Generic Key Value Extractions from Emails

Abstract

Web information extraction systems are widely used to find and understand relevant parts of a text, combine multiple such parts, and produce a structured representation of the information. There are various scenarios where HTML formatted emails are generated by filling a template with user and transaction-specific values from databases. These emails are sent for human consumption among other things. Examples are such emails include flight confirmation emails, restaurant reservation emails, bills, hospital records, etc. In majority of these B2C emails information is presented in the form of key-value pairs i.e., user or transactions specific values are presented in an HTML format with their associated keys. In this paper, we describe a generic method to extract these key-value pairs which can be used for various applications. We analyze the pairs for a number of applications including identifying semantically similar keywords and creating clusters of keywords which then can be used for building information extraction wrappers. We show that just using word-embeddings is a poor substitute for finding similar keys. We use a number of features—types of values, cooccurrence graph of keys, etc., and combine them to present a keyword similarity algorithm which gives more than 50% improvement in homogeneity of the clusters, in comparison to just using word embeddings, using various real-world data.