Two designs for semi-structured communication in a problem-solving environment Alex Feinman Rene Holl Abstract In Vessel World, an environment for solving multi-user problems, communication is essential to coordinate solutions to the problems presented. As previous work has shown, providing structure for this communication may serve both to make coordination easier, and to create more focused collaboration. The purpose of this paper is to present two competing designs for providing such structure to Vessel World, and analyze the pros and cons of these designs. Background One method of altering communcation to be more effective is to allow that communication to be structured or semi-structured. In this case structured means allowing communication only from a small subset of commands and nouns in a particular order, allowing only a communication chosen from a preset list, or some such restriction. This is probably not feasible for most domains. Interestingly, even such a restricted domain does not reduce the problem to the trivial one of enumerating the category for each possible communication. This is due to the fact that most communications can serve more than one purpose, given the context. Humans are exceedingly clever at finding ways to communicate more than a restricted system would ordinarily allow; for example, using timing of a message to denote importance or even reverse the meaning of an utterance. So restricting the users to fully structured communication would harm much more than help. Semi-structured communication is a much less constrictive style of communication where certain portions of the communication may be tagged or otherwise marked, but the actual body of the message is unrestricted. Malone uses e-mail as his example; in a standard email message, as defined by RFC-822 (Crocker82), a message has a number of set headers, each of which has a range of valid values that vary from completely restricted (the date header, restricted to a valid date in a specified format) to completely free (the subject header, which is only restricted by a length limitation). In addition, email has a completely unstructured message body, allowing any amount of information in any form. Work by Malone and others indicates that requiring a certain amount of structure in messages improves human performance even when the structuring information is not acted upon by some external agent. This result is not entirely surprising (as Malone would have it) given that in many cases natural communication acquires a tightly defined structure in situations where performance is the goal. To test these theories out, we are implementing two versions of semi-structured communication for the Vessel World system. Vessel World is a system created to analyze various forms of cooperative planning and collaborative work. It consists of a simulated world where the users operate two sea-cranes to lift the waste and a tug boat to move barges around to transport the waste. The task consists of lifting and loading lodes of toxic waste situated at various places in a harbor, and loading them all onto a central barge. To perform this task efficiently, the users have to coordinate their actions tightly; the waste is scattered about and must be discovered by the agents, some pieces of toxic waste are too heavy for one crane to lift alone, and others require the use of a barge to transport. With all these complications, it is easy to understand why communication is so important. The locations and types of boxes must be communicated, the responsibility for retriving the different boxes must be delegated, and schedule for the retrival must be agreed on. System Design 1: Structured Messages The first design we will present follows the paradigm of modern email. It is a parcel-post system; that is, the sender puts together a message and sends it to the recipients, who receive the entire message at once. The recipients may then reply with messages of their own. There is no real-time interaction. This design decision was made for a few reasons: it made use of a commonly-understood paradigm (parcel-post); it fit in the turn-based system of Vessel World; and it was easy to implement. Shown in Figure 1 are the basic windows for message creation. A message, shown in Figure 1a, consists of a free-text area and a list of attachments. The free text area is used for comments or clarifications, and can also be used for simple messages that do not require use of the structured attachments. The user simply clicks and types to enter information here. This provides users with the capability to communicate simple messages very quickly, and allows flexibility by allowing the users to communcate information in a form that the system design might not otherwise allow. The attachments provide the structure to the message, and are added to message by building them using the attachment palette (Figure 1b). The user selects an attachment from the palette and clicks on the button marked "New Attachment". This brings up a window similar to the one shown in Figure 2. Each attachment contains a list of fields; shown is the attachment for an explicit plan, where the action of all three agents are shown, along with an area for naming the plan. Once the attachment is complete, it can be added to the message and/or saved for future use by clicking the appropriate button at the bottom of the form. The attachment appears in the appropriate area of the message, as shown in the completed message in Figure 3. If the user wishes to edit the attachment at some point, he can bring up the attachment window by double-clicking on the attachment itself. When the user is finished constructing the message, he hits the "Send" button, which distributes the message to its recipients. The recipients see a version of the message which looks identical but is not editable, and has a number of actions associated with receiving it. Drawing on work on commitment networks (Winograd88), we can show that the basic responses to a proposal are acceptance, rejection, or a counter-offer which involves some modification of the plan. The system takes advantage of this conversational structure by providing a simple way to perform these actions in the form of the buttons marked "Accept", "Decline", or "Reply" at the top of the message. Winograd showed a fourth type of response, which he termed "Commit to commit", wherein the recipient postpones replying until he is able to give an answer; this path is implicitly available to the user by simply delaying a response as necessary. System Design 2: Structured chat(?) Rene does this part. Comparison In general, the first design shown offers much less structure than the second design. The attachments are used as an addendum to the free text area, and are themselves only labeled, not truly constrained. The second design, on the other hand, is highly structured; sentences are built up from pre-defined parts and allow only a very limited As can be seen in the first design, most of the flexibility comes from allowing the user to fill in the fields with any text desired. The relatively small number of available attachments means that each attachment must be proportionally more useful than the equivalent sentences in the second system, where the large number of combinations provides most of the flexibility. However, this flexibility only covers the situations which the second system is specifically designed for; that is, it provides variety but not true spontaneity. This is the true danger of fully structured communication: all possible situations must be accounted for, or there will be circumstances where the system is unable to handle the communication necessary. To avoid this pitfall, the second system **** Despite the lack of flexibility in the second system, its highly structured methodology is very useful in both constraining the users to conversation along particular paths (in this case, the path of solving the problem) and providing ability for the system to take advantage of the structure present. Since the contents of each sentence are more or less deterministic, the system can figure out exactly what is in each sentence, and attempt to guess what purpose it has in the conversation. In the first design, it is much more difficult for the system to extract the purpose of more complex messages, although to a certain extent the types of attachments indicate the type of message. Future work Implementation of the first design is nearly complete; when it is, experiments comparing it to a version of Vessel World with completely unstructured communication will begin.