We present a four-step hierarchical SRLstrategy which generalizes the classicaltwo-level approach (boundary detectionand classification). To achieve this, wehave split the classification step by groupingtogether roles which share linguisticproperties (e.g. Core Roles versus Adjuncts).The results show that the nonoptimizedhierarchical approach is computationallymore efficient than the traditionalsystems and it preserves their accuracy.
展开▼