Data Mining for Automatic Linguistic Description of Data
Textual Weather Prediction as a Classification Problem
J. Janeiro, I. Rodriguez-Fdez, A. Ramos-Soto and A. Bugar´ın
CITIUS, University of Santiago de Compostela, Santiago de Compostela, Spain
Keywords:
Linguistic Descriptions of Data, Natural Language Generation, Weather Forecasting, Classification.
Abstract:
In this paper we present the results and performance of five different classifiers applied to the task of automat-
ically generating textual weather forecasts from raw meteorological data. The type of forecasts this method-
ology can be applied to are template-based ones, which can be transformed into an intermediate language
that can directly mapped to classes (or values of variables). Experimental validation and tests of statistical
significance were conducted using nine datasets from three real meteorological publicly accessible websites,
showing that RandomForest, IBk and PARTare statistically the best classifiers for this task in terms of F-Score,
with RandomForest providing slightly better results.
1 INTRODUCTION
Weather forecasting has been one of the most sci-
entifically and technologically challenging problems
around the world in the last century. To make an ac-
curate prediction is one of the major challenges mete-
orologists face on a daily basis. Weather forecasts are
made by collecting quantitative data about the current
state of the atmosphere on a given place and using
scientific understanding of atmospheric processes to
project how the atmosphere will evolve on that place.
Modern weather forecasting is largely based on
numerical weather predictions (NWP), which essen-
tially are massive atmosphere simulations run on su-
percomputers. The output of NWP models is a set of
predictions of meteorological parameters or variables
(wind speed, temperature, precipitation, etc) for vari-
ous spatial locations and at various points in time.
Weather forecasting organizations take NWP data
and modify it according to their local knowledge and
expertise. They also interpolate between the locations
in the source NWP model, again using local knowl-
edge and expertise. The result is a modified set of
predicted numerical weather values, for locations of
interest to their customers.
Initially, the NWP data was used by expert mete-
orologists to manually describe the weather forecast
using texts for different places. With the increasing
accuracy of predictions and the need to generate tex-
tual forecasts for a large number of locations, weather
forecasting organizations require solutions which au-
tomatically build these texts.
There are several official meteorological agen-
cies that offer weather forecast services, such as the
Spanish AEMET (AEMET, 2014), American NWS
(NWF, 2014) or the British Met Office (MetOffice,
2014b). Other private organizations like Weather-
Forecast (WeatherForecast, 2014) or Intellicast (Intel-
licast, 2014) offer their own forecast services. Some
of them provide forecast data for specific domains,
such as skying or surfing, allowing users to find the
best conditions in which to perform this kind of ac-
tivities. Furthermore, due to the need to provide tex-
tual forecasts to an increasing number of locations,
some meteorological agencies started offering auto-
matically generated forecast texts. For instance, in the
1990s, NLG systems such as FoG (Goldberg et al.,
1994) and MultiMeteo (Coch, 1998), were used by
meteorological agencies to provide this kind of infor-
mation services. More recently, the Met Office with
Data2Text (MetOffice, 2014a) or the Galician Meteo-
Galicia with GALiWeather (Ramos Soto et al., 2014)
are also employing this sort of technology to address
the creation of textual forecasts for increasing quanti-
ties of localized data.
Several techniques can be used for automated gen-
eration of weather forecast texts. These techniques
can be divided into two broad categories: knowledge-
intensive (KI) and knowledge-light (KL) approaches
(Adeyanju, 2012). KI approaches require extensive
consultation with domain experts during data analysis
and throughout the text generation approach devel-
556
Janeiro J., Rodriguez-Fdez I., Ramos-Soto A. and Bugarín A..
Data Mining for Automatic Linguistic Description of Data - Textual Weather Prediction as a Classification Problem.
DOI: 10.5220/0005282905560562
In Proceedings of the International Conference on Agents and Artificial Intelligence (ICAART-2015), pages 556-562
ISBN: 978-989-758-074-1
Copyright
c
 2015 SCITEPRESS (Science and Technology Publications, Lda.)

opment process. On the other hand, KL approaches
rely more on the use of automated methods which are
mainly statistical.
The earliest KI systems generated forecast texts
by inserting numeric values in standard manually-
created templates. Multiple templates are created
for each possible scenario and one of them is ran-
domly selected during text generation to provide va-
riety. Other KI systems developed linguistic models
using manually-authored rules obtained from domain
experts and corpus analysis.
The KL approach to generate forecast texts typ-
ically employs machine learning techniques. Train-
able systems are built using models based on statisti-
cal methods such as probabilistic context-free gram-
mars and phrase based machine translation. The ad-
vantage is that systems are built in less time and with
less human effort as compared to the KI approach.
In this paper we consider forecast services with a
KI approach and use these templated textual forecasts
to obtain linguistic predictions presented as a classi-
fication problem to generate natural language (NLG)
descriptions. The paper is organized as follows: in
section 2 we present the problem and the different
types of automatic textual forecasts. In section 3 we
provide the steps needed to solve it using classifica-
tion techniques. In section 4 we explain the five dif-
ferent classification techniques tested, the results ob-
tained for each one and a statistical comparison be-
tween them and, finally, we present the most relevant
conclusions of this approach.
2 LINGUISTIC WEATHER
PREDICTIONS AS A NLG
PROBLEM
The generation of natural language text uses the NWP
data and additional expert information to generate
textual weather forecasts that are issued to the public.
There are two main approaches for generating textual
forecasts automatically (Van Deemter et al., 2005):
• Template-based systems are natural language gen-
erating systems that map their non-linguisticinput
directly to the linguistic surface structure. This
linguistic structure may contain gaps that must be
filled with linguistic structures that do not contain
gaps. For example, a template such as ”[amount]
rain at [time]”, where the gaps represented by
[amount] and [time], can be filled with informa-
tion from the data.
• Standard NLG systems, by contrast, use less di-
rect mapping between input and surface form.
These systems could start from the same input se-
mantic representation subjecting it to a number of
consecutive transformations until a surface struc-
ture results. Various NLG submodules would op-
erate on it, jointly transforming the representation
into an intermediate representation where lexical
items and style of reference have been determined
while linguistic morphology is still absent. This
intermediate representation may in turn be trans-
formed into a proper sentence in one of the avail-
able output languages.
The typical stages of natural language generation
systems (Reiter et al., 2000), are:
• Content determination: Deciding what informa-
tion to mention in the text.
• Document structuring: Overall organizationof the
information to convey.
• Aggregation: Merging of similar sentences to im-
prove readability and naturalness.
• Lexical choice: Mapping words to concepts.
• Referring expression generation: Creating refer-
ring expressions that identify objects and regions.
• Realization: Creating the actual text, which
should be correct according to the rules of syntax,
morphology, and orthography.
The texts generated by these two approaches usu-
ally have a similar structure, from which we can ex-
tract the main information and apply data mining
techniques to the raw data to generate the same fore-
casts. To achieve this, we applied classification algo-
rithms that learn the textual forecasts using data sam-
ples. In the next example, using the temperature data
values, we can learn the forecast text for the weekly
temperature:
• Full forecast: “Mostly dry. Warm. Mainly fresh
winds.”
• Daily Temperature values (
◦
C): 21, 22, 20, 19, 20,
18, 19
• Learned textual temperature value: “Warm”
3 LINGUISTIC PREDICTIONS AS
A CLASSIFICATION PROBLEM
From the two approaches for automatically generate
textual forecasts explained before, we selected the
template based forecasts since they are more abun-
dant and they have a more regular structure that allows
us to extract the relevant information from the text.
To test the classification of these forecasts we need
DataMiningforAutomaticLinguisticDescriptionofData-TextualWeatherPredictionasaClassificationProblem
557

to transform the textual forecast into a class, extract-
ing the relevant information and building descriptive
phrases. We selected three different datasets from the
web that offer NWP data and a descriptive, template-
based textual forecast. Then we transformed these
textual forecasts into classes and used them along
with the raw meteorological data to perform the clas-
sification.
3.1 Weather-forecast Dataset
Weather-Forecast (WeatherForecast, 2014) uses the
Global Forecast System from the National Oceanic
and Atmospheric Administration (NOAA) to get their
raw forecast data and use their own computers to
generate the actual forecasts. Their textual forecasts
include information about precipitation, temperature
and wind, as shown in the example that follows:
Mostly dry. Warm (max 29
◦
C on Tue afternoon,
min 23
◦
C on Wed night). Wind will be generally light.
From this forecast service we can extract three
datasets (one for each of the variables considered),
as indicated in table 1. The selected samples from
this service come from different locations worldwide.
Some examples of the classes considered are:
• Precipitation: “mostly dry”, “light rain”, “some
drizzle”, “moderate rain”.
• Temperature: “warm”, “very mild”, “freeze-
thaw conditions”.
• Wind: “generally light”, “increasing light to
fresh winds”, “mainly fresh ”, “decreasing fresh
to calm”.
3.2 National Weather Service Dataset
The National Weather Service (NWF, 2014) is a com-
ponent of the National Oceanic and Atmospheric Ad-
ministration (NOAA). They provide weather, water,
and climate data, forecasts and warnings for the U.S.
territory. Their textual forecasts include information
about precipitation, cloud coverage and wind:
A chance of showers, mainly before 11pm. Mostly
cloudy, with a low around 60. West wind 3 to 5 mph.
From this forecast we can extract three datasets as
indicated in table 2. The selected samples from this
service come from different locations on the United
States of America. Some class examples considered
from this dataset are:
• Precipitation: “chance showers”, “showers
likely”, “scattered showers and thunderstorms”,
“slight chance showers then slight chance show-
ers and thunderstorms”.
• Cloud coverage: “mostly cloudy”, “partly
sunny”, “mostly clear”, “sunny and hot”.
• Wind: “west”, “calm becoming west”, “north-
west becoming calm”, “west becoming north-
east”.
3.3 Intellicast Dataset
Intellicast (Intellicast, 2014) delivers site-specific
forecasts for 60,000 sites in the U.S. and around the
globe including detailed local forecasts to hurricane
tracks and severe weather warnings to international
conditions. Their textual forecasts include informa-
tion about cloud coverage, precipitation, temperature
and wind, for example:
Partly cloudy skies. Hot. High 93F. Winds WSW
at 10 to 20 mph.
From this forecast service we can extract three
datasets, one of them includes both cloud coverage
and precipitation information as indicated in table 3.
The selected samples from this service come from dif-
ferent locations worldwide. Some examples of the
classes considered are:
• Mixed cloud coverage and precipitation: “partly
cloudy”, “sunshine and clouds”, “partly cloudy
with thunderstorms”, “mix of clouds and sun with
the chance of isolated thunderstorm”.
• Temperature: “hot”, “warm”, “hot and humid”,
“very hot”.
• Wind: “WSW”, “light and variable”, ”S decreas-
ing”.
4 EXPERIMENTAL SETUP
We evaluated the performance of five different classi-
fication techniques for the three datasets introduced
previously using the data mining software “Weka”
(Hall et al., 2009). We selected these ones to test dif-
ferent types of supervised learning techniques which,
in general, provide comprehensible visual models.
Other techniques such as Artificial Neural Networks
were not considered due to its black box structure.
4.1 Classification Methods
The classification techniques applied are:
• J48 (Quinlan, 1993) is an open source Java im-
plementation of the C4.5 algorithm. C4.5 builds
decision trees from a set of training data. At each
node of the tree, C4.5 chooses the attribute of the
data that most effectively splits its set of samples
ICAART2015-InternationalConferenceonAgentsandArtificialIntelligence
558

Table 1: Datasets from Weather-Forecast (WF datasets).
Precipitation
Type Classification Origin Real world
Features 22 (Real / Integer / Nominal) (0/22/0)
Instances 83754 Classes 17
Missing values? No
Temperature
Type Classification Origin Real world
Features 23 (Real / Integer / Nominal) (1/22/0)
Instances 83754 Classes 5
Missing values? No
Wind
Type Classification Origin Real world
Features 15 (Real / Integer / Nominal) (1/13/0)
Instances 83754 Classes 42
Missing values? No
Table 2: Datasets from National Weather Service (NWS datasets).
Precipitation
Type Classification Origin Real world
Features 27 (Real / Integer / Nominal) (13/13/1)
Instances 93896 Classes 122
Missing values? No
Cloud coverage
Type Classification Origin Real world
Features 40 (Real / Integer / Nominal) (1/38/1)
Instances 93896 Classes 26
Missing values? No
Wind
Type Classification Origin Real world
Features 38 (Real / Integer / Nominal) (2/36/0)
Instances 93896 Classes 74
Missing values? No
Table 3: Datasets from Intellicast (ICAST datasets).
Mixed cloud coverage and precipitation
Type Classification Origin Real world
Features 66 (Real / Integer / Nominal) (4/48/14)
Instances 127118 Classes 631
Missing values? No
Temperature
Type Classification Origin Real world
Features 53 (Real / Integer / Nominal) (3/50/0)
Instances 127118 Classes 9
Missing values? No
Wind
Type Classification Origin Real world
Features 26 (Real / Integer / Nominal) (1/13/12)
Instances 127118 Classes 24
Missing values? No
DataMiningforAutomaticLinguisticDescriptionofData-TextualWeatherPredictionasaClassificationProblem
559

into subsets enriched in one class or the other. The
splitting criterion is the normalized information
gain (difference in entropy). The attribute with
the highest normalized information gain is chosen
to make the decision. The C4.5 algorithm then
recurs on the smaller sublists.
• RandomForest (Breiman, 2001) is an ensemble
learning method for classification (and regression)
that operates by constructing a multitude of deci-
sion trees at training time and outputting the class
that is the mode of the classes output by indi-
vidual trees. The training algorithm for random
forests applies the general technique of bootstrap
aggregating, or bagging, to tree learners. Given
a training set, bagging repeatedly selects a boot-
strap sample of the training set and fits trees to
these samples. After training, predictions for un-
seen samples can be made by averaging the pre-
dictions from all the individual regression trees or
by taking the majority vote in the case of decision
trees.
• IBk (Aha et al., 1991) is a K-Nearest Neighbours
classifier (k-NN). In k-NN classification, the out-
put is a class membership. An object is classified
by a majority vote of its neighbors, with the object
being assigned to the class most common among
its k nearest neighbors (k is a positive integer, typ-
ically small). If k = 1, then the object is simply
assigned to the class of that single nearest neigh-
bor.
• BayesNet (Bouckaert, 2004) uses a Bayesian net-
work for classification. A Bayesian network is a
directed acyclic graph that encodes a joint proba-
bility distribution over a set of random variables.
It is defined by the pair B ={G, θ} where G is the
structure of the Bayesian network, θ is the vec-
tor of parameters that quantifies the probabilistic
model and B represents a joint distribution PB(X),
factored over the structure of the network. The
goal of a Bayesian network classifier is to cor-
rectly predict the label for class given a vector of
attributes. It models the joint distribution and con-
verts it to a conditional distribution. The predic-
tion for a class can be obtained by applying an
estimator to the conditional distribution.
• PART (Frank and Witten, 1998) combines C4.5
and RIPPER to avoid their respective problems
since it does not need to perform global optimiza-
tion to produce accurate rule sets and this added
simplicity is its main advantage. It adopts the
divide-and-conquerstrategy in that it builds a rule,
remove the instances it covers, and continue cre-
ating rules recursively for the remaining instances
until none are left. It differs from the standard ap-
proach in the way each rule is created. In essence,
to make a single rule a pruned decision tree is built
for the current set of instances, the leaf with the
largest coverage is made into a rule and the tree is
discarded. This avoid hasty generalization by only
generalizing once the implications are known.
4.2 Classification results
We performed experimentation with a 10 fold cross-
validationusing the previously indicated classifiers on
the datasets considered. The results obtained in each
case are shown in tables 4, 5, and 6 for the F-Score,
since it provides better accuracy measure (Sokolova
et al., 2006), computed as a weighted average of the
precision and recall, and the root-mean-square error
(RMSE).
With the WF datasets (Table 4) we got results with
a score over 0.96 in all cases for the best classifier, that
was RandomForest.
With NWS (Table 5) we got similar results, close
to 1 for the precipitation and cloud coverage datasets
with four of the classifiers, but the results are poorer
for the wind dataset. In this case IBk is the best clas-
sifier except for the Cloud Coverage dataset where is
slightly improved by RandomForest, PART and J48.
With the ICAST datasets (Table 6) the results are
good for the temperature and wind datasets and worse
for the cloud coverage and precipitation mainly be-
cause the high amount of different classes. Random-
Forest is, once again, the best classifier only improved
by IBk in one dataset.
4.3 Statistical Comparison
In order to provide quantitative evidences for support-
ing the results presented in tables 4, 5, and 6, we
used the STAC platform (STAC, 2014) to perform the
tests of statistical significance on the previously pre-
sented experimentation results, with the aim of de-
termining if statistical differences existed among the
performances achieved by the five classifiers. Since
the samples do not follow a normal distribution, a
nonparametric test had to be used, more specifically
the Iman-Davenport test (Iman and Davenport, 1980)
with a significance level of 0.05. The test results are
presented in table 7, showing RandomForest as the
best classifier followed by IBk, as it was expected
analysing the previous results.
Additionally, we needed to verify that Random-
Forest statistically outperforms all the others classi-
fiers. To confirm it, we applied a Finner test (Finner,
1993) with an alpha value of 0.05 to the classifica-
ICAART2015-InternationalConferenceonAgentsandArtificialIntelligence
560

Table 4: Classification results for the WF datasets.
Weather-Forecast
Precipitation Temperature Wind
F-Score RMSE F-Score RMSE F-Score RMSE
J48 0.965127 0.0396 0.999474 0.0121 0.944170 0.0344
RandomForest 0.968698 0.0361 0.999540 0.0112 0.962207 0.0265
IBk 0.963753 0.0396 0.997274 0.0279 0.959188 0.0302
BayesNet 0.949996 0.0477 0.974612 0.0899 0.809128 0.0639
PART 0.965273 0.0382 0.999248 0.0144 0.948923 0.0332
Table 5: Classification results for the NWS datasets.
National Weather
Service
Precipitation Cloud Coverage Wind
F-Score RMSE F-Score RMSE F-Score RMSE
J48 0.974298 0.019 0.993365 0.0216 0.785243 0.0221
RandomForest 0.978431 0.0166 0.994685 0.0196 0.820437 0.018
IBk 0.979700 0.0169 0.982945 0.0362 0.832648 0.0209
BayesNet 0.715449 0.0665 0.869578 0.1125 0.490977 0.0283
PART 0.973348 0.019 0.993160 0.0223 0.782653 0.0225
Table 6: Classification results for the ICAST datasets.
Intellicast
C. Cov. / Precipitation Temperature Wind
F-Score RMSE F-Score RMSE F-Score RMSE
J48 0.883397 0.0178 0.987056 0.0505 0.965454 0.0507
RandomForest 0.906389 0.0146 0.989658 0.0403 0.977720 0.0391
IBk 0.907474 0.0164 0.988727 0.0458 0.963848 0.0537
BayesNet 0.551672 0.0348 0.780627 0.2556 0.758892 0.1308
PART 0.885185 0.0179 0.987364 0.0495 0.965783 0.051
Table 7: Iman-Davenport Test results.
Ranking Algorithms
1.444 RandomForest
2.667 IBk
2.778 PART
3.111 J48
5.000 BayesNet
p-value < 0.001
Table 8: Finner Test results.
Control Method Control Method VS Adjusted p-value Result
RandomForest BayesNet 0.000 H0 is rejected
- J48 0.050 H0 is rejected
- PART 0.097 H0 is accepted
- IBk 0.101 H0 is accepted
tion results. The null hypothesis (H0): “There is no
difference between classifier A and classifier B” was
accepted against IBk and PART and, in consequence,
we cannot conclude that RandomForest is statistically
better than IBk and PART as shown in table 8. For the
other two classifiers the null hypothesis (H0) was re-
jected meaning that RandomForest is statistically bet-
ter than both of them.
5 CONCLUSIONS
In this document we have presented a summary of
the performance of five classifiers over nine different
datasets from websites that provide textual meteoro-
logical forecasts. We tried to find which one of them
obtains better classification results using the raw me-
teorological data to generate these textual forecasts.
DataMiningforAutomaticLinguisticDescriptionofData-TextualWeatherPredictionasaClassificationProblem
561

In general terms all of the classifiers, with the ex-
ception of BayesNet, achieve good results. In terms
of its F-Score results, RandomForest outperforms the
rest, achieving the best F-Score in most of the cases
(6/9), followed by IBk (3/9). According to this criteria
we can sort these five classifiers by their performance:
RandomForest, IBk, PART, J48 and BayesNet.
In order to verify and provide quantitative evi-
dences for supporting these results, tests of statisti-
cal significance were performed to determine if sta-
tistical differences existed among the performances
achieved by the five classifiers. The test results con-
firmed that RandomForest, IBk and PART are sta-
tistically the best classifiers, although RandomForest
achieved slightly better results. On the other hand,
J48 and BayesNet present significant performance
differences, and therefore they do not show to be so
valid from an experimental viewpoint.
In this work we have conceived a feasible model
for providing rapid and accurate linguistic predictions
in an intermediate language composed of linguistic la-
bels. As future work we aim to extend this analysis to
test with more powerful classification methods such
as SVM or Artificial Neural Networks and develop a
NLG-oriented approach which generate textual fore-
casts from the intermediate language obtained by the
classifiers.
ACKNOWLEDGEMENTS
This work was supported by the Spanish Ministry of
Economy and Competitiveness under grant TIN2011-
29827-C02-02. I. Rodriguez-Fdez is supported by the
Spanish Ministry of Education, under the FPU Fel-
lowships Plan. A. Ramos-Soto is supported by the
Spanish Ministry for Economy and Competitiveness
(FPI Fellowship Program). This work was also sup-
ported in part by the European Regional Development
Fund (ERDF/FEDER) under grants CN2012/151 and
GRC2014/030 of the Galician Ministry of Education.
REFERENCES
Adeyanju, I. (2012). Generating weather forecast texts with
case based reasoning. International Journal of Com-
puter Applications, 45.
AEMET (2014). Spanish meteorological agency website.
http://www.aemet.es/, Retrieved: 2014-10-08.
Aha, D. W., Kibler, D., and Albert, M. K. (1991).
Instance-based learning algorithms. Machine learn-
ing, 6(1):37–66.
Bouckaert, R. R. (2004). Bayesian network classifiers in
weka. Department of Computer Science, University
of Waikato.
Breiman, L. (2001). Random forests. Machine learning,
45(1):5–32.
Coch, J. (1998). Multimeteo: multilingual production of
weather forecasts. ELRA Newsletter, 3(2).
Finner, H. (1993). On a monotonicity problem in step-down
multiple test procedures. Journal of the American Sta-
tistical Association, 88(423):920–923.
Frank, E. and Witten, I. H. (1998). Generating accurate
rule sets without global optimization. In Shavlik, J.,
editor, Fifteenth International Conference on Machine
Learning, pages 144–151. Morgan Kaufmann.
Goldberg, E., Driedger, N., and Kittredge, R. I. (1994). Us-
ing natural-language processing to produce weather
forecasts. IEEE Expert, 9(2):45–53.
Hall, M., Frank, E., Holmes, G., Pfahringer, B., Reutemann,
P., and Witten, I. H. (2009). The weka data min-
ing software: an update. ACM SIGKDD explorations
newsletter, 11(1):10–18.
Iman, R. L. and Davenport, J. M. (1980). Approximations
of the critical region of the fbietkan statistic. Commu-
nications in Statistics-Theory and Methods, 9(6):571–
595.
Intellicast (2014). Intellicast website. http://
www.intellicast.com/, Retrieved: 2014-10-08.
MetOffice (2014a). British meteorological office data2text
website. http://www.metoffice.gov.uk/public/weather/
forecast-data2text, Retrieved: 2014-10-08.
MetOffice (2014b). British meteorological office website.
http://www.metoffice.gov.uk/, Retrieved: 2014-10-08.
NWF (2014). National weather forecast website. http://
www.weather.gov/, Retrieved: 2014-10-08.
Quinlan, J. R. (1993). C4. 5: programs for machine learn-
ing, volume 1. Morgan kaufmann.
Ramos Soto, A., Bugarin, A., Barro, S., and Taboada, J.
(2014). Linguistic descriptions for automatic gener-
ation of textual short-term weather forecasts on real
prediction data. IEEE Transactions on Fuzzy Systems,
Early Access.
Reiter, E., Dale, R., and Feng, Z. (2000). Building natural
language generation systems, volume 33. MIT Press.
Sokolova, M., Japkowicz, N., and Szpakowicz, S. (2006).
Beyond accuracy, f-score and roc: a family of dis-
criminant measures for performance evaluation. In AI
2006: Advances in Artificial Intelligence, pages 1015–
1021. Springer.
STAC (2014). Stac: Web platform for algorithms compar-
ison through statistical tests. http://tec.citius.usc.es/
stac/, Retrieved: 2014-10-08.
Van Deemter, K., Krahmer, E., and Theune, M. (2005).
Real versus template-based natural language genera-
tion: A false opposition? Computational Linguistics,
31(1):15–24.
WeatherForecast (2014). Weather-forecast website. http://
www.weather-forecast.com/, Retrieved: 2014-10-08.
ICAART2015-InternationalConferenceonAgentsandArtificialIntelligence
562