ContriSci: A BERT-Based Multitasking Deep Neural Architecture to Identify Contribution Statements from Research Papers

With the rapid growth of scientific literature, it is becoming increasingly difficult to identify scientific contribution from the deluge of research papers. Automatically identifying the specific contribution made in a research paper would help quicker comprehension of the work, faster literature survey, comparison with the related works, etc. Here in this work, we investigate methods to automatically extract the contribution statements from research articles. We design a multitask deep neural network leveraging section identification and citance classification of scientific statements to predict whether a given scientific statement specifies a contribution or not. In the long-run, we envisage to create a knowledge graph of scientific contributions for machine comprehension and more straightforward navigation of research contributions in a particular domain. Our approach achieves the best performance over earlier methods (a relative improvement of 8.08% in terms of \(F_1\) score) for contributing sentence identification over a dataset of Natural Language Processing (NLP) papers. We make our code available at here (https://github.com/ammaarahmad1999/Sem-Eval-2021-Task-A).