The coreference annotation project is being carried out at Raytheon BBN Technologies under the supervision of Ralph Weischedel and Lance Ramshaw
The goal of OntoNotes coreference annotation and modeling is to fill in the coreference portion of the shallow semantic understanding of the text that OntoNotes is targeting. For example, in "She had a good suggestion and it was unanimously accepted", we mark a case of IDENT coreference (identical reference) between "a good suggestion" and "it", which then allows correct interpretation of the subject argument of the "accepted" predicate.
Names, nominal mentions, and pronouns can be marked as coreferent. Verbs that are coreferenced with a noun phrase can also be marked as IDENT; for example "grew" and "the strong growth" would be linked in the following case: "Sales of passenger cars grew 22%. The strong growth followed year-to-year increases." In addition, in 'pro-drop' languages like Chinese, coreference annotation can be applied to a "*T*" element taken from the Treebank parse which serves as a placeholder for the missing pronoun.
In order to keep the annotation feasible at high agreement levels, only intra-document anaphoric coreference is being marked. Furthermore, while annotation is not limited to any fixed list of target entity types, noun phrases that are generic, underspecified, or abstract are not annotated.
Attributive NPs are not annotated as coreference because the meaning in such cases can be more appropriately taken from other elements in the text. For example, in "New York is a large city", the connection between New York and the attributive NP "a large city" comes from the meaning of the copula "is". Similarly, in "Mary calls New York heaven", the connection comes from the meaning of the verb "call". Thus these cases are not marked as IDENT coreference.
Appositive constructions are marked with special labels. For example, in "Washington, the capital city, is on the East coast", we annotate an appositive link between Washington (marked as HEAD) and "the capital city" (marked as ATTRIBUTE). The intended semantic connection can then be filled in by supplying the implicit copula.