Is NER required to resolve Coreference?

Question

... or enough gender information? More specifically, I am interested to know if I can reduce the number of models loaded by NLP Stanford Core in order to extract the basic data. I'm not interested in the actual name recognition.

thanks

Answer 1

According to the EMNLP document, which describes the coref system packaged with Stanford CoreNLP, object name tags are used only in the following coref annotations: precise constructions, easy matching of the head and pronoun (Raghunathan et al., 2010) .

You can specify what passes to use the dcoref.sievePasses configuration property . If you want to use the keyword, but do not want to do NER, you should just start the pipeline without NER and indicate that only annotation passages for which NER labels are not required should be used in the coref system.

However, the resulting coref annotations hit recall . That way, you can do some experimentation to determine if degraded annotation quality is a problem for what you use for the subsequent stream.

Answer 2

In general, yes. First, you need named objects because they serve as candidate antecedents or goals that include pronouns. Many (most?) Systems perform both object recognition and type classification in one step. Secondly, the semantic category (for example, person, organization, location) of objects is important for building accurate account assignment chains.