Real-Time Skill Discovery in Intelligent Virtual Assistants

Solution to many real-world problems often involve the use of expert-level knowledge from various specializations. Such interdisciplinary problems are usually divided into tasks which are then assigned to a set of bots, each specialized in a particular skill. Supervised selection of the right bot each time is cumbersome and not scalable. Hence there is a need for an AI system that identifies the type of task and assigns it to a suitably trained bot. Challenges arise in non-stationary environments when the cost of choosing different bots vary or the bots themselves might evolve in their skills. In this paper, as in Conversational AI, a number of bots are at our disposal, each of which is trained to handle (i.e., answer) a specific type of question in a conversation. We develop a meta-algorithm that learns about capabilities (Skill Discovery) of the available bots in real-time and appropriately selects a relevant bot for the question at hand. We present contextual bandits as a solution in this setting and introduce gradual finetuning of query information to improve Skill Discovery. Using two popular datasets from conversational AI: CoQA and SQuAD, we show promising results of our method on non-stationary environments.