Towards Mutual Theory of Mind in Human-AI Interaction: How Language Reflects What Students Perceive About a Virtual Teaching Assistant

Title of the Paper

Towards Mutual Theory of Mind in Human-AI Interaction: How Language Reflects What Students Perceive About a Virtual Teaching Assistant

Paper Information

• Research Area: Human-AI Interaction, Conversational Agent Design, Online Education
• Keywords: Conversational Agents, Online Communities, Human-AI Interaction, Theory of Mind, Language Analysis, Online Education

Research Background and Problem Statement

• Identified Problems or Challenges:

  • Conversational agents (CAs) have achieved certain success in task-oriented interactions, but they still face significant technical and design challenges in achieving natural and long-term conversations.
  • In scenarios requiring long-term support, such as online education communities, CAs need to understand and adapt to the evolving needs and expectations of users and communities.
  • Existing designs lack a theoretical framework to help CAs perceive users' psychology and expectations. Moreover, research on automatically inferring user perceptions based on language data is relatively limited.

• Importance:

  • The ability to build long-term natural interactions in human-AI interaction is crucial for improving user experience and satisfaction. Particularly, as learning assistants, CAs need to establish trust and acceptance among users.

• Research Motivation and Related Work:

  • In human interactions, "Theory of Mind" (ToM) enables the prediction and understanding of emotions and behaviors between interacting parties.
  • The "Mutual Theory of Mind" (MToM) based on ToM has been applied in robot design, but its application in human-AI interaction, especially in language-based communication contexts, remains limited.
  • Existing studies primarily focus on exploring user perceptions of CAs through surveys and qualitative analyses, which are difficult to scale and automate.

Proposed Solution

• Methods or Solutions:

  • Propose using Mutual Theory of Mind (MToM) as a theoretical framework for designing long-term human-AI interactions.
  • Employ language analysis methods to extract data features reflecting students' perceptions from their interactions with the virtual teaching assistant (Jill Watson, JW).
  • Analyze students' long-term perception changes of JW across three dimensions: "Anthropomorphism," "Intelligence," and "Likeability."

• Innovations:

  1. Theoretical Framework: Introduce the MToM framework to guide the design of CAs that can adapt to users' evolving needs.
  2. Community-Oriented Perception Analysis: Shift the focus from studying individual user perceptions (static) to investigating collective dynamic perceptions of users.
  3. Relationship Between Language Features and Perceptions: Quantitatively reveal the relationship between language features (e.g., readability, diversity, adaptability) and user perceptions of CAs.

• Implementation Steps and Techniques:

  • Data Collection: Deploy JW in an online course with 376 students, collecting interaction records between JW and students in the Q&A forum, along with bi-weekly student perception survey data.
  • Feature Extraction: Extract language features such as verbosity, readability, sentiment, diversity, and adaptability from student responses.
  • Regression Analysis: Build linear regression models to analyze the relationship between language features and changes in student perceptions.

Research Findings

• Specific Findings:

  1. Long-Term Perception Changes: Students' perceptions of JW's anthropomorphism and intelligence significantly changed over the semester, while their perception of likeability remained relatively stable.
  2. Impact of Language Features:
     • Verbosity was negatively correlated with all three perception dimensions.
     • Features such as readability and language adaptability were effective positive predictors of perceptions.

• Comparison with Existing Solutions:

  • Most current studies use qualitative methods to evaluate CA performance, which are difficult to scale and dynamically adapt. This study is the first to introduce the MToM framework and quantitatively analyze user perceptions of CAs using language data, offering a highly automated and scalable solution.

• Experimental or Evaluation Results:

  • Regression analysis results revealed that language readability, adaptability, and semantic diversity features were significantly associated with positive user perceptions, while verbosity was negatively correlated.
  • The three user perception dimensions (anthropomorphism, intelligence, likeability) showed high intercorrelation.

• Limitations and Future Directions:

  1. Limitations:
     • The study data was collected in a controlled academic environment, and the results may not directly generalize to more diverse and informal online communities.
     • The analysis method focused on correlation, without directly uncovering causal relationships.
  2. Future Directions:
     • Explore how CAs can more intuitively express their capabilities through language to help users build accurate mental models of CAs.
     • Extend research to include private conversations and more complex multimodal (e.g., voice, visual) interaction environments.
     • Investigate how different language features reflect CA perceptions in various contexts and tasks.