Text Overlap : an LLM with human-like conversational behaviors

다국어 초록 (Multilingual Abstract)

Traditional text-based human-AI interactions typically follow a strict turn-taking approach. This rigid structure limits conversational flow, unlike natural human conversations, which can freely incorporate overlapping speech. However, a pilot study suggests that even in text-based interfaces, overlapping behaviors such as backchanneling and proactive responses lead to more functional exchanges. This paper introduce text-based overlapping interactions as a new challenge in human-AI communication, characterized by real-time typing, diverse response types, and interruptions. To enable AI systems to handle such interactions, this paper defines three core tasks: deciding when to overlap, selecting the response type, and generating utterances. This paper constructs a synthetic dataset for these tasks and train OverlapBot, an LLM-driven chatbot designed to engage in text-based overlapping interactions.
Quantitative and qualitative evaluations show that OverlapBot increases turn exchanges compared to traditional turn-taking systems, which is perceived as efficient by end-users. This finding supports overlapping interactions and enhances communicative efficiency and engagement.

국문 초록 (Abstract)

전통적인 텍스트 기반 인간-AI 상호작용은 엄격한 턴-테이킹(turn-taking) 접근 방식을 따른다. 이는 오버랩 발화를 자유롭게 수용하는 자연스러운 인간 대화와 달리 대화의 흐름을 경직되게 제한한다. 그러나 본 파일럿 스터디는 텍스트 기반 인터페이스에서도 백채널링(backchanneling)이나 선제적 응답(proactive responses)과 같은 오버랩 행동이 보다 자연스럽고 기능적인 교환을 유도함을 시사한다.
이에 착안하여, 본 논문은 실시간 타이핑 가시성, 다양한 응답 유형, 중단 허용을 특징으로 하는 텍스트 기반 오버랩 상호작용을 인간-AI 커뮤니케이션의 새로운 패러다임으로 제시한다. AI 시스템이 이러한 동적 상호작용을 처리할 수 있도록, 핵심 작업을 오버랩 타이밍 결정, 응답 유형 선택, 발화 생성의 세 가지로 정의하고, 이 작업을 위한 합성 데이터셋을 구축한다. 그리고 이를 기반으로 텍스트 오버랩 상호작용을 수행하도록 설계된 LLM 기반 챗봇인 OverlapBot을 훈련한다.
정량적 및 정성적 평가 결과, OverlapBot은 기존의 턴-테이킹 시스템에 비해 턴 교환 횟수를 크게 증가시켰다. 구체적으로 사용자 턴은 72%, 챗봇 턴은 130% 증가했으며, 최종 사용자들은 이를 효율적으로 인지했다. 이 결과는 오버랩 상호작용이 커뮤니케이션 효율성과 참여도를 효과적으로 향상시킬 수 있음을 입증한다.

목차 (Table of Contents)

• Chapter 1. Introduction 1
• 1. Research Objectives and Methodology 1
• A. Research Background and Core Objective 1
• B. Rationale and Definition of Key Challenges 3
• C. Research Methodology and Structure 5
• 2. Related Work 7
• A. Large Language Models and Conversational Design 7
• B. Overlap in Human Communication 9
• C. Real-time Text Messaging and Synchronicity 10
• Chapter 2. Empirical Foundation and Interaction Definition 12
• 1. Study Design and Procedure 12
• A. Participants and Recruitment 12
• B. Task and Data Collection 13
• C. Data Analysis 13
• D. Observed Conversational Behaviors 14
• E. Interruption and Resolution Mechanisms 14
• F. User Experience and Implications 15
• G. Conclusion of the Pilot Study 16
• 2. Text Overlap Interactions 16
• A. Timing Classification: Deciding When to Overlap 19
• B. Action Classification: Selecting the Response Type 19
• C. Utterance Generation: Producing Coherent Responses 20
• Chapter 3. Experiment and Evaluation 22
• 1. OverlapBot Development 22
• A. Training Strategy and Data Synthesis 22
• B. Synthetic Dataset Construction 22
• C. Model Fine-Tuning: OverlapBot 25
• D. Interface Design and Development 27
• 2. Evaluation Protocol 30
• A. Automatic Evaluation 30
• B. User Evaluation: Comparing OverlapBot and Conventional Chat 30
• 3. Result 33
• 4. Discussion and Future Work 37
• A. Relational Context of Overlap 37
• B. Rethinking the Necessity of Prompting Design 38
• C. User-Customizable Overlap 38
• D. Culturally Adaptive Overlap 39
• E. Computation Inefficiency 39
• Chapter 4. Conclusion 41
• 1. Conclusion 41
• References 42