Category:

Vlog

Duration: 20:11

1,797 views

16

Applications with voice interfaces often work in a "one-shot" manner: the user makes a request, the system carries out the request (if it has been understood) and then the dialogue state is reset. Spoken interactions between humans, on the other hand, can involve a incremental process of establishing and negotiating the participants' goals. In this talk I will discuss statistical models that attempt to reflect the complex nature of human dialogue. We will discuss how recurrent neural networks can detect the user's intentions and track the evolving dialogue state. Evaluations on data from the Dialogue State Tracking Challenges show that these RNN-based models can achieve state-of-the-art performance.