Automated subject indexing is the task of assigning descriptors from a controlled vocabulary to bibliographic records. It is unusually difficult because the label space is enormous, each record has only a few positive labels, and most descriptors appear very rarely in training data. These characteristics make standard classification approaches struggle, especially for the long tail of infrequent terms.

The paper introduces a text-attentive graph reranking method to tackle this problem. Instead of predicting labels directly, it first generates candidate descriptors and then reranks them by considering both the text of the record and the graph structure of the vocabulary or document relationships. This lets the model use contextual and relational signals to promote the correct but rare labels over more common but incorrect ones.

By focusing on reranking rather than one-shot prediction, the approach offers a practical way to improve indexing accuracy in extreme multi-label settings. The authors position the work as a step toward more reliable automated indexing for digital libraries and bibliographic databases, where the long tail of specialized terminology is a persistent challenge.