Detecting Emails Containing Requests for Action

Human Language Technologies: The 2010 Annual Conference of the North American Chapter of the ACL, pages 984–992, Los Angeles, California, June 2010. c⃝2010 Association for Computational Linguistics Detecting Emails Containing Requests for Action Andrew Lampert †‡ †CSIRO ICT Centre PO Box 76 Epping 1710 Australia andrew.lampert@csiro.au Robert Dale ‡Centre for Language Technology Macquarie University 2109 Australia rdale@science.mq.edu.au Cecile Paris CSIRO ICT Centre PO Box 76 Epping 1710 Australia cecile.paris@csiro.au Abstract Automatically finding email messages that contain requests for action can provide valu- able assistance to users who otherwise strug- gle to give appropriate attention to the ac- tionable tasks in their inbox. As a speech act classification task, however, automatically recognising requests in free text is particularly challenging. The problem is compounded by the fact that typical emails contain extrane- ous material that makes it difficult to isolate the content that is directed to the recipient of the email message. In this paper, we report on an email classification system which iden- tifies messages containing requests; we then show how, by segmenting the content of email messages into different functional zones and then considering only content in a small num- ber of message zones when detecting requests, we can improve the accuracy of message-level automated request classification to 83.76%, a relative increase of 15.9%. This represents an error reduction of 41% compared with the same request classifier deployed without email zoning. 1 Introduction The variety of linguistic forms that can be used to express requests, and in particular the frequency with which indirect speech acts are used in email, is a major source of difficulty in determining whether an email message contains one or more requests. Another significant problem arises from the fact that whether or not a request is directed at the recipient of the email message depends on where in the message the request is found. Most obviously, if the request is part of a replied-to message that is contained within the current message, then it is perhaps more likely that this request was directed at the sender of the current message. However, separating out content intended for the recipient from other extraneous con- tent is not as simple as it might appear. Segmenting email messages into their different functional parts is hampered by the lack of standard syntax used by different email clients to indicate different message parts, and by the ad hoc ways in which people vary the structure and layout of messages. In this paper, we present our results in classifying messages according to whether or not they contain requests, and then show how a separate classifier that aims to determine the nature of the zones that make up an email message can improve upon these results. Section 2 contains some context and moti- vation for this work before we briefly review rele- vant related work in Section 3. Then, in Section 4, we describe a first experiment in request classifica- tion using data gathered from a manual annotation experiment. In analysing the errors made …

Content truncated to 500 words