Autotext

This paper will explain how we plan to do, and to some extent already do, to automatically generate a text representation of weather information from a forecast database or directly from a NWP field information.

Introduction The reason to start this project is that we have realised that automatic production of weather information is demanded in order to keep a cost beneficial production. The NWP are getting better and more detailed but the text generators are not getting better in the same pace. We have run in to problems trying to improve the old type of text generators to coop with the needs that we have. We must think along new trails and leave the canned type of text generators behind to history. Our strategy is to split the problem of creating text into several modules. These modules are taking care of subproblems and can be developed more or less independently of each other.

The modules that we are working with are:

The lingual module This module is taking care of the linguistic job of creating full sentences with correct grammar in different languages. For the time being we only use English and Swedish. In this module we convert information from grammatical functional representation into fluent language.

The fact- extraction module The modules extract gridpointed state of atmosphere parameters in time sequences. The module then converts these data to objects. The objects we are looking for are of a kind that can carry weather information to human receivers. I can try to describe the objects as follows. The objects are actors acting on the stage playing a role in a play that tells the weather story. Each object has some attributes and possibilities to be connected to other objects.

The expert module This module selects what objects to look for in certain weather situations to serve a specific kind of consumer with its specific profile of interest and language requirements.

A subpart of the expert module considers how to express the obtained objects. As we all know we can express our knowledge about a certain weather object in many different ways. A second subpart of the expert module is taking care of the co-ordination between different objects. First it takes care of the order in which it should be mentioned and then in some cases if it is possible it tries to express objects reflecting to each other in one sentence.

A third subpart of the expert module is taking care of the time setting in the given information. The time description depends on several various facts like when the forecast is issued, how many changes are being described during the length of the forecast . Also the duration of a change is important to consider when choosing time description.

Up to now till the 1st of May 99 we have implemented parts of some modules in a test environment and some results will be presented during the conference.