Citation Context Analysis

Citation counting treats every citation as an equal, anonymous vote and asks only how many there are. Citation context analysis reads the text around each citation and asks what kind of citation it is: central or perfunctory, agreeing or disagreeing, borrowing a method or merely mentioning a concept. This distinction matters because Moravcsik and Murugesan found that a large share of citations are perfunctory, so two works with equal counts can have very different real influence. Context analysis recovers that lost information at the cost of much greater effort per citation.

They are distinct dimensions. Function describes the rhetorical role a citation plays, for example whether it supplies a concept or a technique, or whether it is essential to the argument or incidental. Sentiment or polarity describes the citing author's stance toward the cited work, positive, neutral, or negative. The classic scheme captures both: confirmative versus negational is essentially polarity, while conceptual versus operational and organic versus perfunctory are functional. Keeping the two dimensions separate avoids the common error of reading a method-borrowing citation as an endorsement, or a neutral mention as influence.

Yes, citation-function and citation-sentiment classification are active computational tasks, using supervised classifiers and language models trained on annotated citation contexts. However, automation does not remove the need for validated ground truth: classifiers are trained and evaluated against human-coded labels, and their performance varies by category, with subtle or rare functions (such as negational citations) being hardest. Whether coding is manual or automatic, the standard of credibility is demonstrated reliability, reported as inter-coder agreement for human coding and as evaluation metrics against a held-out human-labeled set for classifiers.