This survey by Hongshen Chen, Xiaorui Liu, Dawei Yin, and Jiliang Tang, posted to arXiv in November 2017, takes stock of how deep learning had begun to reshape dialogue systems. Its abstract notes that recent progress was “overwhelmingly contributed by deep learning techniques,” and the paper organizes the field around a distinction that has become the standard mental map for conversational AI.
That distinction is between task-oriented systems and non-task-oriented systems. Task-oriented systems are built to help a user accomplish a specific goal, such as booking a flight, reserving a restaurant, or playing a song, and they typically work through a pipeline of understanding the user’s intent, tracking the state of the conversation, deciding what to do next, and generating a reply. Non-task-oriented systems, sometimes called open-domain or chit-chat systems, instead aim to carry on an engaging open-ended conversation with no particular task to complete. The survey reviews how deep learning was applied differently to each, from neural approaches to language understanding in the goal-driven case to generative and retrieval-based models for open conversation.
Written just before large language models began to blur this line, the paper is a useful snapshot of how researchers framed the problem in the era of dedicated dialogue architectures.
For a general reader, this survey is a good map of the conversational-AI landscape: it explains why a flight-booking bot and a casual chatbot were long treated as different engineering problems, a division that today’s general-purpose assistants increasingly try to collapse into one model.