Active Learning Approach for Intent Classification in Portuguese Language Conversations

摘要

Intent classification from conversations is a challenging task especially if messages are collected from real data since messages can contain several grammatical errors, outliers, language slang, and a lot of categories. Most of the intent classification methods use supervised learning approaches and do not consider the lack of labeled data, because supervised learning approach requires a large amount of labeled samples during its training process. In this article, we propose to reduce the sample scale to be labeled and maintain a desired classification effectiveness. Our method is based on active learning to minimizes amount of labeled data required and a convolutional neural network that obtains effective vector representations from BERT to perform accurate classification of messages. Experimental results on a large Brazilian Portuguese corpus suggest that the proposed method can achieve improvements with more than half of the training data, and accurate results with less than half of data in small dataset like as ATIS.