Deep Active Learning for Text Classification

In recent years, Active Learning (AL) has been applied in the domain of text classification successfully. However, traditional methods need researchers to pay attention to feature extraction of datasets and different features will influence the final accuracy seriously. In this paper, we propose a new method that uses Recurrent Neutral Network (RNN) as the acquisition function in Active Learning called Deep Active Learning (DAL). For DAL, there is no need to consider how to extract features because RNN can use its internal state to process sequences of inputs. We have proved that DAL can achieve the accuracy that cannot be reached by traditional Active Learning methods when dealing with text classification. What's more, DAL can decrease the need of the great number of labeled instances for Deep Learning (DL). At the same time, we design a strategy to distribute label work to different workers. We have proved by using a proper batch size of instance, we can save much time but not decrease the model's accuracy. Based on this, we provide batch of instances for different workers and the size of batch is determined by worker's ability and scale of dataset, meanwhile, it can be updated with the performance of the workers.