Learning Natural Language Understanding Systems from Unaligned Labels for Voice Command in Smart Homes

Abstract

Voice command smart home systems have become a target for the industry to provide more natural human computer interaction. To interpret voice command, systems must be able to extract the meaning from natural language; this task is called Natural Language Understanding (NLU). Modern NLU is based on statistical models which are trained on data. However, a current limitation of most NLU statistical models is the dependence on large amount of textual data aligned with target semantic labels. This is highly time-consuming. Moreover, they require training several separate models for predicting intents, slot-labels and slot-values. In this paper, we propose to use a sequence-to-sequence neural architecture to train NLU models which do not need aligned data and can jointly learn the intent, slot-label and slot-value prediction tasks. This approach has been evaluated both on a voice command dataset we acquired for the purpose of the study as well as on a publicly available dataset. The experiments show that a single model learned on unaligned data is competitive with state-of-the-art models which depend on aligned data.