Research On The Classification Technology Of Academic Papers Based On Image And Text Information Processing

Document classification is an important field of natural language processing(NLP).In recent years,huge amount of academic papers are published every year and researchers are eager to get help from intelligent paper classification methods.However existing paper classification models are mostly coarse-grained,and the research on fine-grained academic paper classification is still very limited.In order to solve this problem,we need to pay attention to two characteristics of fine-grained academic papers.One is sample imbalance or data insufficiency.Number of papers in the lower level sub-fields inevitably becomes less.Meanwhile,emerging sub-fields with new discoveries have few papers.The other is that academic papers actually contain multiple modes of information(image and text).The texts and images provide complementary information in classification.If utilized together properly,they can be beneficial to the performance in fine-grained paper classification.To fill the gaps in the dataset of multimodal academic papers,we designed and collected the PaperNet-Dataset that consists of multi-modal data(texts and figures).This datasset contains hierarchical categories of papers in the fields of computer vision(CV)and NLP,2 coarse-grained and 20 fine-grained(7 in CV and 13 in NLP).We also propose a new multimodal model for fine-grained academic paper classification.MobileNetV2 model is used to extract image features,and GRU model is used to extract text features.Then the extracted features are processed by the variable pooling operator for dimensionality reduction.After that,we add them according to different weight ratios to obtain the fusion feature vector,and then get the classification results through the classifier.After comparative tests,we found that the classification accuracy of our proposed method has been improved to varying degrees in both coarse-grained and fine-grained scenarios,especially in the case of a limited number of samples.