1 Introduction to the Transactions on Interactive Intelligent Systems ANTHONY JAMESON, German Research Center for Artificial Intelligence (DFKI) JOHN RIEDL, University of Minnesota This editorial introduction describes the aims and scope of the ACM Transactions on Interactive Intelligent Systems, explains how it aims to constitute a landmark addition to the publication landscape, and shows how the five articles in this inaugural issue fit into the journal s conception. Categories and Subject Descriptors: H.1.2 [User/Machine Systems]; I.2 [Artificial Intelligence] General Terms: Human Factors Additional Key Words and Phrases: Intelligent systems, human-computer interaction, interactive intelligent systems ACM Reference Format: Jameson, A. and Riedl, J. 2011. Introduction to the Transactions on Interactive Intelligent Systems. ACM Trans. Interact. Intell. Syst. 1, 1, Article 1 (October 2011), 6 pages. DOI = 10.1145/2030365.2030366 http://doi.acm.org/10.1145/2030365.2030366 Dear readers, welcome to the inaugural issue of the ACM Transactions on Interactive Intelligent Systems! We d like to remind you what the journal is about, explain why you may find it to be a valuable resource, and give a preview of what you can expect in this issue and subsequent issues. 1. WHAT ARE INTERACTIVE INTELLIGENT SYSTEMS, AND WHY ARE THEY INTERESTING? An interactive intelligent system is an intelligent system that people interact with. An intelligent system embodies one or more capabilities that have traditionally been associated more strongly with humans than with computers, such as the abilities to perceive, interpret, learn, use language, reason, plan, and decide. Systems that exhibit these capabilities mostly use techniques that originated within the field of artificial intelligence, though there is a general tendency (not followed in this journal), once these techniques have become successful and widely deployed, to characterize them in terms of their specific functions (e.g. speech recognition or web search ) rather than as intelligent systems. The technical design of intelligent systems raises fascinating challenges, but a new level of complexity is reached when people interact with these systems: We then have an interaction that involves both human and artificial varieties of intelligence. It is often difficult to understand interactive intelligent systems without examining the intelligent capabilities on both sides of the human-computer divide. Consider, for example, a system that learns how to assist users in performing particular types of tasks. While the system is learning, the users will in general also be learning: about the task itself, about the system and its learning, about how to act in such a way that the Author s addresses: A. Jameson, Intelligent User Interfaces Department, German Research Center for Artificial Intelligence (DFKI); J. Riedl, Computer Science and Engineering, University of Minnesota. Permission to make digital or hard copies of part or all 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 show this notice on the first page or initial screen of a display along with the full citation. Copyrights for components of this work owned by others than ACM must be honored. Abstracting with credit is permitted. To copy otherwise, to republish, to post on servers, to redistribute to lists, or to use any component of this work in other works requires prior specific permission and/or a fee. Permissions may be requested from Publications Dept., ACM, Inc., 2 Penn Plaza, Suite 701, New York, NY 10121-0701 USA, fax +1 (212) 869-0481, or permissions@acm.org. c 2011 ACM 2160-6455/2011/10-ART1 $10.00 DOI 10.1145/2030365.2030366 http://doi.acm.org/10.1145/2030365.2030366
1:2 A. Jameson and J. Riedl
Introduction to the Transactions on Interactive Intelligent Systems 1:3 suggested in the second row of Figure 1, without necessarily considering the relationships between the human and artificial capabilities in detail. In the strongest form of binocularity, shown in the top graphic, research and design aim explicitly to support the interaction between the forms of intelligent processing found on both sides. The journal aims to serve as a uniquely appropriate forum for the publication of substantial research that contributes to this binocular view by taking any of the perspectives shown in Figure 1. 2. RELEVANT RESEARCH AREAS Research on interactive intelligent systems is found in a considerable number of research areas, which illustrate the different uses to which system intelligence can be put in an interactive system. (1) In some research areas, such as recommender systems, information retrieval, or intelligent learning environments, the system s intelligence typically consists in learning, reasoning, or decision making which supports the system s primary function (e.g., suggesting appropriate products or documents; monitoring and supporting a learner s progress). (2) In some other areas, the main contribution of the intelligence is to enhance communication between the system and users, in a way which may not be closely related to the system s main function. This is the contribution most commonly found in the areas of multimodal interaction, natural language processing, embodied conversational agents, computer graphics, and accessible computing. (3) In the areas of model-based user interface design and automated usability testing, the system s intelligence is deployed not at runtime during the interaction with users but earlier, during the process of designing and testing the interactive system. (4) In a number of other relevant research areas, the contributions of the system s intelligence can take two or even all three of the forms just discussed: human-robot interaction; intelligent ubiquitous or mobile computing; user modeling, adaptation, and personalization; AI and games; knowledge capture; and intelligent systems based on semantic technologies. 3. COMMON ISSUES WITH INTERACTIVE INTELLIGENT SYSTEMS Despite the many different research areas and the many forms that interactive intelligent systems can take, there are a number of general issues and challenges that arise again and again. These issues make it interesting and profitable for people who specialize in one particular type of system to read about very different types of system. A first general question is: In what ways can artificial and human intelligence work together effectively? For example, where does intelligent processing yield the greatest benefits for interaction relative to other forms of computation? What patterns of division of processing between the human and the intelligent system, such as forms of mixed-initiative interaction, tend to be successful, and which less so? A second general issue concerns ways of dealing with the possible negative side effects that intelligent system processing can have if it is not designed with careful attention to the cognitive processes of users. For example, when and why is it important for users to be able reliably to predict, understand, and control a system that exhibits intelligence but sometimes errs, and how can we enable them to do so? What are effective strategies for protecting users privacy and avoiding undue narrowing of their experience when they are interacting with intelligent systems that know a lot about them? Third, there are characteristic methodological issues that arise with many types of interactive intelligent system, ranging from ways of understanding users requirements
1:4 A. Jameson and J. Riedl with regard to novel forms of intelligent interaction to ways of evaluating the success of a given combination of intelligent algorithms and interaction design. By bringing together in one venue research that sheds light on these general issues, TiiS aims to create synergies and cross-fertilization that will lead to more successful research in the future. 4. REASONS TO CONTRIBUTE TO TIIS We have aimed to make TiiS not just another journal but rather a prominent addition to the publication landscape that offers significant new opportunities for researchers in the roles of author, reader, and reviewer. (1) TiiS is the only journal that explicitly encourages and supports the perspectives on interactive intelligent systems illustrated in Figure 1, and its editors and reviewers have the knowledge and interests that are required to do justice to articles from all of these perspectives. Hence TiiS enables authors to take the perspective that best fits their research, instead of focusing on particular aspects of the research in order to fit into the scope of a given conference or journal. (2) Authors in this area no longer need to split the publication of their research over several overlapping conference papers, and readers do not need to collect these dispersed papers in an effort to piece together a coherent picture of the research. Within a single article, almost universally available via the ACM Digital Library and in print, authors can situate their work thoroughly within the landscape of related work, discuss both the intelligent technology and people s interaction with it, and provide all of the details that are required to use and build on the work. In addition to being more useful for readers who are interested in the topic in question, such an article can also serve as an introduction to the topic for readers who were previously unfamiliar with it, as is illustrated by the five articles in this inaugural issue. (3) If authors want to introduce a bold and possibly controversial new approach, they have enough space to provide strong arguments and evidence. If the reviewers and editors are not convinced right away, the authors can strengthen their case on the basis of the feedback received, a process that is likely to increase the ultimate impact of the research within the community. (4) TiiS has extended the default reviewing procedure and infrastructure to provide more valuable feedback to authors and reviewers in a way that does justice to their expertise and their investment of time: Before a decision on a manuscript is made, the authors are given an opportunity to point out any misunderstandings that they perceive in the reviews and (if they so choose) to give a brief preview of how they will respond to the reviewers and editors suggestions in a subsequent version of the article. The responsible associate editor has the option of contacting all reviewers to allow them to read each other s reviews (and any author response) and perhaps to change their reviews. Reviewers have access to the final decision letter on any manuscript that they have reviewed, which includes all of the reviews and comments of the authors and the editors. (5) With a policy unusual among journals, TiiS lists all of its reviewers permanently and prominently on a page of the journal s web site (http://tiis.acm.org/ reviewers.html) which encourages readers to see who has been contributing their time and expertise to the advancement of the journal and of the community of researchers on interactive intelligent systems.
Introduction to the Transactions on Interactive Intelligent Systems 1:5 5. PREVIEW OF THE ARTICLES IN THIS INAUGURAL ISSUE Although no set of articles that could fit in a single issue could suggest the full range of topics covered by the journal, the five articles in this inaugural issue do illustrate the binocular approach discussed at the beginning of this introduction. The fourth and fifth articles, by Okita et al. and by Gibet et al., comprise Part 1 of the special issue on Affective Interaction in Natural Environments, whose call for papers encouraged reports on studies that provide new insights into the use of multimodal and multimedia techniques for enabling interaction between humans, robots, and virtual agents in naturalistic settings. In the larger Part 2 of this special issue, which will follow soon, the special issue editors 3 will provide an introduction to the topic that discusses all of the accepted articles, including the two published here. 5.1. Why-Oriented End-User Debugging of Naïve Bayes Text Classification The first article to be published in ACM TiiS, by Kulesza et al., addresses some fundamental issues that arise when people interact with an intelligent system that employs machine learning: Given that machine learning does not in general result in 100% accuracy, how can users deal with the system s incorrect inferences and actions? The authors discussion of previous work can serve as an introduction to this research topic. The solution approach explored by the authors is to enable the user (a) to ask the system to explain one of its actions and then (b) to modify some aspect of the system s learned model by graphically critiquing the explanation. Successfully realizing this general strategy raises various challenges, which are illuminated in the authors detailed analysis of their user study. Overall, this article is a good example of an attempt to find an effective combination of intelligent processing on the part of a system and the humans who interact with it (see the top graphic in Figure 1). 5.2. Active Multiple Kernel Learning for Interactive 3D Object Retrieval Systems Hoi and Jin likewise address the question of how users can help a system to learn how to do useful things for them. In contrast to Kulesza et al., the authors apply the general strategy of asking users to label carefully selected training examples. Since this behavior is more straightforward for users than critiquing a system s explanations, the focus in this second article is appropriately on advances in machine learning that enable the system to acquire the most useful input from the user while requiring a minimum amount of the user s time and attention. 5.3. Recognizing Sketched Multistroke Primitives With the article by Hammond and Paulson, we turn from learning to another typical form of processing in interactive intelligent systems: recognizing input that users provide via human-like channels such as speech, handwriting or, in this case, sketching. After discussing the general benefits to users of sketch recognition, the authors focus on cases where a user sketches a primitive shape (such as an arrow) with more than one stroke. The system s intelligence consists in an algorithm for recognizing multistroke primitives, but this algorithm and its use in an interactive system must be based on a detailed understanding of how users tend to draw and use multistroke primitives in the first place, a question addressed here with a user study. 5.4. Multimodal Approach to Affective Human-Robot Interaction Design with Children Human-like communication modalities likewise figure in the article by Okita et al., but here it is the system that is generating the communication: a robot designed 3 Ginevra Castellano, Laurel Riek, Kostas Karpouzis, Jean-Claude Martin, Louis-Philippe Morency, and Christopher Peters.
1:6 A. Jameson and J. Riedl specifically to interact with children. The key question of interest is how children (of different age levels) respond emotionally and socially to different patterns of robot behavior. The results illustrate vividly how users behavior with an interactive intelligent system can exhibit patterns that would be impossible to predict solely on the basis of an understanding of the system s own capabilities. The article also provides examples of typical methodological issues, such as the use of the Wizard-of-Oz method for experimenting with possible intelligent system behaviors and methods for measuring and analyzing users responses to interactive intelligent systems. 5.5. The SignCom System for Data-Driven Animation of Interactive Virtual Signers: Methodology and Evaluation The article by Gibet et al. also concerns the generation of human-like behavior by a system here, a virtual character that generates French Sign Language for the benefit of people who rely on that language for everyday communication. In contrast to the article by Okita et al., the main focus is on the technical challenges involved in generating natural and comprehensible multichannel sign language. But the authors also report on a user study that sheds light on various aspects of users responses to the virtual characters performances. 6. THANKS TO CONTRIBUTORS The advances described in this introductory essay have been the result of the efforts of hundreds of people. Dozens of distinguished senior representatives of the relevant research areas, including those who now serve as TiiS associate editors, supported the original proposal to ACM s Publications Board to create the Transactions on Interactive Intelligent Systems. The first special issue associate editors 4 gave a big boost to the newly founded journal by creating awareness of it in their communities, attracting and managing dozens of interesting submissions, and helping to extend and optimize the reviewing procedures and infrastructure. Hundreds of expert authors and reviewers have worked together to create a substantial pipeline of articles for the first few issues, typically working hand in hand, despite the veil of reviewer anonymity, to maximize the quality of each revised or resubmitted manuscript. As the next few issues become ready for publication, you will find regularly updated information on the TiiS web site. 5 We hope that these efforts and their results will inspire you to join and support this large-scale effort to raise research on interactive intelligent systems to a new level. 4 http://tiis.acm.org/special-issues.html. 5 http://tiis.acm.org.