A Graph-to-Sequence Model for Joint Intent Detection and Slot Filling in Task-Oriented Dialogue Systems

Effectively decoding semantic frames in task-oriented dialogue systems remains a challenge, which typically includes intent detection and slot filling. Although RNN-based neural models show promising results by jointly learning of these two tasks, dominant RNNs are primarily focusing on modeling sequential dependencies. Rich graph structure information hidden in the dialogue context is seldomly explored. In this paper, we propose a novel Graph-to-Sequence model to tackle the spoken language understanding problem by modeling both temporal dependencies and structural information in a conversation. We introduce a new Graph Convolutional LSTM (GC-LSTM) encoder to learn the semantics contained in the dialogue dependency graph by incorporating a powerful graph convolutional operator. Our proposed GC-LSTM can not only capture the spatio-temporal semantic features in a dialogue, but also learn the co-occurrence relationship between intent detection and slot filling. Furthermore, a LSTM decoder is utilized to perform final decoding of both slot filling and intent detection, which mutually improves both tasks through global optimization. Experiments on benchmark ATIS and Snips datasets show that our model achieves state-of-the-art performance and outperforms existing models.