L. Bourbea.u, D. Carcagno, E. Goldberg,
R. Kittredge and A. Polgu~re
In 1986 the first experiments in text generation applied to weather forecasts resulted in a prototype system (t~AREAS\[6,3\]) for producing English marine bulletins from forecast data. Subsequent work in 1987 added French output to make the initial system bilingual (RAREAS-2111\]). During 1988-1989 a full-scale operational system was created to meet the needs of daily marine forecast production for three regional centres in the Canadian Atmospheric Environment Service 1. In contrast o the earlier systems, the most recent one uses general models for both text planning and sentence realization (see sections 4 and 5 below). This new implementation, dubbed FoG for Forecast Generator, may constitute one of the first "indu'~trial" uses of text generation. FoG is of interest to computational linguists for three additional reasons: ,* the conceptual input to the text generation
process is derived from data that also drive a
graphic display on a workstation for forecasters;
this determination of text frora a selected sub-
set of graphically displayed data represents an
important paradigm for the transformation of
information; • conceptual processing results in an "interlingual"representation, a kind of deep syntactic structure for both English and French in this sublanguage; • sentence generation is carried out using a "streamlined" version of the Meaning-Text linguistic model; this may represent he first time that such a general model has been adapted to the descriptive problems arising in telegraphic sublanguages. i
Operational meteorologists normally work with graphical representations of the information available to them. "Charts" are used to display the large volumes of observational data and also the results of global simulations of the atmosphere. The graphical entities displayed on these charts (such as weather fronts, and low pressure systems) are manipulated to adjust for more recent data, and for perceived errors in the simula£ions. This results in manually created weather depictions which are valid at some future time (24 to 36 hours in the future). The weather situation is always being monitored and updated as new information is received and assimilated. During the normal course of events, much of the communication between forecasters i done using these charts. When it is time to write a forecast for some user community, the forecaster has to extract the pertinent information from these charts and recast them into a structured text. In addition, the 'primary' information taken from the charts has to be modified for local geographic effects. The forecaster appears to do this while the text is being composed. This mental transposition of meteorological information from graphical to text form is believed to be open to a number of subjective errors. In addition, the pressure to compose text often conflicts with the scientific demands of analyzing an emerging weather situation. FoG is part of the recently implemented Forecast Production Assistant (FPA)\[7\], which uses interactive computer graphics to Mlow the meteorologist to view and edit a display of the weather situation. All of the fields produced by the large scale computation are directly available on the FPA together with any manually produced products. This makes it possible to obtain numerical values directly from the charts and to use them in other applications uch as FoG. All of the fields required to produce forecast ext can be obtained from the computer graphics.
A sampling procedure is used to determine values of these fields at specific latitude and longitudes which have been pre-selected as being representative of weather conditions over a specified geographic area. Computer animation techniques are used to interpo~ late between the standard chart times (normally ew. cry 1.2 hours) to whatever time resolution is required for the text product. Currently, charts are available at intervals of three hours through the forecast period. The problem is that this yields nine values for a 24 hour forecast. Practical considerations limit the number of events (e.g. shifts in wind speed or direction) in a forecast o three or four, depending on the severity of the weather. The conceptual phase of the processing treats the sampled data so that only the significant events in time and space are passed on to textual encoding. Conceptual processing involves several stages. 1) Events requiring "weather warnings" are identified and stored before any data smoothing is done, 2) Sampled data is smoothed with respect to time so that only the significant weather changes are retained, 3) Spatial smoothing is done so that areas sharing similar weather conditions can be grouped together in the text. We have noticed, however, that the notion of "significance" is partly dependent on the ability of the lexicon of the forecast language to make semantic distinctions. Thus, a wind change of 30 degrees is more likely to be judged significant when it crosses the boundary between, say, northeasterly and easlerly, than when it stays entirely within the range of one of these terms. The semantic granularity of temporal adverbs has a similar "anticipatory" effect on the way generalizations are made over tirne. This constitutes akind of filter on content determination that precedes formal text planning.
Text planning in l, bG consists of three stages: content determination, text structuring and interlingua production. Content determination covers the problems of (1) converting the smoothed data on significant meteorological events into complex objects appropriate for inferencing, one object for each meteorological event of interest, and (2) using the structured data objets to compute additional concepts needed to talk about transitions between weather events. ':\['he output of content determination is, for each !Forecast area, an enriched data object called a "text content representation". Text structuring consists basically of finding the optimal way of cutting each text content represents4 tion (the conceptual representation corresponding to one future text) into sentencc~sized chunks of infor~ mation ("sentence partitioning") within the complex text structure. The chunks are then linearly ordered according to principles that are sometimes domainspecific, but often more general (e.g., temporal sequence). There is a subsidiary problem of making hill or partial copies of certain concepts to assure continuity of reference between consecutive sentences. The output of the text structuring process gives, fbr each forecast area, a partitioned and possibly enriched structure called the "text representation". The final stage in text planning involves converting the single partitioned text representation i to an a(> tual sequence of conceptual representations for individual sentences. The strong similarity between forecast styles and structures used in Canadian marine forecasts in the two official languages makes it possible to formulate a single interlingual structure, which can map quite directly to the "deep" syntactic structure of the corresponding sentence in either English or French. The primary issue here is the identity of information conveyed in the two parallel sublanguages, and the fact that sentence scoping may be performed in identical ways on the text con-tent representations used for English/I~'ench. There is no guarantee that such an interlingua would suffice for a language using a very different conceptual system or communication style for weather phenomena (e.g., hmktitut).
Realization The last part of forecast generation involves the relatively well-developed technique of sentence realization. By this we mean the conversion of interlingual representations ofEnglish/French sentences into acceptable word strings in one or tile other of these two languages. To guarantee generality and long-term flexibility of linguistic modelling, we have chosen to use the Meaning-Text linguistic theory of Mel'auk\[8\], which has also served as the framework for text generation in other technical sublanguages\[5,4\]. Because of the lack of semantic paraphrase in the forecasting sublanguage considered, however, we have eliminated the semantic net representations from the processing stages, passing directly from interlingual representations to deep syntactic dependency trees\[10\]. We have implemented a fragment of an existing Meaning-Text model for English\[9\] and adapted this model for 15"each. 319
FoG is written in Quintus Proleg and runs on a Hewlett-Packard 9000 workstation as part of the FPA system. The graphics software on the FPA workstation is programmed mostly in C. As of April 1990, the entire FPA is undergoing testing by three regional weather centres in Eastern Canada and is being phased into daily production, initially during one of the three daily work shifts.
Future Plans Since FoG is now configured only to produce marine forecasts for the Halifax, Gander and Great Lakes regions of Canada, an early priority is to adapt the software (mostly the text planner) to the different content and style of forecasting found in Pacific Canada and other marine regions, and to specialized marine forecasts (e.g., for small craft). Concurrently, investigation should continue into extending the system to other forecast ypes, including agricultural and general public forecasts. We expect that our linguistic model will also facilitate the addition of high-quality voice output as an option at some future time. References \[1\] Carcagno D. (1989) Text Planning for Gener-
ation of Weather Forecasts, Technical report,
Odyssey Research Associates, Montreal. \[2\] Carcagno D. and Iordanskaja L. (1989) "Con-
tent Determination and Text Structuring in
GOSSIP", Extended Abstracts of the Second Eu-
ropean Workshop on Nalural Language Genera-
tion, Edinburgh. \[3\] Goldberg E., Kittredge R. and Polgu~re
A. (1988) "Computer Generation of Marine
Weather Forecast Text", Journal of Atmo-
spheric and Oceanic Technology, vol.5, no.4,
pp.473-483. \[4\] Iordanskaja L., Kittredge R. and Polgugre A.
(1988) "Implementing aMeaning-Text Model
for Language Generation" supplement to Proc.
of the 12th Inll. Conf. on Computational Lin-
guistics, Budapest. \[5\] Kittredge R., Iordanskaja L. and Polgu~re A.
(1988) "Multi-Lingual Text Generation and the
Meaning-Text Theory", Proc. of the 2nd Inter-
national Conf. on Theoretical and Methodolog-
ical Issues in Machine Translation of Natural
Languages, Carnegie-Mellon University. 320 \[6\] Kittred~e R., Polguhre A. and Goldberg E.
(1986) "Synthesizing Weather Forecasts from
Formatted Data", Proc. of the 11th Intl. Conf.
on Computational Linguistics, Bonn. \[7\] McLeod J.C. (1990) "The Forecast Production
Assistant" Preprinls Sixth International Con-
ference on Interactive Information and Process-
ing Systems for Meteorology, Oceanography and
Hydrology, Anaheim CA, Amer. Meteor. Soc. \[8\] Mel'6uk I. (1981) "Meaning-Text Models", An-
nual Review of Anthropology, vol.10, pp.27-62. \[9\] Mel'0.uk I. and Percov N. (1987) Surface Syntax
of English, Benjamins. \[10\] Polgu~re A. (1988) A Dependency Grammar
for Marine Weather Forecasts, Technical report,
Odyssey Research Associates, Montreal. \[11\] Polgu~re A., Bourbeau L. and Kittredge R.
(1987) RAREAS.2: Bilingual Synthesis of Arc-
tic Marine Forecasts. Technical report, Odyssey
Research Associates, Montreal.