Without language, neither human intelligence nor our existence as social beings would be conceivable. Language is surely a central characteristic among those that are distinctly human. Therefore, it is no surprise that the study of language has fascinated scholars for a very long time. Indeed, some of the hotly disputed current issues in the study of language, such as the conflict between formal approaches that stress structure and order, and informal, observational approaches that stress the flexibility and context dependency of language, have been well formulated two thousand years ago. So is there nothing new under the sun in the study of language? A perusal of the present volume will convince the reader that this is by no means so, that a number of very important changes have occurred. For almost all of history, philosophers and linguists have focused on language as an object of analysis. They were interested in describing and analyzing the product language, whether this product was spoken discourse or written texts, accumulating a huge amount of important and useful knowledge about language in this way. Only in the last few decades have we begun to consider not just the product but also the processes by which we comprehend and use language. The present volume exemplifies this research strategy. It is a strategy that seeks not to replace linguistic or semantic analysis, but to complement and enrich it. Moreover, it is a strategy that holds a great deal of promise, because it combines the insights arising from the analysis of language as an object with a new set of constraints based on psychological and computational considerations. The data that psychologists have collected in their laboratories about language processing and the results that researchers in natural language processing have amassed about the properties and requirements of computational procedures provide a new look at language and help us to gain a new level of understanding that was not possible before.

Until now the conditions for studying the process of language understanding did not exist. Psychologists needed to develop and refine their research methods, and the computers and computational methods needed simply were not available earlier. In fact, there has been considerable interest in the process of understanding for some time among a school of literary criticism that focused on the reception of a poem or literary text by the reader, that is, the process of understanding the text, rather than on the text itself. However, their only tools were intuition and phenomenological analysis, which were insufficient to build an objective, convincing case. The computational approach, whose strength and promise are well demonstrated by the papers in the present volume, has changed all of that, making it possible to construct processing theories that are objective and testable and that can be powerful tools to investigate language understanding processes.

There are several distinctive features of the articles collected in this volume that set them apart from other work on reading and understanding, even work with a computational emphasis. First, the editors have paid a great deal of attention to the integration of computational approaches and psychological research. Thus, some of the chapters here are explicitly concerned with psychological data and models whose primary goal is to simulate human understanding processes. But even the chapters that focus directly on computational issues are also informed and constrained by what we know about human comprehension. Another important theme that runs through this book is the emphasis on knowledge representation and the role of knowledge in comprehension. Some of the toughest problems for computational models of understanding lie here. The contributors to this volume are vigorously exploring these problems and offer a variety of potential solutions. Finally, special note should be taken of the authors' attempts to deal explicitly with reading goals in the context of computational modeling. People don't read and understand in a vacuum but always for some reason, and their goals are important determinants of the reading process that have all too often been neglected in the past.

Why focus on reading, one might ask, why not study language understanding in general? Are not conversation and discussion equally interesting and important as reading? They certainly are, but there are good reasons for an initial focus on reading because a number of serious scientific problems can be avoided in that way. It is not possible to model conversation successfully without explicitly including the whole physical and interpersonal situation in which the conversation in embedded - which is not easily done. In contrast, written texts are designed to stand on their own, be understandable by readers whom the author cannot know beforehand and in situations the author cannot anticipate. Authors must make assumptions about the knowledge background and goals of their readers, but they can neglect many factors that play crucial roles in conversations. Hence reading comprehension, for all its complexity, is more readily analyzed than a conversation.

For the most part, the papers in this book deal with short and simple texts. Thus, our literary critic might not find much here, for it is a long way from a simple story to high literature. However, this is a young field, and starting with relatively simple though natural text is clearly the right research strategy to use. Furthermore, it is apparent in several of the chapters that the process of scaling up has already begun - both scaling up to complex, even literary texts, and to longer texts. In this effort the present approaches may someday be complemented by another research tradition in natural language understanding that relies on statistical techniques and the analysis of large corpora. Eventually, these different research enterprises might be combined, adding processing constraints to the statistical approaches and enabling computational process models to scale up.

Computational models of understanding are rapidly developing. The present volume will do much to focus this development and even accelerate it. It will also contribute to further the collaboration between computational modeling and the psychological study of language. Reading research is a broad and interdisciplinary field, and computational modeling must find its place in this varied landscape. It has much to offer to the more traditional approaches in the reading area, but it can also profit from the rich experience of the older approaches. The present volume demonstrates that continued interdisciplinary cooperation can pay off handsomely.

Walter Kintsch

Boulder, April 1997