面对面交谈是语言最基本、最主要的用法。大语言模型的语言生成技术成功与否，关键在于能否在人机交互中最大限度地贴近人类语言的社会属性与交际本质。对此，互动语言学比传统语言学更能提供贴近交际本质的指标，它将隐性的言语行为规律与相关交际机制转译为可验证的显性知识，揭示人类如何、也指出机器应该如何操控语言，以达成自然的社会互动。互动语言学的一个核心概念TRP（话轮转换相关位置）已被纳入AI对话系统的标注框架，用来预测可能的话轮结束点，为大模型补齐“何时该接话”的互动感知短板，实现人机交互时话轮的流畅转换。韵律手段、填充成分、词语修复等“不流畅物”也被嵌入AI语言产出，使之显得更为自然。本研究采集大语言模型之间的对话语料，与人类自然口语会话进行对比，进一步考察话语中的“不流畅物”以及基于TRP预测而产生的话语重叠。分析显示，这些语言表现在人类言语中均承载互动功能，且有规律可寻；而在大模型语言产出中却较为稀少甚至缺失。研究结果表明：语言学一方面应深化言语互动行为模式的理论研究；另一方面应推动自然口语语料库的构建工作，打造包含多模态信息的精细标注语料库，以使人类互动复杂微妙的语言组织方式能被充分理解刻画，为人工系统的语言设计提供坚实的经验基石。

Face-to-face conversation is the most fundamental and pervasive form of language use. The success of language generation by large language model (LLM) technology therefore depends crucially on whether it can maximally exhibit the social and communicative properties of human language use. In this regard, interactional linguistics offers more communication-sensitive criteria than traditional linguistic approaches. By rendering tacit regularities of verbal conduct and their underlying interactional mechanisms explicit and empirically testable, it reveals how speakers deploy linguistic resources to accomplish natural social interaction and, by extension, how artificial systems might be designed to do so. In this regard, interactional linguistics offers insights that better profile the nature of communication than traditional linguistic approaches. It uncovers patterns of speech behavior and related interactional mechanisms, revealing how humans manipulate language to achieve natural social interaction, and by extension, how machines should do so. As a matter of fact, some core concepts in interactional linguistics have already been incorporated into the annotation frameworks of AI dialogue systems. TRP (Transition Relevance Place), for one, has been introduced for the prediction of potential turn-ending points, enhancing LLMs’ “interactional awareness” of when to take the floor, thereby enabling smooth turn-taking in human-machine interaction. Moreover, even in the pre-LLM era, AI developers at Google and elsewhere had already figured out that being too fluent is unnatural, and embedded discourse markers and other “disfluencies” into machine language output. In light of the linguistic concepts already adopted by AI developers, this study collects dialogues conducted between large language models and compares them with naturally occurring human conversations. It further examines discourse markers, pauses, elongations, repairs, and other “disfluencies”, as well as turn overlaps arising from TRP-based predictions. The analysis reveals that these linguistic features all carry interactional functions in human speech and exhibit observable regularities, while in LLM language output, they are relatively scarce or even entirely absent. These findings point to two priorities for future research: on the one hand, linguistics studies should place greater emphasis on the recurrent patterns of speech interactional behavior in verbal interaction; on the other hand, greater efforts should be devoted to building multimodal corpora, which will form a solid empirical foundation for language design in artificial systems.