Challenges of Discourse Processing: The Case of Technical Documents