mentionranking

Mention ranking is a task in natural language processing and information extraction that involves ordering candidate mentions of an entity by how likely they are to refer to the same entity in a given context. It is commonly used in two related areas: coreference resolution, where the goal is to identify which mentions refer to the same real-world entity, and entity linking, where mentions are connected to entries in a knowledge base or ontology.

In coreference resolution, the problem is to select the best antecedent for a given referring expression by ranking candidate antecedents with respect to a set of features. Ranking can be based on lexical, syntactic, semantic, and discourse clues, as well as distance between mentions and agreement in gender and number. The highest-ranked candidate is chosen as the antecedent, or a global ranking over candidates is used to form a coherent set of coreferent mentions.

In entity linking, mention ranking is used to link a text span to a knowledge base entry. After generating candidate entities for a span, a ranking model scores each candidate by contextual compatibility, prior probability, and coherence with other linked mentions. The top-ranked candidate is selected for linking.

Methods range from traditional feature-based ranking models, such as logistic regression or SVM-based rankers, to neural approaches using bi-encoders, cross-encoders, or attention-based architectures. Training data comes from annotated corpora for coreference and entity linking, with evaluation focusing on precision, recall, F1, and ranking-specific metrics like mean reciprocal rank or precision@k.

Applications include information extraction, question answering, chatbots, and knowledge base population. Challenges involve pronouns, long-range dependencies, cross-document coherence, multilingual settings, and scalability to large text collections.