语言交流是人与人之间理解和沟通的桥梁，也是其他更为复杂社会活动的基础。然而， 在语言交流的过程中，交流者之间达成共同理解的计算神经机制尚不清楚。以往研究提出, 共享表征与人际预测编码可能是促进共同理解达成的关键机制，但目前仍缺乏相应的实证 研究对这一假设进行验证。针对这一问题，本研究结合计算模型和 fNIRS(functional near- infrared spectroscopy)群脑超扫描方法，对上述假设进行检验。本研究招募了 60 名被试， 随机分为 30 个同性别双人交流小组，每个小组均先完成交流任务，再进行功能定位任务。 在交流任务中，被试需要在每个话轮上通过词汇与同伴进行交流，从而达成特定的交流目 标。交流目标共有三类，分别为说出相同词、说出同类词和说出任意词，第一个为实验条 件，其他两个为控制条件。实验过程中记录被试所说的词，并编码说词、想词和听词的时 间，同时采集大脑活动信号。在功能定位任务中，被试需要被动听词打乱、句打乱以及完 整故事三种类型的音频材料。在听材料的过程中通过 fNIRS 记录被试的大脑活动信号。研 究结果发现，在认知计算层面，共享表征与人际预测以一种交互的方式动态调节交流行为， 进而影响共同理解的达成过程。在神经计算层面，颞叶前部、颞顶联合区以及内侧前额叶 皮层构成了交流相关的神经计算网络。其中，颞叶前部与颞顶联合区分别处于不同的语言 计算层级，并以交互的方式完成层级间信息的传递。具体而言，颞叶前部表征了语言交流 过程中的语义共享表征，这一计算过程对应短时程语言单元的加工;颞顶联合区表征了人 际预测变量，这一计算过程对应长时程语言单元的加工，内侧前额叶整合来自颞叶前部以 及颞顶联合区的交互信息并进一步完成语言输出。本研究的结果支持并拓展了人际交流层 级模型，验证了人际预测与共享表征在交流过程中的重要作用，并进一步给出了这一计算 过程背后的神经网络基础。这些结果对于深入理解语言交流的认知神经机制有重要的理论 价值，同时也为大规模语言模型技术的发展提供了认知计算机制启发和脑科学依据，具有 一定的应用价值。

Language communication is the cornerstone of communication between individuals, as well as the foundation for more complex social activities. However, the neurocognitive computational mechanisms underlying the mutual understanding between communicators in language communication are not yet clear. Previous research has proposed that shared representations and interpersonal predictive coding may be the key mechanisms that facilitate the achievement of mutual understanding, but there is a lack of empirical research to verify this hypothesis. This study combined cognitive computational models with the functional near-infrared spectroscopy (fNIRS)-based multi-brain hyperscanning method to test this hypothesis. The study recruited 60 participants who were randomly divided into 30 same-gender dyads. Each dyad completed a communication task followed by a functional localization task. In the communication task, participants had to communicate with their partners using words to achieve a specific communication goal in each turn. There were three types of communication goals: i.e., saying the same word, saying a word in the same category and saying any words, with the first being the experimental condition and the other two being the control conditions. The words spoken by participants were aurally recorded, based on which the time for speaking, thinking, and listening to words were coded. Brain activity was measured during this process. In the functional localization task, participants passively listened to three types of audio materials: i.e., word- scrambled, sentence-scrambled and intact stories. Brain activity was recorded during the listening process using fNIRS. The results showed that, at the behavioral level, shared representations and interpersonal prediction dynamically regulated communication behavior in an interactive way, thereby affecting the process of mutual understanding. At the neural level, the anterior temporal lobe (ATL), the temporoparietal junction (TPJ), and the medial prefrontal cortex (mPFC) constituted a neural computational network for language communication. Specifically, the ATL represented semantic shared representations, which occurred at the lower level of language hierarchy; the TPJ represented interpersonal prediction variables, which occurred at the higher

level of language hierarchy; and the mPFC integrated interactive information from the ATL and TPJ and further completed language output. The results of this study support and extend the hierarchy model for interpersonal language communication, validate the important role of interpersonal prediction and shared representations in verbal communication, and provide a neural network basis for the computational processes underlying these mechanisms. These findings are important for a deeper understanding of the neurocognitive mechanisms of language communication, and also provide cognitive computational inspiration and brain evidence for the development of large-scale language model technology.