The 5th Annual Canadian Symposium on Text Analysis (CaSTA'06), October 12, 2006, Fredericton, New Brunswick, Canada
Forgetting to attach an attachment to an outgoing e-mail message is at best embarrassing yet the detection of such mistakes is problematic. Existing systems parse a messages body for keywords but typically have limited effectiveness due to their high precision and low recall. An analysis of a small corpus of emails showed that parsing the text of emails- even when supplemented by rules that parsed the high level attributes of the messages only enabled a high recall at the expense of low precision. A system is described that utilises such rules calibrated for each user to determine the likelihood or system confidence - of each rule being correct when fired. The form of user alert is modified according to the system confidence in the rule(s) fired allowing the system recall to remain high without affecting the systems usability.
Proceedings of the 5th Annual Canadian Symposium on Text Analysis (CaSTA'06), October 12, 2006, Fredericton, New Brunswick, Canada.