complex emotions, and improve their psychological state.

• Advantages Compared to Existing Solutions:

  • Provides more flexible and detailed customization options (e.g., voice, visuals, prompts), enabling users to create more diverse chatbots.
  • The customization process goes beyond technical functionality, involving psychological exploration and reflection, which aids in self-discovery and personal growth.
  • Enriches social cues in human-computer interaction, enhancing emotional connection and user experience with support tools.

• Experimental or Evaluation Results:

  • Collected 1,541 dialogue rounds and 178 session logs during participant customization.
  • Embedded diary tracking showed that the design props facilitated users in reflecting on their emotional processes and sparking design inspiration.
  • Participant feedback indicated a desire for future chatbots to be more intelligent, integrating digital traces (e.g., social media activity) or physical states to provide more proactive support.

• Limitations and Future Directions:

  • ChatLab currently lacks integration with voice input, detailed avatar creation, and image generation models. Future research could explore these opportunities.
  • The sample is limited to an Asian cultural context; future studies should expand to diverse cultural samples to examine cross-cultural differences in customization and emotional support needs.
  • The long-term mental health impact remains unclear, requiring further investigation into its effects on loneliness and emotional fluctuations.

Through this study, the authors demonstrate the potential of LLM-driven chatbots in emotional support while uncovering the motivations behind users' desire to explore and express complex emotions through customization.