Research-Paper Classification And Reviewer Recommender System Based On NLP And Arabert Linguistic Model

Abstract

Peer-Review system is a tool for controlling the quality of scientific publications in scientific journals and conferences, with huge increase of submissions, the pressure on the review process has increased, which depends basically on human efforts in sorting papers according to its topic and assigning it to the most appropriate reviewer, which raised the need to develop the traditional systems to avoid slowness, bias, or lack of quality and to keep pace with the continuous scientific acceleration. Recently artificial intelligence techniques have been widely used in most fields, and many libraries were developed to analyzing Arabic texts, which facilitated their use for the development of current systems. In this paper, a multi-label classifier was created based on the deep learning pre-trained model for Arabic language, AraBERT, which was customized by training on data from peer-reviewed journal (Syrian Journal of Agricultural Research) after pre-processing using natural language processing techniques. This model has proven its ability to identify the topic using paper’s title and abstract only, then it was used in developing a reviewers recommender system, which is information retrieval model works on classifying and processing Arabic text semantically based on the AraBERT Linguistic model, this leads to measure the similarity ratio between a paper and collection of publications by reviewers in the same subject to provide top ten suggestions of reviewers with the closest specialization to the paper. The system also provides additional suggestions as new reviewer among previously published articles. This system was tested on data of the journal, and the classification model reached the values 86.5%, 91.8%, 92.4%, and 91.4% for each of the Accuracy, F1-score, Precision, and Recall, respectively. It was also found that 85% of the reviewers suggested by the system are fully suitable for review process, according to the assessment of the journal’s publishing experts, and by applying the recall@10 scale, it showed that 49% of the suggestions submitted by the system matched the actual reviewers. The importance of this system appears as an effective tool to assist editors in scientific journals and conferences in classifying research and selecting the most appropriate reviewers quickly, effectively and without bias.