Continuing on the thoughts described in Part 1, we now describe two trends around dialogue and NLP research. In the previous post, we discussed how data strategy is playing an increasingly important role in the development of modern machine learning models. How that plays out in more detail will be discussed in the future, but today we will look further into the development of dialogue in the past few months as well as other observations.
This is nominally a review of trends from the EMNLP 2020, but also serves as the end of year review. As this (crazy) year comes to a close, we take a moment to reflect on what has happened in the world of NLP and specifically in relation to Dialogue Modeling and Natural Language Understanding. From my (limited) perspective, there are four trends which I noticed, including increased research into (1) data methods rather than models and (2) data efficiency techniques. In the follow-up, we discuss observations around (3) limitations in dialogue state tracking and (4) convergence around similar ideas.
Let’s dive into each one in more detail:
Compared to typical goal-oriented dialogue systems, interactive dialogue agents not only aim to solve the task at hand, but also engage with the user by asking questions. Questions can be used to push the conversation forward or to spark some new ideas, but for now we will focus on the use of questions to clarify understanding. Clarification requests, as they are referred to in the academic literature, come in many forms, but the key issues to solve are when to ask such questions and in what format. Asking questions too often, or at inappropriate times, causes the conversation to feel disjointed or annoying. Asking the wrong type of question causes the dialogue agent to seem incoherent or useless.
Performing active learning on data annotation is to decide when the model should query the expert annotator for more samples. Note the parallel with dialogue, which is to decide when the agent should ask a clarification question to the customer for more details on their intent. Such a policy can be obtained through static data with basic supervised learning or in a more interactive manner through imitation learning using algorithms such as DAgger (Ross 2011).
A key sub-issue in designing conversational agents is being able to reliably calculate uncertainty over the model’s beliefs. In doing so, the model would be able to recognize when it does not understand something, and appropriately ask for clarification. Thus, we can imagine the output of an uncertainty model feeding into a dialogue policy manager which then decides to either retrieve an answer from the knowledge base when it feels fairly certain it knows what the customer wants, or to ask a follow-up question when it is unsure. From a information-theory point of view, this can be seen as a model which asks questions to minimize entropy until it reaches a certain threshold, at which point it will return an answer. Beyond improving model behavior, measuring uncertainty also gives a view into how the model is thinking for improved debugging and enhanced interpretability.