Weather forecasters could find themselves pushed out of a job by artificial intelligence (AI) systems writing human-quality weather forecast texts that have been shown to be preferred by expert forecast readers.
Computer scientists at the University of Brighton have pitted experienced human weather forecasters against computer programmes in a series of experiments designed to test whether expert forecast users such as master mariners prefer human-authored forecast texts or computer-generated ones. The scientists found that despite the generators acquiring all decision-making abilities automatically, their weather forecast texts were judged more appropriate and of higher language quality than forecasts written by experienced meteorologists.
The team developed statistical natural language generation (NLG) systems which convert output data from weather simulators into readable human language. The experiments focused on the part of the forecasts that predicts wind characteristics for the next 15 hours including wind direction, wind speed and gust speed. Five different statistical NLG systems were used to generate forecasts and evaluated in terms of output quality, development time and computational efficiency against human-authored forecasts and traditional (non-statistical) NLG systems.
The scientists found that the statistical systems performed on a par with the traditional hand-crafted system, despite taking a fraction of the time to develop, and that the forecast texts produced by the best statistical system and by the handcrafted system were both judged to be more appropriate overall and of higher language quality than a comparable set of forecasts written by experienced meteorologists.
The research was led by Dr Anja Belz who said:
While these results are exciting they were obtained in the weather forecast domain which is in some ways ideal for automatic language generation, because the language of weather forecasts tends to be very simple in structure and vocabulary. Future research will have to demonstrate the feasibility of statistical language generation systems for more complex domains.
Building language generation systems has traditionally been a time-intensive and expensive business. However, next-generation statistical language generators take much less time to develop, and can be more readily reused and adapted. This makes the commercial deployment of language generation applications much more feasible.
Contact: Marketing and Communications, University of Brighton, 01273 643022