SYSTEM DEMONSTRATION 
GOALGETTER: GENERATION OF SPOKEN SOCCER REPORTS 
Mari~t Theune and Esther Klabbers 
IPO, Center for Research on User-System Interaction, 
Eindhoven University of Technology. 
Abstract - In this paper we describe a demonstration Of the GoalGetter system, which generates spoken soc- 
cer reports (in Dutch) on the basis of tabular data. Two types of speech output are available. The demo runs 
via the web. It includes the possibility of !creating your own match' and having GoalGetter generate a report 
on this match. • . ._ 
1. About the system 
• The GoalGetter system is a Data-to-Speech system which generates spoken soccer reports (in Dutch) on the 
basis of tabular data. The system takes as input data about a soccer match that are derived from a Teletext 
• page. 1 The output of the system is a spoken, natural language report conveying the main events of the match 
described on the Teletext page, 
GoalGetter was developed on the basis of D2S, a generic system for the creation of Data-to-Speech systems. 
The general architecture of D2S is represented in Figure 1. It consists of two modules, the Language Gen- 
eration Module (LGM), and the Speech Generation Module (SGM). The LGM takes data as input .and 
• produces enriched text, i.e., text which is annotated with prosodic markers indicating pitch accents and into- 
national boundarie s. This text is sentas input to the SGM, which turns it into a speech signal. 
Data  n °ag° I SPeech I _- Generation Enriched Module \] " Text ~ Generati°n IModule Speech Signal 
Figure 1" Global architecture of D2S 
Language generation in the LGM is done using syntactic templates, which are syntactic tree structures con- 
taining slots for variable expressions. The selection and ordering of the templates and the filling of their 
slots depend mainly on conditions on the discourse context, which is represented in a Discourse Model. In 
order to achieve variation in the generated texts, the system generates a number of different expressions for 
each piece of information. If various expressions are equally suitable given the current contexL one is cho- 
sen at random. For a detailed description of the syntactic template technique we refer to \[Van Deemter & 
Odijk 97\] and \[Theune et al. 97\]. 
In the Prosody module, accentuation and phrasing are determined using information about the syntactic 
structure of a sentence and about its context. Phrases expressing 'new' or 'contrastive' information are 
accented while those expressing 'given' information are deaccented. Speech generation in the SGM is done 
1. Teletext is a system with which information is broadcast along with the television signal and decoded in 
the receiver. The information is distributed over various 'pages', most of which contain textual informa- 
tion, but some contain tables. 
292 
! 
I 
I 
t 
I 
i| 
by means of either speech synthesis or an advanced form of phrase concatenation, using different prosodic 
versions of otherwise identical phrases•. Which version is chosen depends on the prosodic markings pro- 
vided by the LGM. For Dutch speech synthesis, we use the phonetics-to-speech system SPENGI (SPeech 
synthesis ENGIne) developed at IPO, which employs PSOLA-based diphones. 
2. Outline of the demonstration 
The demonstration of the GoalGetter system runs via the web and consists of three parts. First we present 
an example report that has been generated off-line by GoalGetter. This allows us to prepare an (annotated) 
English translation of the Dutch text in advance. (In the other parts of the demonstration, preparing a trans- 
lation is not possible since reports will be generated on-line and i t is impossible to predict the exact contents 
of the reports, due to the variation mentioned above. In those cases we providea rough translation on the 
spot.) Second, we generate a report on-line, using an existing input table. FinallY, we fill an input table with 
data selected by those attending the demonstration, and generate a report expressing these data. All modules 
depicted in Figure 1 are included in the demonstration. For each 'generation round', first a plain text version 
of the generated report is shown, and then the enriched text with the prosodic markers. Finally, the system's 
• speech output is played. 2 
Teletext page 
Ajax (D 
Goals 
i • " Litmanen (23) 
Scheidsrechter Temmink 
Tceschouwers 25000 
I Kaarten 
i Grim (geeO 
Kaa'ten = Cards 
Vitesse (2) Toeschouwers = Spectators 
Scheidsrechte~ = Referee 
. . geel = yellow 
Crofter (12/per~ rood = red 
Atteveld (45) ed = own goal 
pen = pen~dty 
Output from language generation 
Vite~e ging op be2oek bij Ajax en won met een - twee 
vijfentwintig duizend toe~houwevz kwamen naxr De Meer 
de Vitesse speler Gorter benutte in de twaalfde mintrat een penalty 
in de drieentwintigste mimmt braeht de Ajax zpeler Litmanen de teams op gelijke hoogte 
de verdediger Atteveld maa.kte na vijfenveertig minuten her winnende doelptmt voor Vite~e en bepaalde dawmee 
de einck~t and op een- twee 
Grim van Ajax ontving een gele kaart van ~heldsreehter Temmln~ 
er vielen geenrode kaarten 
Figure 2: Screenshot showing Teletext page and generated report 
Example report - In order to give a first impression of the GoalGetter system, we start the demonstration by 
2. Due to the fact that a few phrases are missing from the phrase database, phrase concatenation is only avail- 
able in the first part of the demonstration. The output from diphone synthesis is available in all parts. 
293 
showing an example 'Teletext page "3 and a report that was generated onthe basis of this page in both plain 
and enriched text format. A written English translation will be provided. We play a spoken version of the 
report using both speech synthesis and phrase concatenation. 
The screenshot in Figure 2 shows a Teletext page and the plain • text version of a generated report. 4 Since 
the screenshot was taken from the English language version of the web demonstration, to the right of the 
Teletext page we can see a translation table of the words and abbreviations occurring on the page. Below 
this table, there is a 'Create report' button, which can be used to generate a new report on the basis of the 
same Teletext page. 
Below the report, there are three more buttons. Clicking the first replaces the plain text version of the report 
with the enriched text version, shown in Figure 3. 5 The other two buttons can be clicked to play the audio 
files that were created when the report was generated, using phrase concatenation and diphone synthesis 
• respectively. The sound files can be either in aiff or in wav format. 
Output from language generation 
I/"Vite~ze ging op "bezoek bij "Ajax/I en "won met "@een H ~ "@twee/H 
"vijfentmintig dulzend "'toe~aouwe~ / kwaman na~ De "Meet I/I 
:le % "Vite~e speler "Gorter/ benutte in de "t-4taalfde "mint.nat een "penalty/// 
in de "drieentwintigste mintrat/brncht de % "Ajax speler "Litmw~n /de teams op "gelijke "hoogte//I 
:le % "verdediger "Arteveld / maa.~te na "vijfenveertig minuten bet "winrm~e doelpuaat / voor "Vite~e Hen 
~*paalde ,:la~'mee de ?elndstimd/op "~n II- "@trace l/I 
"Grim van "Ajax / ontving een "gele "l~art \[ van "~heidsrechter "Temmink//1' 
vielen "geen "rode ~'ten I/! 
Figure 3: Partial screenshot showing generated report in enriched text format} 
Generate a report on-line - The next step is to generate a report on-line. We use the same Teletext page as 
before, which gives us the opportunity tO point out the variation in the generated reports. Although the writ- 
ten report (including prosodic markers) is generated almost instantaneously, the fact that the sound files 
have to be created at the same time causes a small delay of about five seconds. When the report has been 
generated and one of the speech buttons is Clicked, the corresponding sound file has already been created 
and onlY has to be read. 
Creating a match - Finally, the web demonstration allows• one to define the input for GoalGetter by filling 
an empty input table according to one's own preferences. We use this feature to 'create a match' using data 
selected by those present at the demonstration. On the basis of these data, a 'Teletext pag e' is created and 
reports can be generated in the manner explained above. 
In Figure 4 we can see how an input table is filled. First, general information about the match is specified, 
such as the teams involved, the referee and the number of spectators. The teams and the referee are chosen 
from a predefined list, via a pull-down menu. After the general information has been determined, the main 
• 3. :In the web demonstration, we use simulated Teletext pages which are based on the original ones. 
4. Translation: Yitesse visited Ajax and gained a 1-2 victory. Twenty five thousand spectators came to De 
Meer. The Vitesse player Gorter kicked a penalty home in the 12th minute. In the 23rd minute, the Ajax 
player Litmanen leveled the score. After 45 minutes, the defender Atteveld scored the winning goal and 
: thereby brought thefinal score to 1-2. Grim from Ajax received a yellow card from refere e Temmink. There 
were no red cards 
5~ In the enriched text, accents are indicated by double quotes, and intonation boundaries of different 
strengths by a number of slashes. The symbols• @ and % indicate specific pronunciation s for certain words. 
294 
events of the match can be specified in the 'Goals and cards' section. This is done by selecting a player from 
one of the teams, selecting the event this player was involved in (goal, own goal, penalty, yellow or red 
card), specifying the minute the event occurred in (for cards this is optional), and then clicking on the Add 
button to add the 'event' to the input table. This action can be repeated until an interesting match has been 
created. Then clicking the 'Create the Teletext page' button will produce a page similar to the one shown 
in Figure 2, and a report can be generated. 
General information 
Home Team Visiting Team Referee Spectators 
..... 1 
Goals and cards 
de Gl-aaf~:hap Go Ahead Eagles 
i A-al 
Overview 
If i 
i! ! 
i i 
i --f -I 
Card/Goal 
Own go~d 
-- . I C :-:r-== ..................... =--":- ..... 
- ! i Create the.!ele_text pageJ 
.............. i" 
i i Clear the table~ 
............. i 
1 I 
~~~'.-==-]-~="'~:C ~" ;-~i-=r:';'~ '- :-;,~; :'=-= ~k~ '; :'~= ~: ....... ,IE 
Figure 4: Screenshot showing the filling of an input table 
3. Requirements 
The GoalGetter demonstration can be run via the web. The web browser that is used should support frames 
and tables. (In the case of Netscape, this means that version 2.0 or higher is required.) In order to make the 
sound files audible, a sound card must be available. When running the demonstration on a PC, wav format 
should be chosen for the sound files, while for a Unix machine aiff-format is more suitable. The options of 
the web browser should be set so that audio files are played (e.g., using sfplay) instead of being saved to 
•disk. 

References 
[Van Deemter & Odijk 97] Van Deemter, K. and J. Odijk, 1997. Context Modelling and the Generation of 
Spoken Discourse. Speech Communication 21 ( 1/2), 101-121. 
[Theune et al. 97] Theune, M., E. Klabbers, J. Odijk and J.R. de Pijper, 1997. From Data to Speech: A 
Generic Approach. Manuscript, URL: http://www.tue.nl/ipolpeople/theune/manuscript.ps.Z. 
