Towards Deep Understanding in Task-Oriented Dialogue Systems Using Deep Reinforcement Learning

对话是人类获取和吸收信息的自然方式。自20世纪60年代以来，已经开发了几代人工会话代理，但我们距离能够像人类一样灵活和有效地响应的对话系统还有很长的路要走。有监督的端到端深度学习（DL）方法克服了以前管道方法的几个限制，但仍然需要大量的训练集，并且由于无法探索训练数据之外的内容，可能会缺乏概念理解。这项工作旨在通过将接地语言技术纳入深度强化学习（DRL）代理，在基于文本的模拟（如bAbI基准测试）中进行端到端训练，来改进面向任务的对话系统。

Conversation is a natural way for humans to obtain and absorb information. Several generations of artificial conversational agents have been developed since the 1960's, but we are still a long way from dialog systems which can respond as flexibly and usefully as a human can. Supervised end-to-end Deep Learning (DL) approaches overcome several limitations of previous pipeline methods, but still require large training sets and may suffer from a lack of concept understanding due to an inability to explore beyond the training data. This work intends to deliver improvements in task-oriented dialog systems by incorporating grounded language techniques into Deep Reinforcement Learning (DRL) agents, trained end-to-end in text-based simulations such as the bAbI benchmark [1].