Multi-label Classification of Long Text Based on Key-Sentences Extraction

Most existing works on multi-label classification of long text task will perform text truncation preprocessing, which leads to the loss of label-related global feature information. Some approaches that split an entire text into multiple segments for feature extracting, which generates noise features of irrelevant segments. To address these issues, we introduce key-sentences extraction task with semi-supervised learning to quickly distinguish relevant segments, which added to multi-label classification task to form a multi-task learning framework. The key-sentences extraction task can capture global information and filter irrelevant information to improve multi-label prediction. In addition, we apply sentence distribution and multi-label attention mechanism to improve the efficiency of our model. Experimental results on real-world datasets demonstrate that our proposed model achieves significant and consistent improvements compared with other state-of-the-art baselines.