TD-GIN: Token-level Dynamic Graph-Interactive Network for Joint Multiple Intent Detection and Slot Filling

Intent detection and slot filling are two main tasks for building a spoken language understanding (SLU) system. Currently, most work on SLU have focused on the single intent scenario, and paid less attention into the multi-intent scenario, which commonly exists in real-world scenarios. In addition, multi-intent SLU faces an unique challenges: how to effectively incorporate multiple intents information to guide the slot prediction. In this paper, we propose a Token-level Dynamic Graph-Interactive Network (TD-GIN) for joint multiple intent detection and slot filling, where we model the interaction between multiple intents and each token slot in a unified graph architecture. With graph interaction mechanism, our framework has the advantage to automatically extract the relevant intents information to guide each token slot prediction, making a fine-grained intent information integration for the token-level slot prediction. Experiments on two multi-intent datasets show that our model achieves the state-of-the-art performance and outperforms other previous methods by a large margin. Comprehensive analysis empirically shows that our framework successfully captures multiple relevant intents information to improve the SLU performance.