Mixed-Model Text Classification Framework Considering the Practical Constraints

Authors:

Abstract
Text classification is one of the valuable tasks in the industry. Most of the classification tasks are accomplished by supervised learning which has a high demand for labeled data, especially for employing deep learning models. This demand has become the key development bottleneck for industry, especially for domain-specific data. In this paper, we propose a label generation process combining topic modeling and active learning. We review the state-of-art classification models and create a mixed-model approach that leverages the hierarchical structure of labels to maximize the information gained from limited labeled data. We integrate the approach into a text classification framework to help subject matter experts to label data and to train accurate classification models effectively.