A joint framework for identifying the type and arguments of scientific contribution

Scientific contribution is typically embodiment of the value of a scientific publication, which reflects the inspiration, promotion, and improvement of the publication on existing theories or guiding practices. To analyze scientific contribution efficiently, in this paper, we introduce the task of automatically identifying the contribution type and their corresponding arguments. For this novel task, we first construct a new dataset SciContri by manually annotating the contribution type and argument information of 783 scientific articles. And we propose a joint framework named SciContriExt for the scientific contribution extraction task, i.e., classifying the contribution type and extracting corresponding fine-grained arguments. Our proposed framework adopts a deep learning classification model and a extraction model which extracts the contribution arguments from both token and span level. We jointly train the classification and extraction models by performing a weighted summation of the loss functions of the two models. Experiments show that our proposed model outperforms the state-of-the-art approaches on both contribution type classification and argument extraction tasks. The SciContri dataset will be released for future research.