RPI BLENDER TAC-KBP2015 System Description

To extract English name mentions, we apply a linear-chain CRFs model trained from ACE 20032005 corpora (Li et al., 2012a). For Chinese and Spanish, we use Stanford name tagger (Finkel et al., 2005). We also encode several regular expression based rules to extract poster name mentions in discussion forum posts. In this year’s task, person nominal mentions extraction is added. There are two major challenges: (1) Only person nominal mentions referring to specific, individual real-world entities need to be extracted. Therefore, a system should be able to distinguish specific and generic person nominal mentions; (2) within-document coreference resolution should be applied to clustering person nominial and name mentions. We apply heuristic rules to try to solve these two challenges: (1) We consider person nominal mentions that appear after indefinite articles (e.g., a/an) or conditional conjunctions (e.g., if ) as generic. The person nomnial mention extraction F1 score of this approach is around 46% for English training data. (2) For coreference resolution, if the closest mention of a person nominal mention is a name, then we consider they are coreferential. The accuracy of this approach is 67% using perfect mentions in English training data.