口语对话系统是最自然的人机交互界面之一。然而语音识别和口语理解模块带来的级联错误会对用户体验造成很大影响,在嘈杂的环境中更为严重。对话状态跟踪器可根据对话的上下文和可观测到的语音识别、理解结果对各个回合的对话状态做出估计。因此,提出一种由数据驱动基于鉴别式模型的对话状态追踪方法,能够处理更大规模的特征集,特征函数依赖于可观测的全部N-best结果。通过在真实语音数据集上进行评测,实验结果表明,该方法比单纯使用1-best结果的基线系统具有更强的性能。
Spoken dialog system is a natural and intuitive human- computer interface. But the errors resulted from automatic speech recognition and spoken lanuage understanding will harm user experience,which is more serious in noisy environment. Dialog state trackers make estimation of the current dialog states by observations in dialog history. In this paper we propose a data- driven dialog state tracking method by discriminative modeling. This method can handle large feature sets which can utilize full observed N- best results. The proposed method show better performance than baseline operating on 1-best results.