A Joint Model of Natural Language Understanding for Human-Computer Conversation in IoT

Figure 2

Framework of our bidirectional information transfer method. The framework is comprised of three main parts: the shared encoder, the decoder slot filling, and the intent detection. The type-aware method works on the bottom layer of the two decoders, respectively. The bidirectional information transmission module is utilized between the two decoders, which propagates tag information from top layers of the two decoders. For example, the probability distribution of intent label is transmitted to the bottom layer of slot filling decoder. And the probability distribution of slot labels is transmitted to the bottom layer of intent detection decoder simultaneously.