Foreword

Abstract

In preparation for writing this foreword, I looked through old emails (really old emails) dating back to early 1998, when we were planning the “First Workshop on Embodied Conversational Characters.” In and amongst detailed menu planning for the workshop (I haven’t changed a bit since those days) are emails floating the idea of publishing a book with the best papers from the workshop. We were already starting to see a shift in the literature, away from “lifelike computer characters” (Microsoft’s Clippy was presented at a workshop with that name) and “believable computer characters” (characters whose behavior was believable, but that did not do anything for people), and we wanted the book to reflect that shift. We particularly wanted to highlight the fact that embodied conversational characters did not only talk but also listened. They were capable of understanding as well as generating language and non-verbal behavior, and they did so in the service of humans—they were agents, like travel agents or real estate agents. To that end, I sent an email to the chapter authors with the following tidbits. I wrote: Next, a note about terminology. After long debate, we’ve decided to call the book Embodied Conversational *Agents*, and not *Characters* (for marketing reasons, in part) so you might want to follow this terminology in your chapter. Finally, do make sure to focus on the *communicative* abilities of your systems, since this is what distinguishes this work—and this book—from previous volumes on believable characters, software agents and so forth. It’s amusing to read this today when we take for granted the agentive nature of our conversational systems. At this point, we assume that embodied conversation agents (ECAs) are designed primarily to accomplish work for people. We also take for granted that ECAs must both understand and talk. However, when the Embodied Conversational Agents book was conceived, both of those features were only newly possible. In turn, the title of the current volume highlights the most recent technological innovation, which is the ability of the systems not just to do work for humans but to interact socially with them in the process, in many cases using social interaction as a way to bootstrap task performance. It’s illuminating to look at two other debates that took place during this same period. The first concerns what kinds of data are used to create the most natural behaviors for an ECA. The second concerns whether it is ethical to build natural-acting ECAs. While there was beginning to be consensus in the late 1990s on the idea that conversational characters could do more than just look pretty, there were three schools of thought about the proper inspiration for the conversational behaviors of ECAs (as they were called). Some of the authors in the original volume worked with actors to understand what kinds of language and non-verbal behaviors were most evocative of normal human conversation. These researchers hewed to the belief that ECAs should behave in a somewhat exaggerated fashion, like actors on a stage, in order to seem natural to their human interlocutors. Other authors believed that, being native speakers of their own language, and acculturated to the customs of their own society, the simple intuitions of the researcher were sufficient to design human-like conversational behaviors. A third group believed that psychological and linguistic studies of human conversation were the only proper inspiration for the behaviors of ECAs. Today, while a few researchers still work with actors or rely on their own intuitions, the community of researchers in ECAs (and in today’s socially interactive agents) mostly rely on empirical psychological and linguistic studies of human behavior as their inspiration. Some of these researchers carry out their own studies, and some rely on extant literature, but in both cases they rely on normal everyday humans for inspiration rather than actors or computer scientists. The debate is interesting in the face of today’s focus on big data. In fact, the increasing reliance in the field of artificial intelligence (AI) on machine learning techniques to analyze human behavior has led to a parallel increase in ECA systems that rely on deep learning techniques applied to large corpora, often of naturally produced human conversational behavior, to generate appropriate verbal and non-verbal conversational behavior. In other words, AI has brought us closer to the human-inspired ECAs of the past by bringing a focus on corpora of natural behaviors. At the same time, however, it has taken us further away from those human-inspired ECAs of the past because the corpora are too large to be examined by the human eye. Another debate that evoked heated interchanges in the late 1990s, and that is useful to contemplate today, was whether we should even contemplate deploying ECAs as interfaces to computational systems in the first place. Many if not most of the authors in the 1998 volume believed that ECAs represented a more natural way of interacting with computational systems than a keyboard and a mouse. Their work was predicated on the assumption that interacting with a human-like agent was a more intuitive manner of accessing technical systems. To other computer scientists of the era, however, ECAs were downright evil. Perhaps most famously, human–computer interaction researcher Ben Schneiderman saved his strongest invectives for human-like agents and their designers. In 1995, he wrote Anthropomorphic terms and concepts have continually been rejected by consumers, yet some designers fail to learn the lesson. Talking cash registers and cars, SmartPhone, SmartHome, Postal Buddy, Intelligent Dishwasher, and variations have all come and gone. Even the recent variation of Personal Digital Assistant had to give way to the more service oriented name now used in the Apple Newton ads: MessagePad. We’ll leave it to the psychoanalysts to fathom why some designers persist in applying human attributes to their creations … But, possibly, just possibly, all this heated debate is excessive and agents will merely become the Pet Rock of the 1990s—everyone knows they’re just for fun (Ben Shneiderman 1995. ACM Interactions. 2, 1, 13–15). Today, fears about whether robots will steal jobs, and whether machine learning will make it hard to tell who is human and who is an AI, have once again launched debates on whether human-like agents are a good or bad influence on society. These debates have led to a stronger focus on transparency in AI, a concern with bias in data, and a much-needed conversation on the ethics of where ECAs should and should not be used. These contemporary debates, however, and contra Shneiderman’s predictions, show that anthropomorphic agents have stood the test of time. The topic has inspired passion and dedication in a whole new generation of researchers. To that end, here 20 years later is a two-volume follow-up from our 1998 Embodied Conversational Agents book, with more than 25 chapters, showing the depth, breadth, innovation, creativity, and, yes, effectiveness, of human-inspired agents. Justine Cassell