Filed under: email,java,language technology,research,science,search,technology
Posted by: Andrew Lampert
In the early days of email, widely-used conventions for indicating quoted reply content and email signatures made it easy to segment email messages into their functional parts. Today, the explosion of different email formats and styles, coupled with the ad hoc ways in which people vary the structure and layout of their messages, means that simple techniques for identifying quoted replies that used to yield 95% accuracy now find less than 10% of such content.
Many language processing and search tools stand to benefit from better knowledge of the different functional parts of email messages, since this would allow them to focus on relevant content in specific parts of a message. In particular, access to zone information would allow email classification, summarisation and analysis tools to separate or filter out ‘noise’ and focus on the content in specific zones of a message that are relevant to the application at hand. Email contact mining tools, for example, might only access content from the email signature, while tools that attempt to identify tasks or action items in email might restrict themselves to the sender-authored and forwarded content.
Last week, I presented my paper on Segmenting Email Message Text into Zones at the Empirical Methods in Natural Language Processing (EMNLP) conference in Singapore. The focus of this work is Zebra, an SVM-based system that automatically segments and classifies the body text of email messages into nine functional zone types based on graphic, orthographic and lexical cues.
Our set of nine zones includes the following: author, greeting, signoff, quoted reply, forward, signature, advertising, disclaimer and attachment. Zebra currently performs the segmentation and classification of email text into the nine zones with an accuracy of about 87%. When the number of zones is abstracted to two or three zone classes (which is much more likely to be the granularity required for real-world email processing tasks), Zebra’s accuracy increases above 91.5%.
I’m currently working to finish off the Zebra system, as well as to resolve some licensing issues so that the code can be released for other researchers to use. We have, however, already released our annotated email dataset consisting of almost 12,000 lines of annotated email text that we used to train the Zebra system. If you want to know more, you can read our paper, head over to the Zebra website, or just get in touch with me by email or other means.