Multi-sentence and multi-intent classification using RoBERTa and graph convolutional neural network

Citation context analysis (CCA) is a crucial task in scientometrics analysis, examining how and why researchers discuss each other's work in their papers (White 2004; Teufel et al. 2006). Despite its long history, traditional CCA frameworks often make simplistic assumptions, focusing on a single sentence for analysis. This approach overlooks important phenomena, as a research paper frequently contains an extensive discussion of referenced research paper that span multiple sentences and convey multiple intents simultaneously. Consequently, relying on a single citation sentence to capture intent falls short, missing out on many other nuanced discourses. When an author cites another research article, it is often with one or more intentions, such as extending an existing study, utilizing developed methods, data, or approaches, among others. Citations are also crucial for establishing the hypothesis and building the foundation of the literature. It aids in understanding the rationale behind the work & prior contributions in the field. Additionally, citations are helpful for authors to convey how the proposed approach is different from the existing work (Lauscher et al. 2021). In the existing study of citation context analysis, we observed that authors mainly focus on a hard-coded window-based citation context around citing sentences (Athar and Teufel 2012; Ravi et al. 2018; Vyas et al. 2020). This limited approach fails to provide a comprehensive understanding since there may be preceding or following sentences that discuss the same citing paper with different intentions. Consequently, a single-sentence context is insufficient for thorough citation analysis (Budi and Yaniasih 2023). To overcome the limitation posed by a single citation sentence or a fixed window of citation sentences, we advocate to consider multi-sentence-based citation context along with its potential intents (Jurgens et al. 2018). Consequently, we considered analysing a citation context window of variable length and respective intents. As one citation context window can have multiple intents, we considered intent classification as a multilabel classification task. This study can be useful in various scientometric analysis tasks, such as citation search, citation recommendation systems, and citation scoring systems. Examining multi-sentence-based citation context can pave the way for new opportunities and challenges within the scientific citation analysis task.