However one feels about connectionism, one will find much to think about in this book. It nominally describes a language generator named FIG (flexible incremental generator), but in so doing it also discusses such topics as “why previous research has missed the point,” “problems with case grammars,” and “artificial intelligence as an experimental science.” These and other topics discussed should interest anyone studying natural language generation in particular or natural language processing in general.

Based on the author’s doctoral thesis, but going beyond it, this book begins by focusing on the problems that arise in text generation due to conflicting goals, particularly because the choices that must be made depend on each other and on the nuances of meaning in the input to the generator. The preview of FIG in chapter 1 shows how it generates a grammatical English sentence from the representation of a Japanese sentence. Chapter 2 characterizes the generation task and describes the design principles that led the author to a connectionist approach.

Chapter 3 explains how selection of appropriate words is an emergent property in FIG. The connectionist model used here has been modified so that word selection is influenced by nodes representing syntactic constituents that become active in order. Chapter 4 explains how this leads to the output of the selected words in grammatical order as another emergent property in FIG. While the author argues in chapter5 against the use of case grammars to represent relational information, he replaces them with “participatory profiles”: vectors of numerical values representing the strengths of case-like features.

Chapter 6 gives details of the English and Japanese grammars used in FIG. Implementation details of FIG are given in chapter 7. Of particular interest are the special processes that were incorporated to allow for the emergence of proper word selection in the proper order and to prevent uncontrolled feedback in the network. Insights into how the network is fine-tuned are also given.

The next four chapters get more reflective. Chapter 8 discusses the strengths and weaknesses of connectionism and briefly compares FIG with several other connectionist language generators. Chapter 9 reviews some characteristics of human language production, such as pauses, priming effects, and errors, and construes FIG as a running cognitive model having similar characteristics.

In chapter 10, the author explains his ideas on machine translation and how it might be made more natural. He shows the need for new approaches to translation by presenting several Japanese sentences whose natural English translations cannot be generated by structure-preserving translation, parameterized texts, or translation by analogy, but require more creativity than these approaches can provide. The author sketches a connectionist approach to machine translation that he expects will lead to more natural output.

In the final chapter, Ward admits that the present implementation of FIG has limitations, and discusses what has been learned from it and what might be done in the future. An appendix lists FIG’s knowledge network.

This book gives the reader a good understanding of how FIG works without getting mired in the LISP implementation. The many discussions of design issues and problems arising in language generation and translation, while not extensive, suggest some interesting points of view that many readers may find stimulating.