From Human Pragmatic Language Skills to Conversational Agent Design: A Systematic Review of Transfer Strategies

Paper Title

From Human Pragmatic Language Skills to Conversational Agent Design: A Systematic Review of Transfer Strategies

Publication Info

• Topic area: Human conversational skills applied to conversational agent design.
• Keywords: Conversational agents, pragmatic language skills, skill transfer, human-computer interaction, large language models, co-design, verbal skills, paralinguistic skills, nonverbal skills, ethical considerations.

Background and Problem

• Problem / challenge: Existing research on conversational agents (CAs) primarily focuses on semantic and syntactic capabilities, neglecting pragmatic conversational skills. There is no systematic methodology for transferring human conversational skills to CA design.
• Significance: Pragmatic conversational skills are critical for improving user satisfaction, trust, and social presence in human-CA interactions. Their absence leads to inefficiencies, dissatisfaction, and reduced task performance.
• Motivation and related work: Prior studies have explored human-like behaviors in CAs and domain-specific applications but lack a unified framework for skill transfer. Previous reviews have focused on technical implementation or evaluation methods without addressing the design process for anthropomorphic CA development.

Solution

• Proposed approach: A systematic review of 85 studies to identify conversational skills, transfer strategies, implementation methods, and evaluation metrics for CAs. A four-stage design process is proposed for skill transfer.
• Novelty:
  1. Categorization of conversational skills into verbal, paralinguistic, and nonverbal modalities.
  2. Identification of three primary transfer strategies: from dialog data, theories, and via co-design.
  3. Development of a four-stage Double Diamond design process for skill transfer.
  4. Ethical considerations and recommendations for responsible CA design.
• Procedure and key techniques:
  • Systematic review following PRISMA guidelines across six databases.
  • Categorization of skills and strategies using mixed-methods analysis.
  • Proposal of a structured design process integrating divergent and convergent thinking.

Results

• Concrete findings:
  • Verbal skills (e.g., empathy, turn-taking, stylistic variances) dominate CA design, followed by paralinguistic (e.g., prosody, fluency) and nonverbal skills (e.g., facial expressions, gestures).
  • Transfer strategies include dialog data analysis, co-design with human participants, and theory-driven approaches.
  • Implementation methods range from traditional model training and knowledge bases to LLM-based and rule-based systems.
  • Evaluation metrics include user satisfaction, linguistic quality, social intelligence, and task-specific outcomes.
• Advantage over baselines: The review synthesizes fragmented research and provides a structured methodology for skill transfer, addressing gaps in existing literature.
• Experiments / evaluation: Studies evaluated CAs using diverse metrics, including engagement, perceived empathy, and task performance, across domains like healthcare, education, and customer service.
• Limitations and future work: Lack of meta-analysis due to study heterogeneity; Western-centric focus; need for standardized evaluation frameworks and ethical guidelines.

Summary

This paper systematically reviews 85 studies to explore how human conversational skills are transferred to conversational agent designs. It categorizes skills into verbal, paralinguistic, and nonverbal modalities, identifies transfer strategies from dialog data, theories, and co-design, and proposes a four-stage Double Diamond design process for skill transfer. The findings highlight the dominance of rule-based implementations and the emerging potential of LLMs for sophisticated skill deployment. Ethical considerations are emphasized to address risks like bias and user dependency. The research provides actionable insights for creating human-like, socially appropriate, and effective CAs across diverse domains.