An iterative classifier for unsegmented electronic documents is based on machine learning algorithms. The textual strings in the electronic document are segmented using a composite dictionary that combines a conventional dictionary and an adaptive dictionary developed based on the context and nature of the electronic document. The classifier is built using a corpus of training and testing samples automatically extracted from the electronic document by detecting signatures for a set of pre-established classes for the textual strings. The classifier is further iteratively improved by automatically expanding the corpus of training and testing samples in real-time when textual strings in new electronic documents are processed and classified.