In this paper, we propose a method thatexploits full parsing information by representingit as features of argument classificationmodels and as constraints in integerlinear learning programs. In addition, totake advantage of SVM-based and MaximumEntropy-based argument classificationmodels, we incorporate their scoringmatrices, and use the combined matrix inthe above-mentioned integer linear programs.The experimental results show thatfull parsing information not only increasesthe F-score of argument classificationmodels by 0.7%, but alsoeffectively removes all labeling inconsistencies,which increases the F-score by0.64%. The ensemble of SVM and MEalso boosts the F-score by 0.77%. Oursystem achieves an F-score of 76.53% inthe development set and 76.38% in TestWSJ.
展开▼
