Quality Classification of Scientific Publications Using Hybrid Summarization Model

In this paper (Note that the dataset and code to reproduce the results can be accessed at the following URL: https://github.com/slab-itu/hsm), we intend to assess the quality of scientific publications by measuring the relationship between full text papers with that of their abstracts. A hybrid summarization model is proposed that combines text summarization and information retrieval (IR) techniques to classify scientific papers into different ranks based on their abstract correctness. Using the proposed model, we study the relationship between a correctly written abstract (in accordance with full-text) and the scholarly influence of scientific publications. The proposed supervised machine learning model is deployed on 460 full-text publications - randomly downloaded from Social Science Research Network (SSRN). In order to quantify the scholarly influence of publications, a composite score provided by SSRN is used that combines usage indicators along with citation counts. This score is then used to label the publications into high and low ranks. The results determine that the papers having abstracts in accordance with full text also show high scholarly rank with an encouraging accuracy of 73.91%. Finally, 0.701 Area Under the Curve (AUC) for receiver-operating characteristic is achieved that outperforms the traditional IR and summarization models with AUC of 0. 536 and 0.58 respectively. Overall our findings suggest that a correctly written abstract in accordance to its full text have a high probability to attract more social usage and citations and vice versa.