Contextual memory bandit for pro-active dialog engagement

An objective of pro-activity in dialog systems is to enhance the usability of conversational agents by enabling them to initiate conversation on their own. While dialog systems have become increasingly popular during the last couple of years, current task oriented dialog systems are still mainly reactive and users tend to initiate conversations. In this paper, we propose to introduce the paradigm of contextual bandits as framework for pro-active dialog systems. Contextual bandits have been the model of choice for the problem of reward maximization with partial feedback since they fit well to the task description. As a second contribution, we introduce and explore the notion of memory into this paradigm. We propose two differentiable memory models that act as parts of the parametric reward estimation function. The first one, Convolutional Selective Memory Networks, uses a selection of past interactions as part of the decision support. The second model, called Contextual Attentive Memory Network, implements a differentiable attention mechanism over the past interactions of the agent. The goal is to generalize the classic model of contextual bandits to settings where temporal information needs to be incorporated and leveraged in a learnable manner. Finally, we illustrate the usability and performance of our model for building a pro-active mobile assistant through an extensive set of experiments.