End to End Spoken Language Understanding Using Partial Disentangled Slot Embedding

Spoken language understanding (SLU) has switched from pipeline approaches to end-to-end (E2E) ones recently. For most E2E approaches, neural networks are adopted to extract embeddings from the audio signals directly for final intents prediction. In this paper, we explore this method for intent classification on Fluent Speech Commands (FSC) dataset, where intents are formed as combinations of three slots (action, object, and location). The information of different slots will be entangled with each other in the extracted embeddings, which sometimes brings about errors in the prediction of the current slot. To address this problem, we propose partial disentangled slot embedding (PDSE) method through adversarial training. Results show that the proposed method can achieve an error rate of 0.53%, which outperforms the baseline with over 35.3% error rate reduction.