Online Supervised Learning of Non-understanding Recovery Policies

Spoken dialog systems typically use a limited number of nonunderstanding recovery strategies and simple heuristic policies to engage them (e.g. first ask user to repeat, then give help, then transfer to an operator). We propose a supervised, online method for learning a non-understanding recovery policy over a large set of recovery strategies. The approach consists of two steps: first, we construct runtime estimates for the likelihood of success of each recovery strategy, and then we use these estimates to construct a policy. An experiment with a publicly available spoken dialog system shows that the learned policy produced a 12.5% relative improvement in the non-understanding recovery rate.