The thesis is divided into three parts: Part I gives a background survey of SFG and its application to natural language generation, and a review of some related issues in machine translation. Part II describes an implemented and working prototype system which demonstrates an SFG-based approach to the three problems listed above, which could be applied to target language generation in Japanese-English and Japanese-European MT. Part III briefly evaluates the prototype and the SFG-based approach, and suggests some areas in which an exchange of ideas between SFG and other approaches might be fruitful.
In Part I, Chapter 1 reviews the basic concepts of SFG on which the rest of the study is based. These include the organization of functional choices into a network, the way in which these choices are realized in surface structure, the coverage of the wider range of factors (semantic, pragmatic and discourse factors), and the way in which the different factors are combined in generation.
Chapter 2 describes two large computational implementations of SFG applied to English generation: the USC Penman system and the Cardiff COMMUNAL system. The descriptions include the grammars on which the systems are based, the ways in which inquiries for functional choices are made, and the other components of the generation environment in which the grammars are embedded.
Chapter 3 looks at current research on the problems involved in extending SFG-based generation from monolingual English to multilingual generation. Two further projects are described: the Sydney project on SFG-based generation of English, Chinese and Japanese, and the Darmstadt KOMET project on generation of German, English and Dutch. The descriptions include the functional approach to multilingual grammar sharing, the representation of multilingual grammars, and problems in handling grammatical gender agreement in SFG.
Chapter 4 reviews some basic issues in machine translation (MT), including the important differences between import and export translation, and the need to reduce the extreme dependence on skilled post-editors. Some research approaches towards the idea of MT for monolingual users are described. Multilingual MT and multilingual generation are compared, as two alternative solutions to the problems of international organizations. The survey ends with ideas for an SFG-based approach to MT.
In Part II, the demonstration prototype is described. Chapter 5 gives details of the small grammar embedded in the prototype. The organization of the choice networks of the grammar is explained.
Chapter 6 describes the new implementation of the grammar to meet the requirements of the demonstration prototype. The method of representation for the grammar network and the realization rules is shown, and the implementation of the procedures is described, with extracts from key parts of the code.
Chapter 7 explains the design of the Japanese interface, and its use for interactive English generation. Illustrative examples of interactive sentence revision are presented in the form of window images.
Chapter 8 describes the extension of the prototype to include French and German generation. The grammar network is shared by the three languages, but separate realization rules were developed. Examples of multilingual sentence revision using the Japanese interface are presented.
In Part III, Chapter 9 evaluates the prototype. The possibility of applying this approach to knowledge-based post-editing is considered.
Chapter 10 looks at strengths and weaknesses of SFG. These are briefly contrasted with corresponding weaknesses and strengths of HPSG. The thesis concludes with a brief examination of the idea of combining the strengths of the two theories, and the potential benefits of doing so.