BleacherBot: AI Agent as a Sports Co-Viewing Partner

Research Background and Issues

• What problems or challenges did the authors identify?

  • With technological advancements, traditional co-viewing has been extended to new forms of remote and real-time viewing. However, remote viewing still faces numerous challenges, such as coordinating schedules with preferred viewing partners and lacking personalized viewing experiences.
  • Multi-user interaction environments on second-screen devices or platforms often fail to provide a satisfactory viewing experience. Users desire more intimate and personalized viewing modes.
  • Current rule-based AI systems have limited performance, lack real-time contextual adaptability, and fail to establish genuine emotional connections.

• Why is this issue important?

  • Co-viewing scenarios, such as sports events, have significant potential to enhance audience engagement, emotional connection, and social interaction. Effectively addressing these issues can not only improve the viewing experience but also fill existing technological gaps through innovative AI solutions.

• Research Motivation and Related Work

  • The rapid development of large language models (LLMs) has introduced new opportunities for viewing experiences and social interactions, such as providing human-like conversations and emotional responses through natural language processing.
  • Cited literature in the study indicates a growing trend of integrating social TV with personalized media experiences. However, there is still a lack of in-depth research on emotional interaction and co-viewing environments in AI viewing companions.

Solution

• What methods or solutions did the authors propose?

  • The authors developed an AI viewing companion named "BleacherBot," which is based on LLMs and optimizes user interaction through two emotional styles: high-arousal and low-arousal.
  • The system design emphasizes balancing emotional stimulation in sports co-viewing while supporting personalized user preferences.

• What are the innovative aspects of this solution?

  • Introduced an interaction design based on emotional intensity (high-low arousal) variations and implemented dynamic matching of user emotions with game events through large language models.
  • Utilized a two-stage fine-tuning technique to embed sports-specific knowledge into the model and adjust the emotional intensity of language output to suit specific scenarios.
  • Incorporated alignment with the user’s team into the design to enhance resonance and immersion in the viewing experience.

• What are the implementation steps and key technologies used?

  1. Formative Study:
     • Conducted preliminary testing with 10 participants to observe their reactions when interacting with a baseline version of the LLM.
     • Identified three key design elements: emotional arousal levels, response conciseness, and team support.
  2. Model Fine-Tuning:
     • Stage 1: Injected basic baseball knowledge, including game rules, strategies, and contextual understanding.
     • Stage 2: Adjusted emotional outputs to match high-arousal/low-arousal interaction styles, creating four different model versions.
  3. Main Experiment:
     • Designed and conducted an experiment where 27 participants interacted with both high-arousal and low-arousal versions of BleacherBot to evaluate the impact of emotional arousal levels on user immersion, emotional sharing, and social presence.

Research Findings

• What specific results were achieved?

  • Experimental results showed:
    • The high-arousal version of BleacherBot performed better in enhancing emotional engagement, immersion, and overall enjoyment.
    • The high-arousal version scored significantly higher on the "social presence" metric, particularly in collaborative viewing and psychological involvement.

• What advantages does it have compared to existing solutions?

  • Addressed the lack of emotional resonance in rule-driven AI through personalized emotional stimulation design.
  • More accurately reflected the dynamic intensity and emotional resonance of sports events in user interactions, enhancing the sense of authentic interaction between AI and users.

• What were the experimental or evaluation results?

  • The high-arousal AI was perceived as more reflective of human emotional responses, with an average "social satisfaction" score of 3.852 (out of 5), significantly higher than the low-arousal version's 3.148.
  • The high-arousal version significantly outperformed the low-arousal version in dimensions such as "immersion," "emotional sharing," and "entertainment," demonstrating statistical significance in the results.

• Limitations and Future Directions

  • Limitations:
    • The current study is based on pre-recorded game footage, which may not fully capture the higher uncertainty and dynamic interactions of live broadcasts.
    • The scope is primarily focused on baseball scenarios and has not been extended to other sports or non-sports content.
    • The specified LLM (KoAlpaca-Polyglot) completes tasks but tends to exhibit context breakdowns in multi-turn conversations.
  • Future Directions:
    • Expand to other live content domains (e.g., movies, other team sports) to validate cross-context applicability.
    • Investigate the long-term impact of extended interactions on user engagement.
    • Explore more advanced language models to improve multi-turn interaction capabilities in complex scenarios.
    • Deepen research into the long-term social impact of AI co-viewing systems (e.g., effects on real interpersonal relationships).

Through this study, the authors not only provide practical insights into AI-enhanced social viewing but also propose a new set of design principles that can contribute to broader applications in the future!