Robust Dialogue State Tracking with Weak Supervision and Sparse Data

Abstract Generalizing dialogue state tracking (DST) to new data is especially challenging due the strong reliance on abundant and fine-grained supervision during training. Sample sparsity, distributional shift, occurrence of concepts topics frequently lead severe performance degradation inference. In this paper we propose a training strategy build extractive DST models without need for manual span labels. Two novel input-level dropout methods mitigate negative impact sample sparsity. We model architecture with unified encoder that supports value as well slot independence by leveraging attention mechanism. combine strengths triple copy matching benefit from complementary predictions violating principle ontology independence. Our experiments demonstrate an can be trained strategies improve robustness towards concepts, topics, leading state-of-the-art range benchmarks. further highlight our model’s ability effectively learn non-dialogue data.