Tuesday, 11 January 2000: 9:14 AM
Weather forecasts are made to be communicated. Natural language is a basic element of human discourse, and is especially effective for communicating summarized explanatory information. However, much of the forecast data in a weather office is visualized in chart form. Extracting information content from a picture and expressing it in flowing sentences is so far a skill dominated by humans. For many years, a Natural Language Generation (NLG) project associated with the Forecast Production Assistant (FPA) has attempted to use computational linguistics to accelerate the text communications output of weather forecasters. Out of several prototypes, a bilingual (English and French) marine forecast writing program called FoG (Forecast Generator) achieved operational implementation in the early 1990's in Canada.
A new marine forecast generator has been developed to replace FoG. A significant change is a move to simple ad hoc grammar rules to achieve linguistic rendering. This change involves a tradeoff between grand linguistic theory and the need for easy external configurability. Maintenance is reduced because decisions about locations and lexical choice can be made by administrators in the field with a minimum of training. With this approach, it should be possible to transplant pre-encoded styles while accommodating modest regional variations. Far from repudiating linguistic theory, the new text generator draws on current debate in the NLG community concerning deep vs. shallow grammars.
![](images/flag.png)
![](images/blue_ribbon.png)