Semantic Integration in Learning from Text

We define learning as the generation of meaningful knowledge representations which can be utilized in future decision making. Optimal learning entails that these knowledge representations be integrated with prior knowledge. In this paper, we introduce a knowledge representation based on an integration of a variety of shallow semantic parsing techniques. Entity detection, event detection, semantic role labeling and temporal relation identification are combined to produce graph-like structures which represent the most important semantic components of a text and the relations between these components. We show how new entities, events and relations can be successfully integrated into this representation using features derived from lexical and dependency-based sources.