Understanding Sequence and Reply Relationships within Email Conversations: A Mixed-Model Visualization1Gina Danielle Venolia Microsoft Research One Microsoft Way Redmond, WA 98052 USA +1 425 703 2891 [email protected] Carman Neustaedter University of Calgary Department of Computer Science Calgary, AB, T2N 1N4 Canada +1 403 210 9404 [email protected] INTRODUCTIONThe mismatch between the user interfaces for email clients and user needs for handling email has been documented numerous times [1, 8, 14, 4]. This disparity has stimulated proposals for a plethora of client user interface design changes. One recurrent theme is that messages should be viewed as elements of a conversation rather than as independent elements [14, 10, 9]. A conversation, also known as a thread, is typically defined as the tree of messages that grows with the reply operation. One of plausible benefits of viewing messages as conversations is that the related messages will provide better local context, which can help one better understand the meaning of the message. Although this context is preserved to a limited extent by current email programs when they automatically include the text of the original message when replying, this method breaks down when a message receives multiple replies, creating a complex, branching reply tree. Subsequent replies provide additional context but these are not captured by quoting. TWO MODELS OF CONVERSATION Implicit in the designs of email clients and related research projects are two models of conversation that appear to be in conflict. On one hand, a conversation is a simple sequence of turns; on the other, a conversation is a branching tree. Sequential Model of Conversation A visualization supporting the sequential model can answer certain questions about a conversation: A. Which of these two messages was sent first? B. Which messages were sent before this one? C. Which messages were sent after this one? An interface supports the sequential model to the extent that it can answer these questions at a glance. Note that displaying the message “sent date” on a non-chronological list of messages does not satisfy the “at a glance” requirement as reading and comparing dates is substantial cognitive act. Typical email clients in a view sorted or grouped by conversation or subject, typical IM clients and Coterie [2] show a chronological list of messages clustered by conversation, thus strongly supporting the sequential model. Typical chat clients, typical email clients in normal operation, the interface proposed by Rohall et al. [10] and the Loom thread view [3] are chronological but messages are not clustered by conversation, undermining the sequential model because messages in the conversation are more likely to be scrolled out of view. Netscan [13] offers an interesting design point, being chronological by day but mixed within a day. Lotus Notes in its “Thread View”, ConverSpace [9], Threaded Chat [12] and typical Usenet newsgroup browsers use a schematic tree view where replies to a message are sorted chronologically; such a tree view can answer some of the above questions some of the time but cannot be relied upon to answer all of the questions all of the time. Conversation Map [11] does not represent the chronological sequence of messages. Tree Model of Conversation A visualization supporting the tree model of conversation can answer these questions about a conversation: D. Which message is the root of the conversation tree? E. Which message is this one a reply to? F. Does this message have any replies? G. Which messages are replies to this one? An interface supports the tree model to the extent that it can answer these questions at a glance. Lotus Notes in its “Thread View,” ConverSpace and Threaded Chat show the messages in a tree view, and thus strongly support the tree model. Netscan, Conversation Map, the work of Rohall et al., the Loom thread view and typical Usenet newsgroup browsers each shows a schematic tree that reflects the selection of a single message to be viewed, but cannot answer questions E or G at a glance; all these interfaces would typically be used with quoting, so question E would be answered by the message content, leaving G unanswered. Typical email clients rely only on quoting, thus they answer only questions D and E and support the tree model weakly. Coterie and typical chat and IM clients do not support the tree model at all. Models in Conflict? The degree to which each model is supported can be taken as two orthogonal axes. The interfaces discussed here may then be placed approximately in the resulting space captured graphically in Figure 1. Note that the top-right corner of the space is empty. 1This is an abbreviated version of a paper submitted to CHI 2003. See http://research.microsoft.com/~ginav/mixed_mode.pdf for the complete version.Is it possible for a visualization to fully support the sequential and tree models simultaneously? We set out to design one. VISUALIZATION DESIGN In preparation for designing a mixed-model conversation visualization that fully supports both the sequential and tree models, we gathered our design requirements. To begin with, it had to answer all seven questions above at a glance. The “at a glance” requirement ruled out dependence on interaction, e.g. selection or mouse-over highlighting (though we knew we might later add those to reinforce the visualization). That same requirement also ruled out a separate message viewing pane: the message content had to be present in the visualization. From examining many conversation trees, we were aware that they tend to be narrow rather than bushy – that is to say that a message is much more likely to get one reply than two or more – so chains of replies-to-replies should be visualized cleanly. We chose to list the messages in a chronological, vertical list for three reasons. First, a chronological list of messages supports the sequential model trivially. Second, the messages could be reflowed to fit the available width. Third, scrolling (if necessary) would be in one dimension only. We chose to list the messages from old (top) to new (bottom) so it could be read like a script. But what about supporting the tree model of conversations? The root is always the first chronologically, so Question D is answered, leaving Questions E, F and G open. Specifically we had to design a visualization that showed the reply relationships. We suspected that