Msharifi writeup of Jansche and Abney

This is a review of Jansche_2002_information_extraction_from_voicemail_transcripts by user:msharifi.

Paper describes a loglinear model applied to the task of extracting caller phrase and caller name from a voicemail corpus. Main focus is on feature engineering and some of them are interesting (e.g., the beginning words of conversation). Also, the success of rule based systems comes from this regularity in the data which is crude way of using features.

While it may be necessary to do two stage filtering but the systems trying to combine and optimize both stages simultanuously are more prefered these days.