Another quarter, another class project on task oriented dialogue agents! This quarter I completed a paper studying the details of Intent Tracking within UW Graduate Machine Learning - CSE 546. And this time around, the paper won the award for best paper in the class!
What exactly is the task we are trying to solve in task-oriented dialogue modeling? Is there a different goal for academic research vs industry application? First, note that I am already segmenting into chit-chat vs. goal-oriented dialogue. Despite this, even with the realm of goal-oriented chat, there seems to be different categories.
Problem: Training function is imprecise (i.e. BLEU score)
Thoughts on some of the latest and greatest as of mid-2018.
Once again building on End-to-End Task Oriented Dialog. Now with a Transformer: