Not all citations are equal: teaching AI to read why scientists cite
Citations are the currency of science. They show how ideas spread, and they feed the indicators used to evaluate papers, researchers and institutions. But a citation can mean very different things: a paper may be cited as background, because its method was used, or to compare results. Counting every citation the same way hides that difference.
We built a compact AI model, based on generative adversarial networks (GANs), that reads the sentence around a citation and identifies its intent. It matches state-of-the-art accuracy with far fewer parameters, so it is cheaper to run at scale. Applying it to a large network of arXiv papers, we found that which papers look most central changes significantly depending on which kinds of citations are counted, especially for papers that bridge different fields.
Why it matters. Rankings of papers and researchers often rely on raw citation counts. This work shows that those rankings depend on why papers are cited, and offers an efficient tool to take citation intent into account: a step toward fairer, more meaningful research evaluation.
