A transfer learning approach to interdisciplinary document classification with keyword-based explanation

With the exponential increase of interdisciplinary research, identifying accurate disciplines of scientific documents has become increasingly important in various research management tasks. Interdisciplinary classification, which classifies documents into multiple disciplines, is essential for multidisciplinary research development. Due to the scarcity of labeled multidiscipline data, existing scientific document classification methods can't solve the interdisciplinary issue. Most of them also have the problem of explainability with curtly providing classification results. This study proposes an explainable transfer-learning-based classification method for interdisciplinary documents. First, we trained a single-discipline classification model using existing labeled single-discipline documents. Then, we transfer the knowledge learned from single-discipline classification to interdisciplinary classification to address the scarcity of labeled interdisciplinary data. We also added discipline co-occurrence information into our proposed model. Finally, we obtained our final model by training the transferred model with interdisciplinary data. In addition, keyword-based explanations for classifying texts are provided by employing layer-wise relevance propagation. Experiments on real-life NSFC data show the effectiveness of the proposed method, which can promote interdisciplinary development by constructing an efficient and fair classification for interdisciplinary review systems.