Dialogues with Deep Learning
Problem being addressed
Can machine learning hold up a conversation?
In this research paper, a deep reinforcement learning approach is taken to the problem of dialogue generation. Typically, chat bot dialogues tend to be short term generated with the focus on the next line of dialogue, rather than on a whole conversation. This paper attempts to address the question of generating a longer, more rich, dialogue. Three key properties are zoomed in on: information content, coherence and ease of answering; and the dialogue is represented as an alternating sequence between two agents in which information flows between them. Metrics are created for each of these properties; and a supervised and curriculum learning approach is taken for training an AI to perform against these metrics.
Advantages of this solution
This research seems to lead to better dialogues that have a more human "smooth" feel to them. The result is an improved experience when talking to a chat bot, for example.
Solution originally applied in these industries
Possible New Application of the Work
Online education programs do not normally use chat bots because they don't have the nuance of language required to hold up a student conversation. However, this approach seems to enhance the conversation, and could make for an interesting school teacher AI.
It would be interesting to see if an AI could create a movie script with real sounding dialogue. This could prove very useful for soap opera writing, for example, where the same themes are played out repeatedly.
Remote diagnostics are becoming important healthcare, as is managing patient flow in hospitals. An automated self service dialoguing approach could make it easier for a pre-diagnostic to occur, before a patient sees a medical practitioner.
Source URL: #############