INTELLIGENT HANDLING OF WEATHER FORECASTS 
Stephan Kerpedjiev 
Institute of Mathematics 
Acad. G. Bonchev St., bl.8 
1113 Sofia, BULGARIA 
Veska Noncheva 
Laboratory of Applied Mathematics 
No Vapzarov St., 15 
4000 Plovdiv, BULGARIA 
ABSTRACT 
Some typical cases of intelligent 
handling of weather forecasts such as 
translation, visualization, etc. are 
decomposed into two subprocesses ~ 
analysis and synthesis. Specific 
techniques are presented for analysis 
and synthesis of weather forecast 
texts as well as for generation of 
weather maps. These techniques deal 
with the weather forecasts at 
different levels ~ syntactic, 
discourse and semantic. They are based 
on a conceptual model underlying 
weather forecasts as well as on formal 
descriptions of the means of 
expression used in particular natural 
and cartographic sublanguages. 
I. INTRODUCTION 
Weather forecasts (WF) are the 
subject of various manipulations. 
Evolving as sets of numerical data and 
qualitative estimations they grow into 
complete images of expected 
meteorological situations represented 
in the form of texts in NL, weather 
maps, data tables or combined 
information objects. Then the WF could 
be translated from one language to 
another, transformed from verbal to 
pictorial form or vice versa, etc. 
Such kind of manipulations are often 
performed by forecasters, translators, 
editors in order to obtain better 
representation of the WF from the 
point of view of its perception and 
understanding. 
There is some experience with such 
kind of automatic handling of differ~ 
ent classes of specialized texts, The 
HETEO system \[7\] translates WFs from 
English to French. It is developed on 
the basis of the TAUM system. A tech~ 
nJque for creating computer animation 
scenarios of storie is described in 
\[6\]. The SOCCER system \[1\] interprets 
video recordings of soccer games and 
comments on them in NL. 
The purpose of the present project 
is to develop techniques for various 
manipulations of WFs including transla- 
tion from one natural sublanguage into 
another, transformation from textual 
to cartographic form, retelling WFs in 
the same NL but according to other 
discourse structure° Each of these 
manipulations can be decomposed into 
analysis of the source WF and 
synthesis of the target WF. These two 
processes are mediated by a uniform 
internal representation which is 
language independent. The example in 
Fig.l illustrates the basic processes 
and objects involved in the project. 
2.CONCEPTUAL MODEL 
Our conceptual model was created on 
the basis of both conceptual analysis 
of WFs from mass media and the 
recommendations for formulating WFs 
given in \[2\]. 
In our conceptual model each WF is 
a set of assertions having the form 
( W , T , R ). 
TRANSLATIONN ~=>~= ~=~= .......... ==~ ................ '=~ Today mostly sunny 
~.~i.~ ~-~- ...... weather, flore sig- 
~IHeo le ,peoG,~a-| I(CLOUDINESS=SUNNY, ness expected in 
~aBa C~bHqeBO ~ I TIME=TODAY, _ I SYNTHESIS East Bulgaria with 
BpeMB. (lo-~Ha~i-~ I REGIDN=BUE.GARIA) ~some showers in 
T{;~Ha 06~aNHOCT L_ANALYSIS ~(OL~\[4ESS=SIGNIF, i the afternoon. 
me MMa B MJTO'IHa ......... a TIME=TODAY, I , 
cAe~ o6~ ~e I I( PRECIPITATION=RAIN, I SYNTHESIS 
npeBaA~. I I TIHE=TODAY.AFTERNOON,~___~- 
-.. _J I REGION:EAST_BUL) j .--, .-. 
i,,, . ........... -___I_____A f <__... ./ ..... --. 
VISUALIZATION kl.. 0 .... /< )'77" \[-" 
Fig.l. Different types of transformation can be represented 
as a composition of analysis and synthesis 
379 
It is read as "Weather characteristics 
W are expected in the region R during 
the period l". For each of the items 
W, T, R a taxonomy is created that 
defines the quantities used together with 
their scales. Each taxonomy rep- 
resents a particular submodel. The 
present model does not reflect the 
degree of uncertainty in WFs. 
The internal representation of any 
WF is a consistent set S of assertions 
formulated in terms of the conceptual 
model. 
3. TEXT ANALYSIS 
Analysis is the process of 
transforming the text of a given WF into its 
internal representation. It 
consists of the following steps: 
data extraction -- identifying 
tl~e strings in the text that represent 
the values of the assertion items 
(such strings are called fillers of 
the corresponding quantity); 
- data translation -- representing 
the fillers in terms of the conceptual 
model; 
- ellipsis resolution -- completing 
elliptical phrases with information 
extracted from their contexts; 
- inconsistency elimination -- re- 
placing contradictory assertions by 
non-contradictory ones. 
3.1. Data extraction 
The technique for data extraction 
is based on a formal description of 
the typical phrases of WFs by means of 
transition networks called analyzers 
\[3\], The conditions along the arcs of 
the network are checks for certain 
semantic primitives defined in a 
dictionary or comparisons with certain 
strings. Some of the nodes are marked 
by two types of symbols mb(x) and me{x), 
where x is the identifier of a 
quantity in the conceptual model, mb 
means that a fiiler of x begins, me 
marks the end of the filler of x. 
3.2. Data translation 
Those entries of the dictionary 
that may occur as fillers have 
sections where the meaning of the 
filler is defined by the corresponding 
term of the conceptual model. Thus, 
the data translation process consists 
of looking up in the dictionary for 
the current filler, picking out the 
contents of the corresponding section 
and replacing the filler by the term 
extracted. Numerical values such as 
temperature degrees are calculated 
automatically from the text. 
3.3. Ellipsis resolution 
A good deal of sentences in WFs are 
e11i@tical. For example, consider the 
WF in Fig.l, where the region is 
missing from the first assertion and 
tlle time period is not specified in 
the second assertion. In order to 
complement them a context analysis is 
performed. It is based on certain 
heuristics represented as rules. For 
the particular case of medium-range 
forecasts in Bulgarian such rules are 
given in \[4\] where four variables are 
used, namely a short--term and a 
long-term backgrounds of both the 
time-period and the region. The rules 
specify the conditions under which 
these variables change their values as 
well as the variable whose value 
should fill in the current elliptical 
assertion. 
3.4. Inconsistency elimination 
TWO assertions (W~ , ~ ,R 4) and (~,~,R z) 
are said to be 
contradictory if they predict 
inconsistent weather characteristics 
(e.g. sunny weather and overcast) for 
overlapping regions and time periods. 
The following heuristic is applied to 
eliminating contradictory assertions. 
The more specifically the information 
is given, the higher priority it has. 
In terms of the conceptual model it 
means that if W i and are 
~nconsistent, O,c~ and 4c~en the 
~econd assertion is replaced by the 
assertion (W z,~-~,~-~), where &-5 
and Rz-Riare calculated on the basis 
of the corresponding submodels. 
4. TEXT SYNTHESIS 
Synthesis is the process of trans= 
forming the internal representation of 
a given WF into a text in a certain NL 
and according to a certain scheme 
(discourse structure). It consists of 
the following stages: 
- generation of words and phrases 
representing the items of the 
assertions; 
- generation of sentences 
describing the assertions; 
generation of text fragments 
containing all the weather information 
related to a certain region, time 
period or weather phenomenon depending 
on the scheme adopted. 
The three generators are called 
phrase generator, sentence generator 
and discourse generator, respectively. 
They are presented in detail in \[5\] 
for the case of short-range forecasts 
in Bulgarian. 
4.1. Discourse generator 
The process of generation is a 
Lop-down one. First, the assertions 
are arranged according to the 
discourse structure, defined 
preliminarily in terms of the 
conceptual model. For example~ one 
feasible discourse structure is the 
requirement to present the WF in four 
sections containing information for 
the regions North, East, South and 
West Bulgaria, respectively. In each 
section the information should be 
arranged according to weather 
phenomena in the following order: 
cloudiness, precipitation, wind and 
380 2 
temperature. The discourse generator 
scans the internal representation and 
sorts out the assertions into the 
predefined sections, whereby some 
~ssertions (e.g. those referring to 
the whote country) may fall into more 
than one section. Inside each section 
the assertions are reordered by their 
~eather characteristics. Other discourse 
~tructures can be processed as well. 
4,2. Sentence generator 
The generation of simple sentences 
representing single assertions and 
~ompound sentences representing groups 
of assertions is performed according 
to certain rules. The simple sentence 
~)tructure is defined by means of 
patterns. The choice of the actual 
pattern is made on the basis of the 
values of the current assertion. 
Compound sentences are used when 
certain patterns of assertions occur 
:in the final form of the internal 
representation. 5ix specific rules for 
constructing compound sentences are 
defined, one of them is shown below: 
(w,, R, =BUt) ===> °wI en" 
4.3. Phrase generator 
After the sentence structure has 
been settled, the phrase generator is 
invoked to form the phrases describing 
the assertion items. It replaces the 
terms of the internal representation 
by corresponding phrases ~n NL. For 
this purpose a reversed dictionary is 
u~;ed. 
5. MAP GENERATION 
Map generation is the process of 
transforming the internal representa- 
tion of a glven NF into a weather map. 
WFs are expressed on the map by means 
of stylized depictions as pictograms 
and symbols. The form and possibly the 
color of the depictions reflect the 
characteristics of certain weather 
phenomena while the positions of the 
depictions on the map correspond to 
the regions where these phenomena are 
expected to happen. There are no 
established rules for representing 
time on the map. The most simple 
solution from such a point of view 
consists of creating different maps 
for each of the periods of a given 
scale. More sophisticated solutions 
provide with specific means of time 
expression such as animation, symbolic 
representation of processes, explicit 
time notation of the phenomena being 
represented, etc. 
The technique for map generation 
m~ployed in this project consists of 
two steps: 
- time analysis of the internal 
ri~presentation resulting in a proper 
division of the assertions into groups 
according to the model adopted for 
t~me representation; 
scanning tho assertions of each 
group and translating them into depic. 
tions corresponding to the weather 
items and arranged in some positions 
on the map depending on the regions. 
The arrangement is based on a 
preliminarily designed template of the 
map. The template contains the spots 
where depictions could potentially 
appear° Together with the template a 
mapping M is defined. For any pair' 
(W,R) fl determines the set of spots 
where the corresponding depiction is 
to be placed. The depictions used are 
prepared in advance through a 
specialized graphical editor. 
6. CONCLUSION 
The three techniques presented here 
have been implemented on an IBM PC/AT 
microcomputer with EGA monitor for the 
case of medium-range WFs for Bulgaria 
issued by the Office of Hydrology and 
fleteorology in Sofia. 
The technique for text analysis was 
estimated by analyzing 100 WFs from 
Bulgarian newspapers. About 85-90~ of 
the data items were extracted and 
translated correctly by the system. 
The subsystems for text synthmsis and 
map generation produce WFs of 
acceptable quality. 
ACKNOWLEDGEMENTS, This project was 
supported by the Presidium of the 
Bulgarian Academy of Sciences (grant 
No 1001003) and by the Committee for 
Science and Higher Education (grant No 
607). The authors would like to thank 
Peter Barney for his encouragement and 
Ivan Bosov who helped in the implemen- 
tation of the map generator. 

REFERENCES 

I.E. Andre, G. Herzog and Th. Rist, On 
the Simultaneous Interpretation of 
Real World Image Sequences and their 
Natural Language Description: the 
System SOCCER, Proc. ECAI 88, p.449. 

2.E.M. Dobryshman, Review of Forecast 
Verification Techniques, World Mete- 
orological Organization, Technical 
Report No 120, Geneva, 1972. 

5.S. Kerpedjiev, A Technique of Auto- 
matic Data Extraction from Special- 
ized Texts, SERDICA Bulg. math. 
public. 13 (1987), p. 239. 

4.S. Kerpedjiev. Transformation of 
Weather Forecasts from Textual to 
Cartographic Form, Computing Tech- 
niques in Physics, Proo. 8th Euro- 
pean Summer Schools,1989, p. 212. 

5.V. Noncheva, Synthesis of a weather 
forecast text. Mathematics and Educa- 
tion in Mathematics, Proo. 19th Conf. 
of Union of Bulg. Math.,1990, p.269. 

6.H.Shimazu, Y.Takashima and M. Tomono, 
Understanding of Stories for Anima- 
tion, Proo. Coling B8, p. 620. 

7.B.Thouin, The METEO System, in: Prac- 
tical Experience of Machine Transla- 
tion, ed. V. Lawson,p.39. 
