Automatic Semantic-Role Annotation for Portuguese

The paper presents and evaluates a parsing system for the automatic annotation of Porguguese text with semantic role tags. All in all, 38 different categories, like agent, patient, location etc. are distinguished. The annotater uses a grammar of 500 hand-written Constraint Grammar rules and exploits syntactic dependency links as well as semantic prototype classes and syntactic function. The interaction with lexical information like verb-argument frames and the ensuing boot-strapping problem are discussed. The system achieved promising results in an evaluation with treebank input from the Floresta Sintá(c)tica, with an average recall of 86.6% and a precision of 90.5%.