Multi-task learning model for citation intent classification in scientific publications

Citations play a significant role in the evaluation of scientific literature and researchers. Citation intent analysis is essential for academic literature understanding. Meanwhile, it is useful for enriching semantic information representation for the citation intent classification task because of the rapid growth of publicly accessible full-text literature. However, some useful information that is readily available in citation context and facilitates citation intent analysis has not been fully explored. Furthermore, some deep learning models may not be able to learn relevant features effectively due to insufficient training samples of citation intent analysis tasks. Multi-task learning aims to exploit useful information between multiple tasks to help improve learning performance and exhibits promising results on many natural language processing tasks. In this paper, we propose a joint semantic representation model, which consists of pretrained language models and heterogeneous features of citation intent texts. Considering the correlation between citation intents, citation section and citation worthiness classification tasks, we build a multi-task citation classification framework with soft parameter sharing constraint and construct independent models for multiple tasks to improve the performance of citation intent classification. The experimental results demonstrate that the heterogeneous features and the multi-task framework with soft parameter sharing constraint proposed in this paper enhance the overall citation intent classification performance.