Making email actionable: the identification and use of obligation acts in workplace email / by Andrew Lampert.
thesisposted on 28.03.2022, 13:24 by Andrew Thomas Lampert
"Email is a key communication medium in business environments, where it is often used to assign and delegate tasks. Existing research has established that task-oriented communication is built upon the exchange of request and commitment speech acts collectively, obligation acts between interlocutors, but email software has so far ignored this insight; it has not adapted to support task management, despite its popularity as a medium for such workflows. The lack of task awareness in email software has been repeatedly highlighted as a key factor in the 'information overload' that burdens many email users. In particular, the difficulty of distilling tasks from the ever-increasing email flow leads to obligations that remain unfulfilled. This thesis explores how to address this problem by making email more actionable. We begin by analysing data from a series of annotation experiments through which we gathered independent human judgements about requests and commitments across a collection of more than 2000 real-world email messages. These annotated messages provide insight into how obligation acts are realised and interpreted. We identify and analyse a range of complex phenomena involved in these speech acts, and provide definitions for identifying them in email. Building on this analysis, we then present effective computational techniques for detecting obligation acts at three levels of granularity within email messages: the message, paragraph and sentence levels. Message-level identification determines whether or not an email message contains obligation acts, and aims to assist users to triage their messages by focusing on those containing actionable content. Paragraph-level identification builds on this to classify each paragraph in the same manner; this enables, for example, the production of extractive summaries of messages. Finally, sentence-level identification classifies each sentence in an email message, and allows requests and commitments to be extracted to external task lists. We use our annotated email data to train supervised machine learning algorithms for each of these classification tasks. These classifiers also exploit a novel classification system that segments the text of email messages into different functional zones, identifying material such as signatures, advertising, and quoted reply content. This enables our obligation classifiers to focus on only relevant email text when identifying requests and commitments. In sum, this thesis provides both theoretical and practical foundations for managing obligation acts in real email data. Our empirically-grounded analysis and the software tools developed on the basis of this analysis demonstrate how it is possible to make email actionable today, as well as providing a platform for future task-related email research." -- Abstract.