The Text Classification Active Learning Using Transfer Learning Methods

In recent years, convolutional neural networks, or CNN have made it feasible to do text categorization with great accuracy. Zhang et al. used character-level decomposition and a CNN with reasonably deep layers to achieve high accuracy in text categorization. Real-world text-classification issues might be difficult to solve without a large enough labelled sample set. One solution is transfer studying, which use a network trained on any job as the seed network for another. While it is well-established that transfer learning works well for image identification, it is less clear what kinds of data it works best with and how successful it is for tasks in the processing of natural languages, such as document categorization. In this study, we first present a character-level convolutional neural network (CNN) that utilises a residual network's architecture to build a network with more advanced layers specifically for Japanese text categorization. We then show that by applying transfer learning across two specific datasets, we can enhance classification accuracy. We also present a method for measuring the success of transfer instruction and apply it to the assessment of our model.