International Journal of Information Science and Management (IJISM)

Abstract The pattern of the in-text citation is significant in determining its value. This study focuses on advancing multi-task learning (MTL) techniques for citation-context classification: sentiment, role source, and function. Specifically, the research explores the integration of shared-trunk, cross-stitched, and shared-private networks with deep learning (DL) models, namely a convolutional neural network (CNN), a long short-term memory (LSTM), and a bidirectional LSTM. The dataset comprises 8566 sentences that include in-text citations of journal articles originating from Indonesia. To address the issue of unbalanced data, distinct loss weights were assigned to each output. Additionally, Bayesian optimization was employed to determine the hyperparameter sizes. The findings of the optimization process confirm that the learning rate is the most influential hyperparameter. The architectural complexity of the three MTL types varies, but this does not affect time efficiency, which is heavily influenced by the DL model employed. By achieving an F1 score of 0.87 for role source, 0.90 for function, and 0.81 for sentiment classification, the shared-private CNN model surpasses the current state-of-the-art model. The proposed model may improve citation analysis, making it beneficial for developing citation recommendation systems and conducting science evaluations.