VIO: a mixed-initiative approach to learning and automating procedural update tasks John Zimmerman, Anthony Tomasic, Isaac Simmons, Ian Hargraves, Ken Mohnkern, Jason Cornwell, Robert Martin McGuire Carnegie Mellon University {johnz, tomasic}@cs.cmu.edu ABSTRACT Today many workers spend too much of their time translating their co-workers’ requests into structures that information systems can understand. This paper presents the novel interaction design and evaluation of VIO, an agent that helps workers translate request. VIO monitors requests and makes suggestions to speed up the translation. VIO allows users to quickly correct agent errors. These corrections are used to improve agent performance as it learns to automate work. Our evaluations demonstrate that this type of agent can significantly reduce task completion time, freeing workers from mundane tasks. Author Keywords agents, interaction design, mixed initiative. ACM Classification Keywords H.5.2 User Interfaces: interaction design INTRODUCTION Today many workers in companies spend time translating requests into language and structures that information systems can understand. Consider the task of transferring a student from a waitlist to a course. The requester, a professor, has an intent that matches a common work task. The professor expresses her intent in an email to the department coordinator with relevant information such as the student’s and course’s names. The coordinator then logs in to the appropriate systems and makes the changes, translating the request into information the system can understand. Organizations address translation tasks by assigning a human-service-agent, such as administrative assistants, webmasters, network administrators, purchasers, etc., who perform procedural translation tasks on behalf of coworkers or customers. Procedural translation tasks are good candidates for automation because the input is easily captured, the output is structured, and the tasks are repeatedly executed. In order to study this opportunity we have developed a machine-learning-based agent and mixed- initiative interface. Called VIO, our agent takes on the role of a webmaster’s assistant (Figure 1). Requesters email requests (i.e., updates for a website) to the webmaster using natural language. VIO preprocesses the requests and pre- fills website update forms with suggestions. These pre- filled forms are presented to the webmaster for approval. The forms are an augmentation of a traditional direct- manipulation interface that allow the webmaster to quickly recognize the task and repair mistakes made by VIO. Our interaction design focuses on making repairs easy because (i) we accept that agents make errors, and (ii) having an interface that lets webmasters correct errors by doing their regular work and without generating additional work allows VIO to be deployed with little or no training. Through the process of repairing and approving forms, webmasters provide training data, allowing VIO to “learn in the wild,” that is, directly from use. This frees the webmaster to then focus on non-procedural tasks that require more human skill. Figure 1. Webmaster repairs and approves the task form causing the web database to update and VIO to learn from the addition of a new training example. Casting VIO as a webmaster’s assistant is a first step to concretely test our ideas. However, the design principles of VIO generalize to a much larger set of procedural tasks found within organizations. The design of VIO raises several fundamental HCI research questions including: (1) How effective is a human-service- agent collaborating with an agent that has had little training compared to a traditional direct manipulation interface? (2) How effective is VIO if it performs perfectly? (3) How do VIO’s errors impact overall performance? In this paper we address all these questions. We begin by describing our novel interaction method that combines natural language interaction—in this case the preprocessing of incoming email requests—with existing direct manipulation tools, and a feedback loop to the machine-learning algorithms. In Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. CHI 2007, April 28–May 3, 2007, San Jose, California, USA. Copyright 2007 ACM 978-1-59593-593-9/07/0004...$5.00. CHI 2007 Proceedings • Programming By & With End-Users April 28-May 3, 2007 • San Jose, CA, USA 1445