Sublanguage Engineering In The Fog System

FoG currently produces bilingual marine and public weather forecasts at several Canadian weather offices. The system is engineered to reflect "good professional style" as found in human forecasts. However, some regularization and simplification of the output has been needed. Sublanguage engineering issues include tradeoffs in coverage and style, handling variation and evolution of sublanguages, "legislating" lexical semantics and assuring a language model able to accomodate new text types and support spoken output. 1 Background and System Overview FoG (for Forecast Generator) was developed during 1985-89 (Kittredge et al., 1986; Bourbeau et al., 1990). After tests at Environment Canada during 1989-91, FoG entered regular use during 1991-92, first for marine forecasts, and more recently for public forecasts. Forty percent of the operational marine forecasts (roughly half of all marine forecast text) in Canada is now produced using FoG. Meteorologists have been very receptive to using FoG, which is now a "back-end" facility of the FPA graphics workstation. The FPA supports the graphical analysis of weather while providing the rulebased concept formation needed to drive both text generation and non-linguistic applications. Meteorologists now concentrate on weather analysis and give less thought to how forecasts should be verbalized. Still, it has taken much time and effort to fit text generation into their work environment, and respond to new requirements. Operational experience has shown that some linguistic refinements first proposed during design were of low priority to users, compared with other features which were not originally anticipated. Early work on FoG set up a specialized sublanguage grammar for marine forecasts, based on analysis of more than 100,000 words of archived English text. Corpus analysis of each forecast type has been supplemented by interviews and other means to check validity of categorizations and rules. In contrast to earlier work on machine translation (TAUMMETEO), where sublanguage grammars had to be relatively complete to recognize each possible human input text, generation of forecasts from concepts provides an opportunity to "engineer out" infrequent words and sentence patterns as long as each intended text content coming into the text planner is sayable in good quality text. The simplification and regularization of sublanguage grammars raises questions for engineering design. We are no longer just engineering the system to fit the sublanguage, but also engineering the output sublanguage itself to achieve goals such as simplicity (without significant loss of expressiveness) and clarity. Different types of forecasts have differences in word usage, grammatical patterns and text structure, but the similarities are strong enough for them to be treated withing the same grammatical framework. This means that they require different text planners (and lexicons), but can use similar grammatical realizers. Significantly, English and French forecasts issued in Canada use the same principles for determining sentence boundaries, ordering and combining clauses, and formatting the output text. This greatly simplifies the problem of bilingual forecast generation in FoG. FoG uses three major stages to compose forecasts: (1) graphically mediated content determination, (2) text planning resulting in interlingual forms, and (3) realization of English and French texts from the interlingua. Details are given in (Kittredge and