Semantic Role Labelling withTree Conditional Random Fields

摘要

In this paper we apply conditionalrandom fields (CRFs) to the semanticrole labelling task. We define a randomfield over the structure of each sentence’ssyntactic parse tree. For each nodeof the tree, the model must predict asemantic role label, which is interpretedas the labelling for the correspondingsyntactic constituent. We show howmodelling the task as a tree labellingproblem allows for the use of efficientCRF inference algorithms, while alsoincreasing generalisation performancewhen compared to the equivalentmaximum entropy classifier. We haveparticipated in the CoNLL-2005 sharedtask closed challenge with full syntacticinformation.