Large-Scale Information Extraction from Emails with Data Constraints

Email is the most frequently used web application for communication and collaboration due to its easy access, fast interactions, and convenient management. More than 60% of the email traffic constitutes business to consumer (B2C) emails (e.g., flight reservations, payment reminder, order confirmations, etc.). Most of these emails are generated by filling a template with user or transaction specific values from databases. In this paper we describe various algorithms related to extracting important information from these emails. Unlike web pages, emails are personal and due to privacy and legal considerations, no other human except the receiver can view them. Thus, adapting extraction techniques used for web pages, such as HTML wrapper-based techniques, have privacy and scalability challenges. We describe end-to-end information extraction system for emails-data collection, anonymization, classification, building the information extraction models, deployment, and monitoring. To handle the privacy and scalability issues, we focus on algorithms which can work with minimum human annotated samples for building classifier and extraction techniques. Similarly, we present algorithms to minimize samples for human inspection to detect precision and recall gaps in the extraction pipeline.