文章导读 柴语生是2022年11月投放市场的新型聊天机器人，支持多任务连续对话和问答，其特点为：可进行多轮对话、语言自然流畅、会承认错误、能质疑不正确的前提等。柴语生由大规模语言模型GPT–3.5支持，并在2023年3月升级到GPT–4，其服务也相应升级。GPT全称为生成式预训练转换器模型（Generative Pre-trained Transformer），是一种大规模语言模型（Large Language Model，LLM，以下简称大模型）。典型的大模型还包括BERT、RoBERTa、T5、XLNet和悟道等规模不一、技术路线各异的模型。与一般语言模型相比，大模型的训练语料规模巨大，参数急剧膨胀。GPT–3的参数规模达到1750亿，GPT–4已突破万亿，悟道2.0和Switch Transformer则达到了惊人的1.75万亿和1.6万亿。
超大规模的训练数据和参数量，使得大模型产生了知识涌现能力。在面对提问时只需给出几个任务实例，大模型即可通过推理给出举一反三式的回答，且具备将复杂问题拆分开来的步骤推理能力。因此，它在阅读、答问、摘要、翻译等173项语言相关测试上均有优异表现，甚至在编程和画图上也可圈可点（Wei et al. 2022；Nanda et al. 2023）。
Abstract：Over the past six months, the application of large language models such as ChatGPT has drawn international attention and sparked critical reflection in the whole world. In this paper, it is argued that these large language models should be viewed as instrumental tools that bring about benefits with their technological development as well as risks in the application. Consequently, their governance should be focused less on technological intervention, and more on language resources vital for their development and application. Regarding the governance of language resources in large language model development, efforts should be made to break down the data silos of Chinese language resources, develop distributed model construction technologies through federated learning, establish open-accessed national knowledge data mechanisms, and expand the open and efficient language data exchange markets. These efforts are aimed to promote Chinese expression of world knowledge and facilitate the development of Chinese large language models. Since the large language models are an important language resource in nature, their fundamental resource status should be emphasized in the application, and perspectives of standardization, evaluation, and ethical regulation should be taken in their governance.
饶高琦，胡星雨，易子琳. 语言资源视角下的大规模语言模型治理[J]. 语言战略研究, 2023, 8(4): 19-29.
Rao Gaoqi, Hu Xingyu and Yi Zilin. Governance of Large Language Models from the Perspective of Language Resources. , 2023, 8(4): 19-29.
