{Anonymous}

Abstract

Intent detection and slot filling are two important basic tasks in natural language understanding. Actually, there are multiple intents in an utterance. How to map different intents to corresponding slot becomes a new challenge for recent research. Existing models solve this problem by using neural layers to adaptively capture related intent information for each slot, which the process of intent selection is not clear enough. It is observed that there is strong consistency between intents and topics of a sentence, thus we exploit topic information for joint intent detection and slot filling via a topic fusion mechanism, where token-level topic information take the place of intent information to guide slot prediction. In addition, sentence-level topic information is also utilized to enhance the intent detection. Experiment results show explicit improvements on two public datasets, where provide 4.8% improvement in sentence accuracy on MixATIS and 0.7% improvement in intent detection on MixSNIPS.