Uncovering the to-dos hidden in your in-box.

INTRODUCTION

The beginning of the 21st century has been marked by an explosion of electronic information. We often find ourselves inundated with information that reaches us through a variety of channels, including e-mail accounts, voice-mail recorders, and most recently, text-messaging clients on our cellular phones. This amounts to a deluge of requests, notices, scheduling invitations, and the like. We owe the information explosion to the proliferation of inexpensive and readily available technology that has led to what most would agree is both a blessing and a curse. Whereas information facilitates efficient business and social networking, it also has become a major burden. Humans simply cannot accommodate massive information input.

For economic reasons, e-mail is the most acute and widespread facilitator of this explosion due to the virtually free cost of generating and distributing e-mail messages. Even when we ignore spam and focus on legitimate e-mail (from trusted and welcome senders), we find that e-mail is still a major problem. (1-3) One user reports receiving well over 70 legitimate messages per day. (4) Anecdotal evidence suggests that his experience is not unusual. In an enterprise environment, a significant portion of these legitimate e-mails are associated with tasks that must be acted upon by the recipient.

A key challenge posed by e-mail inundation is how to effectively manage the tasks and activities that are associated with e-mail messages. Herein lies the goal of our work: to help users manage their tasks effectively. We consider a task to be a particular kind of activity. Moran defines "activity" as a set of mental or physical actions carried out by persons. (5) Through composition, activities can contain subactivities, which can themselves contain subactivities. In this vein, we define a task to be an atomic level activity, one that may not contain subactivities. We focus only on tasks that are communicated by e-mail messages, such that there is at most one task per message. As an example, the process of bidding for a product on eBay ** is an activity containing many subactivities. One e-mail associated with this process alerts the recipient that he or she has won an auction. The task for the recipient contained in this e-mail is to initiate a payment to the seller.

In this work, we focus on the population of business managers who receive daily a large number of legitimate, machine-generated e-mails, such as the ones that are generated by business processes within a large enterprise. We present here a practical solution for dealing with such e-mail in the form of a task management tool called SCOUT, (6) which uses contextual information about the user and the environment to recognize, filter, sort, organize and execute tasks associated with e-mails. By using information from pervasive sources (i.e., ubiquitous computing devices), SCOUT alleviates some of the problems associated with e-mail overload by presenting the core information to the recipient in an efficient and well-organized fashion.

We hypothesize that tasks contained within e-mail messages can be automatically identified for presentation within SCOUT. Tasks can be contained in one of two types of e-mail messages: human-generated and machine-generated. For our purposes, the salient difference is that human-generated messages tend to be unstructured, whereas the contents of machine-generated messages have a regular structure. To simplify the problem, we focus on machine-generated messages. We assume that every machine-generated message is associated with some business process (e.g., the eBay bidding process or the expense reimbursement process in an enterprise), that we only have access to e-mail messages generated by business processes, and that other than inspecting the e-mail messages themselves, we have no knowledge of the syntactic structure used in these messages. Furthermore, we assume that we make no modifications to messages or to the business processes that generate them.

SCOUT tracks a set of registered task types, each of which corresponds to a business process. When SCOUT identifies an e-mail message associated with a business process, the task contained within that e-mail message is specified in a document by using an Extensible Markup Language (XML) dialect called TaskML. A task description contains the following attributes:

* Type--the task type represented by a label unique to a business process or transaction associated with the task (e.g., a bidding transaction on eBay, a password update at Amazon.com Web site).

* Subject--a summary description of the task (e.g., you have won the auction)

* Person--an optional list of persons associated with the task (e.g., a collaborator who can help complete a task)

* Deadline--an optional deadline by which the task must be completed

* Thread--the set of related messages associated with the activity containing this task

* Comments--free-form comments associated with the task

* Status--the state of completion of the task

By automatically identifying tasks within e-mails generated by business processes, SCOUT helps make users aware of the tasks awaiting their attention. Furthermore, by pulling these e-mail messages into a task management system, it reduces the number of legitimate e-mail messages the user must process each day.

SCOUT provides three main functions: e-mail analysis, context-based task presentation, and context-based task reminding.

1. E-mail analysis: An e-mail analysis engine recognizes incoming e-mails as being associated with known business processes. Such e-mails are then parsed and further analyzed to extract task information relevant to that process.

2. Context-based task presentation: SCOUT uses context associated with a task so that it can be presented in a graphical interface that is customized according to the viewer.

3. Context-based task reminding: To extend SCOUT beyond the desktop, context-based reminders enable task-related messages to be sent to users on pervasive devices. Users can specify contextual criteria to trigger the reminding process (e.g., if my task is to pick up a package, alert me when I am in the vicinity of the mail room; if my task involves Steve, alert me when we are both available).

The e-mail analysis function is implemented using Unstructured Information Management Architecture (UIMA) annotators. UIMA (7) is a component-based software framework used for the development of applications that process unstructured information. It focuses on text analysis and isolates the core algorithms that perform text analytics from system services such as storage of data, communication between components, and visualization of results. By offering a framework with well-defined application programming interfaces (APIs), UIMA allows developers to share and combine text analysis algorithms in order to build complex applications.

The rest of the paper is organized as follows. In the next section, we review related work. In the following section we introduce the SCOUT application, describe the way in which the application requirements were defined, and describe the two interfaces to SCOUT, the Web portal and the e-mail client. Next we present the context information used by SCOUT, the sources of that information, and the way in which additional context is derived. We then describe the SCOUT architecture and give an overview of the e-mail analysis components. We present results of a pilot study and conclude with some final comments, including ideas for future work.

RELATED WORK

Moran and his colleagues identified several metatasks required for efficient task management (2,5):

* Creating awareness of the core task and related metatasks

* Prioritization of tasks

* Scheduling of task appointments

* Completion of task prerequisites

* Monitoring of task status

* Notification/reminders of partially completed tasks

* Delegation of tasks through reassignment

An important focus in task management is the awareness aspect. Although task management has received a great deal of attention in the literature, (8-11) most approaches tend to disregard the awareness problem. A notable exception is the work of Cortson-Oliver et al., (12) which deals with general e-mails. They propose SmartMail, a prototype task-extraction system that uses linear support vector machines (machine-learning method used for classification) and linguistic rules to analyze unstructured e-mails. Their technique produces task-focused summaries of action items detected in e-mails. With such a wide scope on general e-mails, their solution has had only modest predictive success.

Another exception is the work of Bennett and Carbonell (13) describing a system that tries to identify the action items contained in unstructured e-mails. They compared a standard unigram (1st order Markov) approach to an n-gram (n--1 order Markov) approach applied at both the document and sentence level. They found that n-grams applied at the sentence level are most effective, achieving accuracies of 0.8092, 0.8145 and 0.8173 for a k-nearest neighbor, naive Bayes, and support vector machine classifier, respectively. In contrast to this work, SCOUT limits the e-mail that it considers to those items that arrive from semi-structured business processes. In the case of one SCOUT user with an e-mail corpus consisting of 2,269 messages, the observed accuracy was 0.9996; similar results were obtained for other SCOUT users.