Semi-supervised learning for spoken language understanding semantic role labeling
Gökhan Tür, Dilek Zeynep Hakkani-Tür, Ananlada Chotimongkol · 2005
In a goal-oriented spoken dialog system, the major aim of language understanding is to classify utterances into one or more of the pre-defined intents and extract the associated named entities. Typically, the intents are designed by a human expert according to the application domain. Furthermore, these systems are trained using large amounts of data manually labeled using an already prepared labeling guide. In this paper, we propose a semi-supervised spoken language understanding approach based on the task-independent semantic role labeling of the utterances. The goal is to extract the predicates and the associated arguments from spoken language by using semantic role labeling and determine the intents based on these predicate/argument pairs. We propose an iterative approach using the automatically labeled utterances with semantic roles as the seed training data for intent classification. We have evaluated this understanding approach using two AT&T spoken dialog system applications used for customer care. We have shown that the semantic parses obtained without using any syntactically or semantically labeled in-domain data can represent the semantic intents without a need for manual intent and labeling guide design and labeling phases. Using this approach on automatic speech recognizer transcriptions, for both applications, we have achieved the 86.5% of the performance of a classification model trained with thousands of labeled utterances