EAGER: Language Learning through Machine Theory of Mind

随着自然语言系统变得无处不在（例如电话树、聊天机器人和智能家居），它们必须学会通过将每个用户建模为具有不同能力、知识和品味的个体来适应用户。心智理论是人类对他人隐藏的心理状态进行推理的能力，但与其他更基本的沟通技能相比，它是一种复杂的现象，直到儿童发育的后期才出现。这个EAGER项目感兴趣的问题是：(1)儿童很难学习这项技能的原因，(2)计算机可以从儿童的教学方式中学到什么，以及(3)机器学习模型可以以何种方式提供对人类发展的洞察。这个项目位于机器学习、发展心理学和教育学的交叉点。该项目包括基于共享参考游戏的信息共享和教学的正式模型。智能体和儿童的任务是要求指导者有效地区分相似的物体——这一任务需要理解共同点并识别不同的特征。虽然学习者经常会做出模棱两可的陈述，但老师会提供纠正和指导，以指导学习过程。这个公式允许在与成功沟通相关的几个维度上发生变化：工作记忆、视觉和词汇复杂性以及教学的特殊性。以儿童为对象的实验将为比较计算智能体提供基准，而以智能体为对象的实验将使我们能够分解这些因素对发展心智理论的难度的贡献。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

As natural language systems become ubiquitous (e.g. phone trees, chatbots, and smart homes) they must learn to adapt to users by modeling them each as individuals with different abilities, knowledge, and tastes. Theory of mind is the human ability to reason about the hidden mental states of others, but is a complex phenomenon that does not emerge in children until late in their development compared to other more basic communicative skills. The questions of interest to this EAGER project are: (1) what makes this skill hard for children to learn, (2) what can computers learn from how children are taught, and (3) in what ways can machine learning models provide insight into human development. This project sits at this intersection of machine learning, developmental psychology, and pedagogy. This project includes formal models of information sharing and teaching grounded in shared referential games. Agents and children are tasked with asking an instructor to efficiently distinguish similar objects -- a task which requires understanding common ground and identifying distinguishing features. While the learner will often make ambiguous statements, the teacher will provide corrections and instruction to guide the learning process. This formulation allows for variation along several dimensions of relevance to successful communication: working memory, visual and lexical complexity, and specificity of instruction. Experiments with children will provide benchmarks against which computational agents can be compared, and experiments with agents will allow us to decompose the contribution of each of these factors to the difficulty of developing a theory of mind.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.