Mining Script-Like Structures from the Web

This paper presents preliminary work to extract script-like structures, called events and event sets, from collections of web documents. Our approach, contrary to existing methods, is topic-driven in the sense that event sets are extracted for a specified topic. We introduce an iterative system architecture and present methods to reduce noise problems with web corpora. Preliminary results show that LSA-based event relatedness yields better event sets from web corpora than previous methods.