A Cognitively Inspired Multi-granularity Model Incorporating Label Information for Complex Long Text Classification

Because the abstracts contain complex information and the labels of abstracts do not contain information about categories, it is difficult for cognitive models to extract comprehensive features to match the corresponding labels. In this paper, a cognitively inspired multi-granularity model incorporating label information (LIMG) is proposed to solve these problems. Firstly, we use information of abstracts to give labels the actual semantics. It can improve the semantic representation of word embeddings. Secondly, the model uses the dual channel pooling convolutional neural network (DCP-CNN) and the timescale shrink gated recurrent units (TSGRU) to extract multi-granularity information of abstracts. One of the channels in DCP-CNN highlights the key content and the other is used for TSGRU to extract context-related features of abstracts. Finally, TSGRU adds a timescale to retain the long-term dependence by recuring the past information and a soft thresholding algorithm to realize the noise reduction. Experiments were carried out on four benchmark datasets: Arxiv Academic Paper Dataset (AAPD), Web of Science (WOS), Amazon Review and Yahoo! Answers. As compared to the baseline models, the accuracy is improved by up to 3.36%. On AAPD (54,840 abstracts) and WOS (46,985 abstracts) datasets, the micro-F1 score reached 75.62% and 81.68%, respectively. The results show that acquiring label semantics from abstracts can enhance text representations and multi-granularity feature extraction can inspire the cognitive system’s understanding of the complex information in abstracts.