Re iews

Analogical Natural Language Processing aims to challenge the current hegemony of the rule-based paradigm in NLP. Traditional NLP decomposes languages into atomic units, whereas example-based NLP centres around the re-use of language fragments. The book consists of six chapters : a short introduction, chapters on background material, analogical machine translation, stochastic and analogy-based NLP, some experiments in analogical cloning and a conclusion. We shall now look at the chapters more closely. The first, introductory, chapter outlines the scope of the book, namely analogical approaches (including the use of parallel distributed processing systems) to NLP and the suitability of a functional grammar representation for the task of storing example texts. Analogy-based machine translation, in particular the matching of source and target languages, is used as an example throughout the book. The basic analogical machine translation architecture, involving matching, recombination and generation is also presented. The second chapter provides the background material which is necessary for the understanding of the subsequent chapters. The basic ideas behind example-based machine translation (EBMT), analogical modelling, functional grammar, parallel distributed processing, the PHRAN natural language front end and rhetorical structure theory are briefly introduced before the process of EBMT is looked at more closely. The decomposition\transfer\ composition model of EBMT is presented, and both Sato and Nagao's and Sumita's recombination metrics are discussed, along with their pros and cons. The chapter concludes by examining functional grammar and its application to machine translation. Chapter three centres on analogical translation. It covers the representation of translation examples and the nature of source examples, as well as the basic process of analogical translating. This involves using a shadow representation in functional grammar, linked to the example language text and using a match between these examples to provide the translation. The need to consider not only sentence structure but also text structure is made clear and amplified by means of examples. This not only leads to translations which are functionally closer to the rhetorical goals of the source text, but can be used to restrict the search for examples to those whose rhetorical function is closest to that of the original text. The problem with the need for the hand-coding of such relations is raised before a detailed example of EBMT is presented. Various problems are discussed at the close of the chapter, such as the problem with optional and obligatory functions, e.g. not all languages assign optional and obligatory …