With sophisticated enough natural language processing, relationships between concepts and actions that can be performed with individual objects can be extracted and modeled as an ontology (sort of like object orienting the concepts in an email). Spam can then be classified based on how well the concepts fit into the ontology (unless it includes valid text, which some now does). An ontology trained on valid email would contain a wide array of concepts and actions that a spam mail would not possess, or would not follow.

For example, “Bob said” may be parsed into “Bob”, an entity, performing an action called “said”. “Bob enlarged”, however, is an action much more likely to be found in a spam email than in most people’s ham. The associations between concepts and actions can be used as features in spam classification.