Wide and Deep Learning for Spoken Language Understanding

Anonymous

Keywords: spoken language understanding, intent classification, slot filling, joint model

TL;DR: We build a joint model for intent classification and slot filling using constructed wide lexical features and deep semantic features.

Abstract: Spoken language understanding (SLU) is one of the essential parts in smart voice assistants, which typically includes intent classification (IC) and slot filling (SF) tasks to interpret user utterances. Deep models jointly trained for the two tasks show more promising results compared with single-task models. However, these models always learn semantic representations for tokens and utterances but ignore their lexical information. Although they can generalize better to unseen tokens and utterances from low-dimensional dense semantic features, they also suffer from over-generalization when training data is limited. On the other hand, sparse lexical features such as word ngrams are good to memorize existing data correlations but fail for generalization. In this paper, we propose an approach leveraging lexical and semantic features to jointly learn IC and SF. The aim is to combine the benefits of memorization and generalization for SLU. Evaluating on a couple of domains from a large-scale smart voice assistant, results show our approach significantly improves IC and SF compared with several strong baselines.