Stream mining with integrity constraint learning for event extraction in evolving data streams

An event is a structured interaction of objects, generally involving an agent (actor) who acts on target entities, possibly with associated instruments. The problem of event extraction is to identify events of interest from a variety of information sources to be stored in a knowledge base for subsequent retrieval and analysis. Two major problems in event extraction are (1) adapting to concept drift and (2) maintaining the consistency of the event knowledge base. In this paper, we address these problems in a stream mining framework, making use of an ontology for events that represents background knowledge and integrity constraints derived empirically from an evolving data stream. We introduce a first-order language FLORE (Formal Language for Ontologies and Representing Events), which is used to encode an ontology of event types and objects, and to represent individual events categorized using concepts from the ontology. We present a multi-layered stream mining method for event extraction, where the first layer consists of a pool of simple learners, and the second layer learns an evolving set of integrity constraints to ensure the ongoing consistency of the extracted events. One intended application of this approach is conflict monitoring, and the domain of the Afghanistan conflict is used to illustrate the approach. Experimental results confirm that our multi-layered approach achieves higher recall and F1 than event extraction baselines on the event extraction task and on the subtasks of event detection and argument detection and classification.