Mf-cite: citation intent classification in scientific papers based on multi-feature fusion

Citations are crucial in scientific works. Citation analysis techniques help in literature search, citation recommendation, scientific assessment and other research works. Citation intent classification has proved to be useful as an important branch of citation analysis techniques, which categorizes the role that citations play in research works. However, scientific papers usually contain words that are difficult to understand and semantically uncertain, and they are generally more complex in structure and more rigorous in logic compared to ordinary texts. Previous studies have not fully explored to find useful information in citation contexts. Meanwhile, we find that classification label is more related to the part-of-speech feature of the word in the citation contexts, and usually nouns, verbs, adjectives and adverbs have a greater impact on the classification results. Therefore, in this paper, we propose a citation intent classification model called MF-Cite that combines citation context feature, WordNet-based semantic feature, and part-of-speech feature. It fuses them for scientific text representation, enabling the model to enhance the understanding of specialized domain terms and accurately comprehend the grammatical information of sentences. Experiments show that our method achieves more favorable results on the ACL-ARC and SciCite datasets. We conduct extensive ablations to analyze the effectiveness of the proposed model fusion method.