Detecting Action-Items in E-mail

E-mail users have a difficult time managing their inboxes in the face of mounting challenges. These include prioritizing e-mails from a variety of senders, filtering junk e-mail, and quickly taking action on items that demand the user’s attention. Automated action-item detection targets the third of these problems by detecting e-mails which require an action or response, and within those e-mails, highlighting the specific text that indicates the request. In contrast to action-item detection which aims at locating exactly where the action item requests are contained within the email body, typical text categorization (TC) merely assigns a topic label to the entire message — whether that label corresponds to an email folder or an indexing vocabulary [8]. In further contrast to TC, action-item detection attempts to recover the sender’s intent, i.e. whether she means to elicit response or action on the part of the receiver. Whereas TC by topic [5, 6, 9], TDT [1], and even genreclassification [7] work well using just individual words as features, we believe that action-item detection is the first TC task where we clearly must move beyond bag-of-words — albeit not too far, as bag-of-n-grams seems to suffice.