Abstract: Building universal dialogue systems that can seamlessly operate across multiple domains/APIs and can generalize to new ones with minimal supervision and low maintenance is a critical challenge. Recent works have leveraged natural language descriptions for schema elements to build such systems. However, descriptions only provide indirect supervision for downstream tasks, while still requiring effort to construct. In this work, we propose Show, Don't Tell, which uses a short labeled example dialogue to show the semantics of a schema rather than telling the model about the schema elements via descriptions. While requiring similar effort from service developers, we show that using short examples as schema representations with large language models results in stronger performance and better generalization on two popular dialogue state tracking benchmarks: the Schema-Guided Dialogue (SGD) dataset and the MultiWoZ leave-one-out benchmark.
Paper Type: short
0 Replies
Loading