Preserving Privacy in Spoken Language Databases

Goal-oriented spoken dialog systems aim to identify intents of humans, expressed in natural language, and take actions accordingly, to satisfy their requests. State-of-the-art data-driven spoken dialog systems are trained using large amounts of task data which is usually transcribed and then labeled by humans, a very expensive and laborious process. Hence sharing and reuse of this data has extreme importance for research and development of spoken language processing systems. On the other hand these utterances may include confidential personal information about the speakers, such as the social security numbers or credit card numbers. In this paper, we describe data sanitization approaches for natural language utterances to protect the privacy of the speakers. The challenge in sanitization is ensuring that the performance of the spoken dialog system models trained using the sanitized data is as good as the ones before the sanitization. We show that, by hiding task-dependent named entities we can preserve the privacy of the speakers, and still achieve a comparable accuracy.