Unsupervised Learning of KB Queries in Task-Oriented Dialogs

Task-oriented dialog (TOD) systems often need to formulate knowledge base (KB) queries corresponding the user intent and use query results generate system responses. Existing approaches require datasets explicitly annotate these KB -- annotations can be time consuming, expensive. In response, we define novel problems of predicting training agent, without explicit annotation. For prediction, propose a reinforcement learning (RL) baseline, which rewards generation those whose cover entities mentioned in subsequent dialog. Further analysis reveals that correlation among attributes significantly confuse memory augmented policy optimization (MAPO), an existing state art RL agent. To address this, improve MAPO baseline with simple but important modifications suited our task. train full TOD for setting, pipelined approach: it independently predicts when make (query position predictor), then at predicted uses (next response predictor). Overall, work proposes first solutions problem, highlights research challenges