An Ensemble Deep Active Learning Method for Intent Classification

Abstract

Intent classification plays a primary and critical role in intelligent dialogue systems. However, faced with the lack of labeled data, the training of robust intent classification model is time-consuming and costly. Thanks to the powerful pre-trained model and active learning, it's possible to construct an integrated method to fulfill this task efficiently. Therefore, we propose an ensemble deep active learning method, which constructs intent classifier based on BERT and uses an ensemble sampling method to choose informative data for efficient training. Experimental results on both Chinese and English intent classification datasets suggest that the proposed ensemble deep active learning method can achieve state-of-the-art performance with less than half of the training data. In addition, the performance of the proposed method is stable and scalable for both datasets. In general, the proposed method shows substantial advantages in building intent classifier across different datasets.