Leveraging Web Query Logs to Learn User Intent Via Bayesian Discrete Latent Variable Model

摘要: A key task in Spoken Language Understanding (SLU) is interpreting user intentions from speech utterances. This considered to be a classification problem with the goal of categorizing given utterance into one many semantic intent classes. Due substantial var, significant quantity labeled utterances needed build robust detection systems. In this paper, we approach as two-stage semi-supervised learning problem, which utilizes large number unlabeled queries collected internet seach engine click logs. We first capture underlying structure using bayesian latent feature model. then propagate onto obtain quality training data via graph summarization algorithm. Our improves compared comparison our baseline, uses standard model actual features.