System Demonstration 
Multilingual Weather Forecast Generation System 
Tianfang Yao Dongmo Zhang Qian Wang 
Department of Computer Science and Engineering, Shanghai Jiao Tong University 
1954 Hua Shan Road, Shanghai 200030, Shanghai, • •China 
E-mail: yao-tf{zhang-dm, wangqian}@cs.sjtu.edu.cn 
Abstract 
The MLWFA (Multilingual Weather Forecasts Assistant) system will be demonstrated. It is 
developed to generate the multilingual text of the weather forecasts automatically. The raw data 
from the weather observation can be used to generate the weather element chart. According to the 
weather change trend, the forecasters can directly modify the value Of the element on the chart, 
such as the center .point value, the isoline and the isocircle. After •that, the modified data are 
stored as the input for the system. The system can select a schema depending on the input or the 
requirement from the users. The schema library can be conveniently maintained, such as the 
schema modification or extension. Through optimizing and mapping the schema •tree, the 
microplanner constructs the brief and coherent internal text structure for the surface generator. 
After the processing of the generator, the muitilingual •weather forecasts used for the broadcast 
program are generated. 
Keywords: Multilingual Generation, Weather Forecast Assistant. 
., 1 Introduction 
The MLWFA ~system is developed as the first application of the project ACNLG \[Huang et 
al. 97a & b\] which is an international •cooperation between the German Research Center for 
Artificial Intelligence (DFKI) and Shanghai••Jiao Tong University (SJTU). The system mainly 
consists of four components: the graphic processor, the macroplanner, the microplanner and the 
surface generator. The graphic, •processor is used to adjust weather forecasts data by the 
forecasters. The technique •adopted for the macroplanner is based on the schema approach 
\[McKeown 85\], but we expand the operator of schema. The microplanner is based on the 
sentence structure optimizing which is independent of the language and language resource mapping 
-which is associated with the language. On the basis of the FB-LTAG (Feature-based Lexicalized 
-Tree Adjoining Grammar) \[Joshi 85, XTAGRG 95\], the surface generator identifies the feature of 
the nodes, compounds the grammar-trees and finally generates Chinese, •English and German 
weather forecasts. 
2 Architecture 
The architecture of the system is shown in figure 1. The macroplanner obtains original 
predicted weather data either from the interface or from the users. These data mainly include 
weather status, wind direction, wind force, temperature and so on. 
In the stag e of macroplanning, we adopt the schema approach and introduce the "It" operator 
• which indicates that the order of the predicates can be exchanged. The macroplanning procedure 
296 
II 
I 
I 
iJ 
I 
i 
'l 
sI 
i| 
i 
! 
i 
! 
! 
I 
I 
i 
I 
! 
! 
I 
i 
I 
I 
I 
I 
I 
contains five major steps: inputting meteorological data, building knowledgebase dynamically, 
initializing predicate, selecting schema and filling schema \[Wang et al. 97\]. 
The users can dynamically define objects and input their property value. Then the system will 
dynamically build a hierarchical knowledgebase. With the help of that, the system transforms 
original data to the conceptual value. After producing the conceptual value, the system initializes 
the predicates with the conceptual value. The process of initialization is to fill the conceptual 
value in thepredicate by matching the semantics of the conceptual value and tl~e argument. We 
add the layout information on the predicate and define some valid argument combinations that 
enhance the expressible ability of the predicate. In the macroplanner, System selects schema 
according to three conditions. There are selecting information carried by the schema item, big 
grain information that is deduced by several conceptual values according to the domain conditions 
and the selecting priority. 
input and requirement s from users 
soeol 1 Macroplarmer 
Language I \[ Resource ~-~ Microplanner 
Knowledge I I " " ' 
Mult ilingdat Text Internal Structure 
L I Language " 
IOptimizing IKnowledge 
l Libra D 
I Grammat , ~ 1 Flule ~ Surface Gener~or 
\[Library\[\[ Mult,i~gual 
Text 
Machine 
3ict io nary 
Fig. 1 MLWFA System Architecture 
The work of the microplanner can be divided into two closely related parts: the sentence 
structure optimizing (SSO) and the language resources mapping (LRM). SSO aims at removing 
redundant information in the schema tree so as to make the sentence natural, fluent, and readable, 
and a semantic-based rule driven method is adopted in this component. LRM adds the function 
words into it in conformity to the specific language syntax, and enriches them with the specific 
language resources so that every expression in the output is flexible and grammatically correct. 
In accordance with the weather forecast domain, 5 types of optimizing tasks are classified: 
the3' are the aggregation of the proposition with the common parts, the constitute conjoining, the 
297 
domain restriction, theconstitute omission and the reference planning. Each stands for a group of 
the optimizing operations. In this idea SSO can be confined within the semantic level which is 
irrelevant with any specific language, so a single operation can be done for all languages. 
As to LRM, the sentence structure classes are used to represent the detailed Sentence 
structures as well as the syntax and semantic information for every composition of the sentence. 
There are several sentence structure classes defined for every predicate. Considering text 
requirements, one of the classes above is chosen and determined to construct the corresponding 
text structure tree. Furthermore, according to the characteristics of the different language, the 
dynamic information is added to the relevant nodes of the tree. As the definition differs widely for 
different target languages, we implement this module for different languages respectively. 
The input of surface generator is the output of the microplanner, which is a dependent tree 
structure including not only the information of the text content, the sentence syntax and the 
sentence semantics, but also the result of the optimizing. 
• We have used the FB-LTAG as a generation grammar for our generator [Yao et al. 97, Wang 
97]. The algorithm of the surface generator can be summarized as: 
• Instantiate the trees and compute all legal subtree compositions; 
• Pick out the first •tree from the set of acceptable derived trees, which satisfies the feature 
constraints and has a complete structure; 
• • Traverse the selected tree and output the Sentences (do some morphology according to 
the •feature values, if necessary). 
.3 Demonstration 
The main window of the system during the macroplanning is shown in figure 2. 
L 
:~oI xl 
;i 
~ ~rldi¢t%¢ r*O.¢u4 $~tict 0113133 
~ &ttri~ut* 
• L Itedi~ re¢~ s~*ti*= 
i ' .ArCaa~t ?i** Yd.~¢ Otri~ 
p~'~~~i(~'_ ~ '~ , .. ~ .... ~.., ,,. 
i 7] #riai¢itl l~l~ub 
t ~l*lld~llj*¢t I~lstirdi71 
-~ -_~ "-" r/ 7f ~ t.:, ~ ~.,~j ....... ~ ..... 
v. ='"" ...... - 
• ~J- D~aaicObj let vii~te 
Fig. 2 The Main Window 
At the top of the screen is the main menu of the system. In the system submenu the user can 
maintain all the database in the system. In the data submenu the system can generate Chinese, 
English or German text, through executing the macroplanning, the microplanning and the surface 
generation procedure in turn. In the prediction submenu the user can change options of the 
system, either produce forecasts step by step or do it uninterruptedly. The left window on the 
middle of the screen shows the weather element chart, the user can modify the element value on 
the chart of that window directly. The right window shows the result of the macropianning, we 
split that window into two parts: One shows the tree structure of the macroplanning result; the 
298 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
•I 
other shows the detail information of the focusing node. 
4 Conclusion 
We have implemented a prototype of MLWFA system consisting of four components in 
VC++ under Windows 95. Currently we are replenishing more optimizing rules and sentence 
structure classes, and further modifying some existing ones. After doing that, we will plan to put 
this prototype into test use. 
Acknowledgements We want to thank sincerely Dr, Xiaorong Huang for his concrete 
instruction and valuable suggestion. Thanks also to the people• who took part in the project or 
made valuable suggestions. 
This work is supported by the VW-Stiftung, Germany, the National Science Foundation, 
China, and the Science and Technology Commission of Shanghai, China. 

References 
\[Huang et al. 97a\] Xiaorong Huang, Tianfang Yao and Guodong Gao. 1997. Generating Chinese Weather 
Forecast with Stylistic Variations. In Proc. 17th International Conference on Computer Processing of 
Oriental Language, Hongkong, 733-738 
\[Huang et al. 97b\] Xiaorong Huang, Tianfang Yao, Huanye Sheng. 1997. The Project ACNLG: Applying 
Natural Language Generation in China. In Proc. DFKI Workshop on Natural Language Generation, 
Saarbruecken, Germany. 
\[Joshi 85\] Aravind K. Joshi. 1985. An Introduction to Tree Adjoining Grammars. Technical Report MS-CIS- 
86- 64, LINCLAB-31, Department of Computer and Information Science, Moore School, University of 
Pennsylvania. 
\[Mckeown 85\] Kathleen R. McKeown. 1985. Text Generation. Cambridge Press, Cambridge, UK. 
\[Wang et al. 97\] Qian Wang and Tianfang Yao. 1997. Design and Implementation of Macroplanner for Chinese 
Weather Forecast Automated Generation System, In Proc. of the Joint Symposium Computational 
• Linguistics '97, Beijing, China, 176-181 
\[Wang 97\] Yufang Wang. 1997. A Tree Adjoining Grammar for Chinese Weather Forecasts. In Proc. of the 
Joint Symposium Computational Linguistics '97, Beijing, China, 207o213 
\[XTARG 95\] The XTAG Research Group. 1995. A Lexicalized Tree Adjoining Grammar for English. Technical 
Report IRCS 95-03, Institute for Research in Cognitive Science, University of Pennsylvania. 
\[Yao et al. 97\] Tianfang Yao, Xiaorong Huang and Huanye Sheng. 1997. The MLWFA System. In Proc. of the 
Natural Language Processing Pacific Rim Symposium 1997. Phuket, Thailand, 609-612.
