Ensembling Approaches to Citation Function Classification and Important Citation Screening

Compared to feature engineering, deep learning approaches for citation context analysis have yet fully leveraged the myriad of design options for modeling in-text citation, citation sentence, and citation context. In fact, no single modeling option universally excels on all citation function classes or annotation schemes, which implies the untapped potential for synergizing diverse modeling approaches to further elevate the performance of citation context analysis. Motivated by this insight, the current paper undertook a systematic exploration of ensemble methods for citation context analysis. To achieve a better diverse set of base classifiers, we delved into three sources of classifier diversity, incorporated five diversity measures, and introduced two novel diversity re-ranking methods. Then, we conducted a comprehensive examination of both voting and stacking approaches for constructing classifier ensembles. We also proposed a novel weighting method that considers each individual classifier’s performance, resulting in superior voting outcomes. While being simple, voting approaches faced significant challenges in determining the optimal number of base classifiers for combination. Several strategies have been proposed to address this limitation, including meta-classification on base classifiers and utilising deeper ensemble architectures. The latter involved hierarchical voting on a filtered set of meta-classifiers and stacked meta-classification. All proposed methods demonstrate state-of-the-art results on, with the best performances achieving more than 5% and 4% improvements on the 11-class and 6-class schemes of citation function classification and by 3% on important citation screening. The promising empirical results validated the potential of our ensembling approaches for citation context analysis.