Predicate-specific Annotations for Implicit Role Binding: Corpus Annotation, Data Analysis and Evaluation Experiments

摘要

Current research on linking implicit roles in discourse is severely hampered by the lack of sufficient training resources, especially in the verbal domain: learning algorithms require higher-volume annotations for specific predicates in order to derive valid generalizations, and a larger volume of annotations is crucial for insightful evaluation and comparison of alternative models for role linking. We present a corpus of predicate-specific annotations for verbs in the FrameNet paradigm that are aligned with PropBank and VerbNet. A qualitative data analysis leads to observations regarding implicit role realization that can guide further annotation efforts. Experiments using role linking annotations for five predicates demonstrate high performance for these target predicates. Using our additional data in the SemEval task, we obtain overall performance gains of 2-4 points F_1 -score.