ExploreSelf: Fostering User-driven Exploration and Reflection on Personal Challenges with Adaptive Guidance by Large Language Models

Research Background and Issues

• Issues or Challenges: The paper highlights that while expressive writing has been proven to improve psychological and physical health by alleviating stress and emotions, many individuals face difficulties in engaging with reflective writing interventions. This is primarily due to the heavy cognitive burden of organizing thoughts and emotions, which may lead individuals to abandon the process. Furthermore, traditional writing support methods (e.g., predefined prompts) lack flexibility and fail to adequately consider users' specific contexts.
• Importance: Reflective writing facilitates self-understanding and helps individuals find meaning in negative events, which is crucial for mental health and personal growth. However, current methods fall short in supporting diverse reflective pathways and risk losing user engagement due to inappropriate guidance.
• Research Motivation and Related Work:
  • Utilizing large language models (LLMs) to provide personalized guidance during the writing process has shown potential. Existing research primarily focuses on creative writing, with limited studies on LLMs for self-expression and mental health support.
  • Unlike previous chatbot-based approaches, ExploreSelf emphasizes user-driven writing and exploration rather than simple conversational formats.

Solution

• Proposed Solution:
  • An interactive writing system named ExploreSelf is designed, combining the generative capabilities of LLMs to support users in exploring and reflecting on personal challenges. Users can expand on initial narratives and delve deeper through reflective questions generated by the LLM.
  • The system offers keywords and instant feedback as auxiliary guidance and can generate summaries of thematic exploration, helping users organize and refine their thoughts.
• Innovations:
  • Compared to traditional writing support tools, ExploreSelf prioritizes user autonomy, enabling users to freely choose their reflective directions.
  • By dynamically generating questions and prompts, the system avoids direct intervention in content, encouraging users to express their unique linguistic styles and thought structures.
  • Inspired by psychological counseling and user-driven design, the system incorporates flexible navigation components (e.g., topic selection, self-questioning style prompts, keywords, and summaries).
• Implementation Steps and Key Technologies:
  • Functional Design:
    • Users input an initial narrative.
    • The system generates relevant topics, allowing users to choose exploration directions.
    • Multiple questions are generated for each topic to guide users in multi-angle thinking.
    • Keywords and instant feedback are provided to reduce emotional burden.
    • Summaries based on user records are generated to help organize thoughts.
  • Technical Implementation:
    • The backend utilizes OpenAI's GPT-4 API to manage key generative tasks.
    • "Chain-of-thought prompting" is employed to optimize LLM generation, enhancing contextual relevance and output quality.
    • The frontend is built using React to deliver a smooth user experience.

Research Outcomes

• Specific Results:
  • Experiments involving 19 participants revealed that ExploreSelf helped users engage more meaningfully in the reflective writing process, significantly enhancing their sense of control over the exploration process and deepening their self-reflection.
  • Participants could independently decide exploration paths, either delving into existing topics or exploring new ones. Additionally, they expressed satisfaction with the dynamic guidance provided through keywords and instant feedback.
• Advantages Compared to Existing Solutions:
  • ExploreSelf places greater emphasis on user autonomy, achieving thematic and question flexibility through open-ended exploration, unlike traditional fixed-prompt tools or conversational mechanisms.
  • Compared to previous LLM-driven methods, ExploreSelf reduces reliance on single-session conversations or output quality, allowing users to switch and expand across multiple pathways.
• Experiment or Evaluation Results:
  • The Pathways Subscale was used to assess the impact of visual design on users' perceived sense of control, with average scores increasing significantly after exploration (from 22.32 to 24.95).
  • Interaction logs revealed personalized usage patterns (e.g., topic selection, question generation, and keyword unlocking), validating the system's diverse interaction features.
• Limitations and Future Directions:
  • Limitations:
    • The study was based on a single experiment and did not explore the effects of multiple sessions or long-term use.
    • The research focused on Korean users, leaving the system's applicability and potential biases in other cultural contexts unexamined.
    • The safety of reflective emotional engagement for users with severe psychological issues (e.g., trauma or depression) has not been thoroughly investigated.
  • Future Directions:
    • Explore designs supporting multi-session and long-term use, including cross-session memory and data summary optimization.
    • Develop personalized solutions for users from diverse cultural and linguistic backgrounds to reduce potential cultural biases in language models.
    • Design safety mechanisms for complex or sensitive challenges, such as incorporating psychology-informed fuzzy guidance strategies or multi-choice prompts.

In summary, ExploreSelf demonstrates the potential of a reflective writing system that combines user autonomy with adaptive LLM support, offering new perspectives for the development of future mental health technologies.