Discriminative methods for statistical spoken dialogue systems

摘要

Dialogue promises a natural and effective method for users to interact with and obtain information from computer systems. Statistical spoken dialogue systems are able to disambiguate in the presence of errors by maintaining probability distributions over what they believe to be the state of a dialogue. However, traditionally these distributions have been derived using generative models, which do not directly optimise for the criterion of interest and cannot easily exploit arbitrary information that may potentially be useful. This thesis presents how discriminative methods can overcome these problems in Spoken Language Understanding (SLU) and Dialogue State Tracking (DST).A robust method for SLU is proposed, based on features extracted from the full posterior distribution of recognition hypotheses encoded in the form of word confusion networks. This method uses discriminative classifiers, trained on unaligned input/output pairs. Performance is evaluated on both an off-line corpus, and on-line in a live user trial. It is shown that a statistical discriminative approach to SLU operating on the full posterior ASR output distribution can substantially improve performance in terms of both accuracy and overall dialogue reward. Furthermore, additional gains can be obtained by incorporating features from the system's output.For DST, a new word-based tracking method is presented that maps directly from the speech recognition results to the dialogue state without using an explicit semantic decoder. The method is based on a recurrent neural network structure that is capable of generalising to unseen dialogue state hypotheses, and requires very little feature engineering. The method is evaluated in the second and third Dialog State Tracking Challenges, as well as in a live user trial. The results demonstrate consistently high performance across all of the off-line metrics and a substantial increase in the quality of the dialogues in the live trial. The proposed method is shown to be readily applied to expanding dialogue domains, by exploiting robust features and a new method for online unsupervised adaptation. It is shown how the neural network structure can be adapted to output structured joint distributions, giving an improvement over estimating the dialogue state as a product of marginal distributions.