Abstract: The automatic identification of citation intentions in academic literature contributes to a deeper understanding of academic paper content and facilitates the development of a more equitable framework for scientific research evaluation. To improve the accuracy and generalizability of citation intention recognition in low-resource scenarios, this study proposes a novel framework that integrates prompt learning and multi-task learning, First, the main task of automatic citation intent recognition is jointly trained with two auxiliary tasks: citation value identification and citation section identification. All three tasks are unified within a P-tuning-based prompt learning framework. In this setup, discrete prompt tokens are first transformed into continuous vectors by a multi-layer perceptron. These vectors are then combined with the input text to form the input sequence, effectively converting the classification task into a cloze-style prediction. Furthermore, the label word set is expanded, and a weighted averaging method is applied to select the label with the highest score as the final prediction. Finally, a soft parameter sharing mechanism is employed to achieve collaborative optimization across the three tasks for citation intent recognition. Experimental results on two publicly available datasets, ACL-ARC and SciCite, demonstrate that the proposed model significantly outperforms baseline and other prompt learning models under varying sample sizes. Additionally, the two auxiliary tasks effectively improve the accuracy and generalizability of citation intent recognition.

Key words: Citation intent identification, Prompt learning, Multi-task learning, Text classification, SciBERT