A Joint Model based on CNN-LSTMs in Dialogue Understanding

Abstract

In Task-oriented Dialogue System, intent recognition and slot filling are two key subtasks of dialogue understanding (DU) module. Considering the strong relationship between intent and slots, this paper proposes an encoder-decoder architecture (using CNN-LSTMs) which based on attention mechanism to jointly model the two subtasks. Meanwhile, this paper also discusses the performance impact of the emitted slots information on the recognition of intent when jointly modeling. Our proposed model obtains 1.31% accuracy promotion on intent recognition and 0.90% gain on slot filling over the baseline model.