相互参照の解像度

Texts often introduce an entity by name and refer to it later through pronouns or descriptions. Coreference resolution groups mentions that identify the same entity in the discourse. In Maria said she was tired, a system can investigate the link between Maria and she; a different context may require a different interpretation. Resolving a pronoun back to an earlier referring expression is a familiar form of pronominal anaphora. Not every relation between expressions is identity coreference, so the task should be defined carefully. Winograd-schema examples illustrate a difficulty beyond grammatical agreement. In the trophy and suitcase example, knowing what cannot fit when it is too large helps interpret the pronoun. Grammar alone permits both noun phrases as candidates. Incorrect mention links can also affect a summary or answer: confusing a company with its product can change who performed an action. Influential span-ranking models score candidate mentions and potential earlier antecedents, constructing clusters through predicted links. This is a description of a model family, not every possible coreference algorithm. Distance, agreement, and contextual information can contribute evidence, without making a surface cue definitive. Entity linking adds another task: associating a textual mention with an entry in a knowledge base. Grouping Maria and she within a document does not necessarily identify the relevant external database record. Evaluation should also check biases. WinoBias contains occupation-based, Winograd-style examples that compare pro-stereotypical and anti-stereotypical pronoun associations. It examines whether stereotype-consistent choices receive an advantage in the evaluated systems rather than assuming occupation determines gender.