Synthesizing Weather Forecasts from Formatted Data

This paper describes a system (RAREAS) which synthesizes marine weather forecasts directly from formatted weather data. Such synthesis appem~ feasible in certain natural sublanguages with stereotyped text structure. RAREAS draws on several kinds of linguistic and non-linguistic knowledge and mirrors a forecaster's apparent tendency to ascribe less precise temporal adverbs to more remote meteorological events. The approach can easily be adapted to synthesize bilingual or multMingual texts. 1. Natural Language Report Synthesis We use the term "natural language report synthesis" (NLRS) to describe the process of creating well-formed text which summarizes formatted data in a given domain using a style which mirrors the conventions of professional report writers for that domain. NLRS for highly reslricted domains was first demonstrated in the work of Kukicb_ (1983) on "knowledge-based generation" of stock market reports. Kukich's ANA system produces professionalsounding stock market stmlmaries using a daily trace of Dow Jones' half-hourly quoudions for the market average and major indices. Both ANA and the analogous FRANA system for French (Contant 1986) have used a phrasal lexicon approach (Beeker 1975) which limits the generality of the linguistic component, but which seems to suffice for small and stereotyped domains. The work described below represents a more modular approach to NLRS as well as a new application domain. 2. Synthesis of Arctic Marine Weather Forecasts The RAREAS system was developed during a five-month effort to explore the feasibility of synthesizing marine weather bulletins from formatted weather forecast data. The particular task was to produce Arctic marine forecasts for five forecast areas to the east of Baffin Island (known as FPCN25 forecasts). Marine forecasts are one of several types of weather bulletin based on the same basic weather data, each type emphasizing the conditions of interest to a particular corurrmnity of users. In the case of marine bulletins, linguistic emphasis is placed on wind direction and speed, dangerous wind and freezing spray conditions, etc. RAREAS is designed to be sufficiently modular and flexible so as to allow easy extension and adaptation to other types of weather bulletin (e.g., agricultural bulletins, public weather forecasts). Although the current project seems to have proved the feasibility of automatically synthesizing weather forecasts, extensive testing and refinement is required before RAREAS or any successor can be introduced into daily use. The RAREAS system is the natural language component of the MARWORDS project, which envisages automating the process of creating bulletins from meteorological information. In the current manual procedure all the available meteorological information (observations, radar and satellite imagery, and numerical weather prediction products) is made available to the weather forecaster. The weather forecaster must correctly diagnose the meteorological processes which will affect his particular area of interest throughout the forecast period, and then translate this knowledge into appropriate textual forecasts for various users. In the proposed automated process, MARWORDS will use predicted values for meteorological parameters such as wind speed and direction, cloud cover, and others. In some cases, these predictions could be obtained directly from numerical weather prediction products. In most cases though, they would still be the result of a manual (i.e., human) forecasting procedure. MARWORDS will significantly reduce the workload on the forecaster, making it possible to focus more attention on meteorological problems. In the normal course of events, the predicted values make up a continuum in both time and space. For simplicity, values are often given at regular steps in time (e.g., hourly) and space (either at grid points, or at weather observing sites). Alternatively, forecast parameters may be given in terms of significant changes only. MARWORDS is flexible enough to accept both types of data description. In fact, the structure and nature of the required data is a problem which needs more work to resolw~. 3. Design of the RAREAS system A major task in designing RAREAS was the definition of an input data format which properly divides the work between the MARWORDS expert system, which computes predicted values of weather parameters based on large-scale observations, and RAREAS itself, which interprets that data under local conditions for the purpose of marine forecasts. The format and its permissible content should be sufficiently rich in expressive power to reflect the nuances found in natural language forecasts. Ideally, the expert system should be kept as independent of forecast purpose as possible. RAREAS should therefore take care of all matters related to subjective evaluation of the data (e.g., importance of individual parameters of marine forecasts), as well as the linguistic expression of data values and data relations. In its current implementation RAREAS reads the formatted forecast data and carries out (sequentially) the following major operations: reading and parsing of formatted input data, with the interpretation of certain coded values; checking of data for consistency and plausibility, using databases of geographical and meteorological information; insertion of default values when needed; detection of conditions which are hazardous for marine operalions (e.g., freezing spray, calculated as a function of forecast wind speed and air temperature, and of a seasonally and regionally adjusted water temperature taken from the database); "merging of areas", namely, a check for similarity in the data for contiguous forecast areas; when similarity threshold conditions are satisfied a single report formula is created for the merged areas under a header which lists those areas;