Natural Language Does Not Emerge ‘Naturally’ in Multi-Agent Dialog
Satwik Kottur1 José M.F. Moura1 Stefan Lee2,3 Dhruv Batra3,4 1Carnegie Mellon University, 2Virginia Tech, 3Georgia Tech, 4Facebook AI Research
Abstract
A number of recent works have proposed techniques for end-to-end learning of communication protocols among cooperative multi-agent populations, and have simultaneously found the emergence of grounded human-interpretable language in the protocols developed by the agents, learned without any human supervision!
中文速览
多智能体系统能否自发"聊"出一套人类可理解的组合性语言?研究者以两个强化学习智能体(Q-bot 和 A-bot)玩"任务与对话"参考游戏为实验台,系统考察了这一问题。结果发现,智能体总能发明出高效的通信协议、以接近满分的准确率完成任务,但这些自发涌现的语言几乎都不具备组合性(compositionality)或人类可解释性——词汇量足够大时,A-bot 会直接用唯一符号枚举每个物体,完全无需真正的对话。只有在对词汇量和通信方式施加层层限制之后,智能体才被"逼"出越来越接近人类语言结构的组合性表达。这项研究揭示了近期多智能体语言涌现文献中被忽视的关键盲点:任务成功并不等于语言可解释,自然语言绝不会"自然地"涌现,未来研究需要更明确地将组合性作为设计约束,而非期待它自动出现。
原文 arXiv:1706.08502;中英对照 + 大白话阅读 https://aha.fim.ai/paper/1706.08502v3