A Multilingual News Summarizer 
Ilsin-Hsi Chen 
l)epartlnent of Computer Science and 
Information Engineering 
National Taiwan University 
Taipei, TAIWAN, R.O.C. 
hh chen @csie.ntu.edu.tw 
Chuan-Jie Lin 
Defmltment of Computer Science and 
Information Engineering 
National Taiwan University 
Taipei, TAIWAN, R.O.C. 
cjli n @ nlg2.csie.ntu.edu.tw 
Abstract 
Huge multilingual news articles are reported 
and disseminated on the Internet. ltow to 
extract the kcy information and savc the 
reading time is a crucial issue. This paper 
proposes architecture of multilingual news 
sumlnarizer, including monolingual and 
multilingual clustering, similarity measure 
among lneaningful ullits, and presentation of 
summarization results. Translation anlong 
news stories, idiosyncrasy among s, 
itnplicit information, and user preference are 
addressed. 
Introduction 
Today many web sites on the lnternet provide 
online news services. Multilingual news articles 
are reported periodically, and across geographic 
barrier to disseminate to readers. Readers can 
access the news stories conveniently, but it takes 
much time l'or people to read all tile news. This 
paper will present a personal news secretariat that 
helps on-line readers absorb news information 
from multiple sources in different s. 
Such a news secretariat eliminates the redundant 
information in tile news articles, reorganizes tile 
news for readers, and helps them resolve the 
 barriers. 
Reorganization of news is sonic sort of 
document summarization, which creates a short 
version of original document. Recently, many 
papers touch on single document summarization 
(ltovy and Marcu, 1998a). Only a few touch on 
multiple document sulnmarization (Chen and 
Huang, 1999; Mani and Bloedorn, 1997; Radev 
and McKeown, 1998) and multilingual document 
summarization (Hovy and Marcu, 1998b). For 
multilingual multiple news summarization, several 
issues have to be addressed: 
(1) Translation among news stories in 
different s 
The basic idea in multiple doculnent 
sulnmarizations is to identify which paris of news 
articles present similar reports. Because the 
news stories are in different s, seine kind 
of Iranslation is required, e.g., term translation. 
Besides the problem of translation ambiguity, 
different news sites often use difl'erent names to 
refer tile same entity. The translation o1' named 
entities, which are usually ttnknown words, is 
another probleln. 
(2) Idiosyncrasy among s 
1)ifferent s have their own specific 
features. For example, a Chinese sentence is 
composed of characters without word boundary. 
Word segmentation is indispensable for Chinese. 
Besides, Chinese writers often assign l~unctuation 
ntarks at randonl, how to determine a mealfingful 
unit for similarity checking is a crucial issue. 
Thus seine tasks may be done for specific 
s during SUlnmarization. 
(3) hnplicit information in news reports 
Some information is ilnplicit in news stories. 
For example, the name of a country is usually not 
mentioned in a news article reporting an event that 
happened in that country. On the contrary, the 
country name is important in foreign news. 
Besides, time zone is used to specify date/time 
implicitly in the news. 
(4) User preference 
When users want to read documents in their 
familiar s, news fragments in some 
159 
9 ANt r::: 
Nl.lnlnlary N illllmal'y Sullllllary ,~ 11111111 al'y 
ltlr l'venl 1 for l ivent 2 for l ivem 3 for Event m 
Figure 1. Architecture of 
Our Multilingual Sunmmrization System 
s are preferred to those in other s. 
Even machine translation should be introduced to 
translate news fragments. Besides, if a user 
prefers the news from tile view of his country, or 
more precisely, of some news sites, we should 
meet his need. 
Figure 1 shows the architecture of a 
multilingual summarization system, which is used 
to sulnmarize the news from multiple sources in 
different s. It is composed of three 
m~tior components: several monolingual news 
clusterers, a multilingual news clusterer, and a 
news summarizer. Tile monolingual news 
clusterer receives a news stream from multiple on.- 
line newspapers in its respective , and 
directs them into several output news streams by 
using events. The multilingual news clusterer 
then matches and merges the news streams of the 
same event but in different s in a cluster. 
The news summarizer summarizes the news 
stories for each event. 
The possible tasks for each component 
depend on the s used. Some major tasks 
of a monolingual clusterer are listed below. 
(1) identifying word boundaries for Chinese 
and Japanese sentences, 
(2) Extracting named entities like people, 
place, organization, time, date and monetary 
expressions, 
(3) Clustering news streams based on 
predefined topic set and named entities. 
The task for the multilingual clusterer is to 
align the news clusters in the same topic set, but in 
different s. It is similar to document 
alignment in comparable corpus. Named entities 
are also useful cues. 
The major tasks for the news summarizer are 
shown as follows. 
(1) Partitioning a news story into several 
meaningful units (MUs), 
(2) Linking the lneaningful units, denoting 
the salne thing, from different news reports, 
(3) Displaying the summarization results 
under the consideration of  type users 
prefer, information decay and views of reporters. 
1. Clustering 
1.1 Monolingual Clustering 
We adopt a two-level approach to cluster the 
news t)o111 multiple sources. At first, news is 
classified on the basis of a predefined topic set. 
Then, tile news articles in the same topic set are 
partitioned into several clusters according to 
named emities. Classification is necessary. Oil 
tile one hand, a famous person may appear in 
many kinds of news stories. For example, 
President Clinton may make a public speech 
(political news), join an international meeting 
(international news), or even just show up in the 
opening of a baseball game (sports news). On 
the other hand, a common name is flequently seen 
but denotes different persons. Classification 
reduces the ambiguity introduced by famous 
persons and/or common names. 
An event in a news story is characterized by 
five basic entities such as people, affairs, time, 
places and things. These entities form important 
cues during clustering. Systems for named entity 
extraction in a famous lnessage understanding 
competition (MUC, 1998) demonstrate promising 
performances for English, Japanese and Chinese. 
In our multilingual summarization system, we 
focus on English and Chinese. Gazetteer 
approach is adopted to deal with English news 
articles. Comparatively, Chinese news articles 
are segmented at first. Then, several types of 
inforlnation fiom character, sentence and text 
levels are employed to extract Chinese named 
160 
entities. These tasks are similar to tile 
approaches ill tile papers (Chen and Lee, 1996; 
Chen, el al., 1998a). 
1.2 Multilingual Clustering 
Tile multilingual clusterer takes input from 
the lnonolingual clusterers, and determines which 
news clusters ill which s talk about tile 
same story. Recall that a news cluster consists of 
several news articles reporting tile same event, and 
one news cluster exists lbr one event ariel 
monolingual clustering. Ill this way, there is at 
most one corresponding news cluster ill another 
. Therefore, the main task of the 
multilingual news clusterer is to lind tile 
matchings among tile clusters ill different 
s. Figure 2 shows an example, ill 
Topic !, cluster cHl is aligned to c itr, and cluster 
Cil 2 is aligned to c.ili. Clusters cii~z arid cjl 2 are 
left unaligned. That means the denoted events 
arc reported ill only one . 
Similarity of two clusters is measured based 
on verbs, named entities, and the other nouns. 
Because Chinese words are less anibiguolls tMn 
English ones (Chen, Bian anti Lin, 1999), we 
translate nouns and verbs in the Chinese news 
articles into English. If a word Ms more than 
one translation, we select high fl-equent English 
translation. For tile named enlities not listed ill 
tile lexicon, name transliteration similar to tile 
algoritlnn (Chen, el al., 1998b) is introduced for 
matching in non-alpMbetic (e.g., Clfinese) and 
alphabetic s (e.g., English). 
Alignment is made under the same topic. A 
news chlster c i is aligried to another cluster cj if 
their similarity is above a threshold, and is tile 
highest between q and the other clusters. If tile 
similarity of q and the other clusters is less than a 
given threshold, c i is not aligned. It is possible 
because local news is reported only ill tile 
restricted areas. 
2. Similarity Analysis 
2.1 Meauingful Units 
The basic idea during smnmarization is to tell 
which parts of the news articles are similar in the 
same event. The basic unit tbr similarity 
measure may be a paragraph or a sentence. For 
Language l., 
Language 1i 
Topic I Topic i 
I Topic I Topic t 
Figure 2. Matching among tile Clusters 
in Two Languages 
tile t'ormer, text segmentation is necessary for 
documents without paragraph markers (Chcn and 
Chen, 1995). For the latter, text segmentation is 
necessary ibr s like Chinese. Unlike 
English writers, Chinese writers often assign 
punctuation marks at random (Chen, 1994). 
Thus the sentence boundary is not clear. 
Consider the following Chinese example (C l): 
(Central News Agency, 1999.12.02) 
(Although they were undeterred by mass arrests 
and a police crackdown, anti free-trade protesters 
still marched on downtown Seattle today. The 
protesters, carrying signs and chanting, opposed 
lhc global trade liberalization being worked on at 
a meeting of h+ade lninisters flom tile World Trade 
Organi zat ion.) 
It is composed of four sentence segments 
separated by commas. 11' a sentence segment is 
regarded as a unit for similarity checking, it may 
contain too little information. On tile contrary, if 
a sentence is regarded as a unit, it may contain too 
much M'ormation. Here we consider a 
meaningful unit (MU) as a basic unit for 
measurement. A MU is composed of several 
sentence segments and denotes a complete 
meaning. We will find two MUs shown as 
follows for (C 1): 
(Although they were undeterred by mass arrests 
and a police crackdown, anti free-trade protesters 
still marched on downtown Seattle today.) 
161 
-~.~ .e- ~ 4-/5,-- " 5, 
(The protesters, carrying signs and chanting, 
opposed the global trade liberalization being 
worked on at a meeting of trade ministers fl'om the 
World Trade Organization.) 
In our summarization system, an English 
sentence itself is an MU. Comparatively, it is a 
little harder to identify Chinese MUs. Three 
kinds of linguistic kuowledge- punctuation marks, 
linking elements and topic chaius, are proposed. 
(1) Punctuation marks 
There are fourteen marks in Chinese (Yang, 1981). 
Only period, question mark, exclamation mark, 
comma, semicolon and caesura mark are 
employed. The former three are sentence 
terminators, and the latter three are segment 
separators. 
(2) Linking elements 
There are three kinds of linking elements (Li and 
Thompson, 1981): forward-linking elements, 
backward-linking elements, and couple-linking 
elements. A segment with a forward-linking 
(backward-linking) elemeut is linked with its next 
(previous) segment. A couple-linking element is 
a pair of words that exist in two segments. 
Apparently, these two segments are joined 
together. Examples (C4)-(C6) show each ldnd of 
linkings. 
(C4) T~,~:-~,,..,-~ ' q~&d# g#~ ° 
(After school, I wanted to see a movie.) 
(I wanted to see a movie, but I couldn't get a 
ticket.) 
(C6) N -h&a-~ ~. ~'; g ' ,~a.rx & a_~-.-24 "~d 
:V4 o 
(Because I couldfft get a ticket, (so) 1 didn't 
see a movie.) 
(3) Topic chains 
The topic of a clausal segment is usually deleted 
under the identity with a topic in its preceding 
segment. The result of such a deleting process is 
a topic chain. We employ part of speech 
information to predict if a subject of a verb is 
missing. If it does, we postulate that it must 
appear in the previous segment and the two 
segments are connected to form a larger unit. 
Consider example (C1). The word "f'$ @" 
(although) is a forward linking element. Thus 
the first two segments are connected together (C2). 
The last segment does not have ally subject, so 
that it is connected to the previous one by topic 
chain (C3). In summary, two MUs are formed. 
2.2 Similarity Model 
Tile next step is to find the similarity among 
MUs in the news articles reporting the same event, 
and to link the similar MUs together. We 
analyze the news stories within the same , 
and then the news stories among different 
s. The key idea is similar at these two 
steps. That is, predicate argument structure 
forms the kernel of a sentence, thus verbs and 
nouns are regarded as important cues for similarity 
measures. The difference between these two 
steps is that we have to translate nouns and verbs 
in one  into another . The 
approach of select-high-frequent translation and 
name transliteration shown in Section 1.2 is 
adopted here too. Consider (MUI) - (MU3). 
The former two are in Chinese and the last one is 
in English. They denote a similar event 
"Seattle's Curfew Hours". Each noun (verb) is 
enclosed by parentheses and assigned an index. 
There are 9 common terms between (MUI) and 
(MU2); 10 common terms between (MUI) and 
(MU3); and 8 common terms between (MU2) and 
(MU3). Note the time zones used in (MU2) and 
(MUI) are different, so are (MU2) and (MU3). 
(MU 1 ) .g (1 ~-J 5J~ N )(2 ~ ~ )(3 ~'~ ~ )(4 ~ )~I ~- ) ~'~ (5 
2-~-(11~)~%) ( 12"qc2 }\]~)(13~\]~)(14>J" ;~v)"~ ' kl5 ;~ 
G}I~"(16.T-.~")(,7,{g'a~) 1" (,s'?O" fl" '~) o 
(Chinatimes, 1999.12.02) 
(MU2) (, N~IflN)(2~-~v)(4Gdff), ~(5"1~)(6~}.~/N,~)(7 
(Formosa Television, 1999.12.02) 
(MU3) GSeattle) (2Mayor) (2sPaul) (3Schell) has 
Gdeclared) a (sState) of (scivil) (Temergency) and 
(13imposed) a (m7 p.m.) to (267:30 a.m) ((2v10 p.m.) 
EST - (2s10:30 a.m.) EST) (,4curfew) on 
(2,downtown) (29areas) of the (30city). 
(Reuters) 
162 
(s2) 
(s3) 
($4) 
once. 
(ss) 
Several strategies lnay be considered in 
similarity measure: 
(SI) Nouns in one MU are matched to nouns in 
another MU, so are verbs. 
The operations in (1) are exact matches. 
Thesauri are employed dtu-ing matching. 
Each term specified in (S 1) is matched only 
Tile order of llOUllS and verbs in MU is not 
considered. 
($6) 'File order of nouns and verbs in MU is 
critical, but it is relaxed within a window. 
(S7) When continuous terms are matched, an 
extra score is added. 
($8) When tile object o1: transitive verbs arc not 
matched, a score is subtracled. 
($9) When date/time expressions and monetary 
and percentage expressions are matched, an extra 
score is added. 
l;ive models shown below are collstrtlcted 
under different combinations of tile strategies 
specified in tile above. 
(M t) (S 1)+($3)+($4)+($5) 
(M2) (S 1)+($3)+($4)+($6) 
(M3) (S 1)+(S3)+(S4)+(S5)+($7)+($8) 
(M4) (S 1)+($3)+($4)+($5)+($7)+($8)+($9) 
(M5) (S I)+($2)+($4)+($5)+($7)+($8)+($9) 
3. Experiments 
3.1 l'reparation el'Testing Corpus 
Six events selected from Central l)aily News, 
China I)aily Newspaper, China Times Interactive, 
and FTV News Online in Taiwan arc used to 
lneasure tile performance of each lnodel. They 
are shown as follows: 
(1) military service: 6 articles 
(2) construction permit: 4 articles 
(3) landslide in Shah Jr: 6 articles 
(4) Buslfs sons: 4 articles 
(5) Typhoon Babis: 3 articles 
(6) stabilization fund: 5 articles 
The news events are selected from different 
editions, including social edition, economic 
edition, international edition, political edition, etc. 
An annotator reads all tile news articles, and 
connects tile MUs that discuss the same story. 
Because each MU is assigned a unique ID, the 
links among MUs form the answer keys for the 
performance evaluation. 
Table I. Perfiwmance of Similarity of MUs 
Model 
M I 
M2 
M 3 
M4 
M5 
Precision Rate 
0.5000 
0.4871 
0.5080 
0.5164 
0.5243 
Recall Rate 
0.5434 
0.3905 
(/.5888 
0.6198 
0.5579 
3.2 Resulls 
Traditional precision and recall are computed. 
Table 1 lists the perfornmnce of these five models. 
M I is regarded as a baseline model. M2 is 
different l'ronl M1 in that the matching order of 
nouns itl\](l verbs are kept conditionally. It tries to 
consider the subject-verl>object sequence. The 
experiment shows that tile performance is worse. 
The major reason is that we c~ltl express the same 
meaning using different syntactic structures. 
Movement transformation may affect tile order of 
sulkiest-verb-object. Thus in M3 we give up the 
order criterion, but we add an extra score when 
continuous terms are matched, lind subtract some 
score when tile object of a transitive verb is not 
matched. Compared with M1, the precision is a 
little higher, and tile recall is improved about 4.5%. 
If we further consider some special named entities 
such as date/time expressions and monetary and 
percentage expressions in M4, tile recall is 
increased about 7.6% at no expense of precision. 
M5 tries Io estimate tile function of tile thesauri. 
It uses exact matching. Tile precision is a little 
higher but the recall is decreased abollt G% 
compared with M4. 
Several m~\ior errors affect tile overall 
performance. Using nouns and verbs to find the 
similar MUs is not always workable. Tile same 
meaning may not be expressed in terms of the 
same words or synonymous words. Besides, we 
can use different format to express monetary and 
percentage expressions. Word segmentation is 
another source of errors. Two sentences 
denoting tile similar meaning may be segmented 
differently clue to tile segmentation strategies. 
Unknown words generate many single-character 
words. After tagging, these words tend to be 
nOUllS and verbs, which are used in computing tile 
scores for similarity measure. Thus errors may 
be introduced. 
163 
4. Presentation Model 
Two models, i.e., focusing inodel and 
browsing model, are proposed to display the 
sumlnarization results. In the focusing model, a 
SUlnlnarization is presented by voting fi'om 
reporters. For each event, a reporter records a 
news story from his own viewpoint. Recall that 
a news article is composed of several MUs. 
Those MUs that are similar in a specific event are 
COlnmon focuses of different reporters. In other 
words, they are worthy of reading. In the current 
ilnplementation, the MUs that are reported more 
than once are our target. For readability, the 
original sentences that cover the MUs are selected. 
For each set of similar MUs, the longest sentence 
in user-preferred  is displayed. The 
display order of the selected sentences is 
determined by relative position in the original 
news articles. 
In the browsing lnodel, the news articles are 
listed by information decay. The first news 
article is shown to the user in its whole content. 
In the latter shown news articles, the MUs 
denoting the inforlnation mentioned before are 
shadowed (or eliminated), so that the reader can 
focus on the new information. The alnount of 
information in a news article is lneasured in terms 
of the number of MUs, so that the article that 
contains lnore MUs is displayed before the others. 
For readability, a sentence is a display unit. In 
this model, users can read both the COlnmon views 
and different views of reporters. It saves the 
reading time by listing the colnlno11 view only 
once. 
5. Evaluation of Sumnmrization Results 
The same six events specified in Section 3.1 
are used to measure the performance of the two 
summarization models. Three kinds of metrics 
are considered - say, the document reduction rate, 
the reading-tilne reduction rate, and the 
inforlnation carried. The higher the document 
reduction rate is, the more time the reader may 
save, but the higher possibility the ilnportant 
information may be lost. Tables 2 and 3 list the 
document reduction rates for focusing and 
browsing summarization, respectively. Only 
focuses are displayed in focusing sutnmarization, 
Table 2. Reduction Rates for 
Focusing Summarization 
Event Name l)ocl,en Sum Len Sum/l)oc \[ Reduction 
mililary service 7658 2402 0.3137 68.63% 
construction permit 4182 1226 0.2932 70.68% 
laMslide in Shah ,It" 5491 1823 0.3320 66.80% 
Busies sons 6186 924 0.1494 85.06% 
Typhoon Babis 4068 1460 0.3589 64. I 1% 
stabilization ftmd 8434 2243 0.2659 73.41% 
Average 36019 10078 0.2798 72.02% 
Table 3. Reduction Rates 
for Browsing Summarization 
Event Name Doc Len Sum Len + Sum/l)oc Reduction 
military service 7658 2716 0.3547 64.53% 
construclion permit 4182 2916 0.6973 30.27% 
landslide in Shah Jr 5491 2946 0.5365 46.35% 
Buslfs sons 6186 5098 0.8241 17.59% 
Typhoon Babis 4068 2270 0.5580 44.20% 
stabilization fund 8434 4299 0.5(197 49.03% 
Average 36(/19 20245 0.5621 43.79% 
Table 4. Assessors' Evaluation 
Event Name Document Question- Reading-Time 
Reduction Answering Reduction 
Rate Correct Rate Rate 
military service 64.53% 10(1% 45.24% 
3/I.27% 33.33% 33.54% construction permit 
landslide in Shah J|" 46.35% 80% I 10.28% 
gush's soils 17.59% 100% I 36.49% 
Typhoon Babis 44.20% 100% 35.10% 
stabilization fund 49.03% 100% 18.49~ 
Average 43.79% 88.46% 3(/.86% 
so that the average doculnent reduction rate is 
higher than that of browsing summarization. 
Besides the document reduction rate, we also 
measure the correct rate of question-answering, 
and reading-time reduction rate. Assessors read 
the highlight parts only in the browsing 
summarization, and answer 3 to 5 questions. 
Table 4 lists the evaluation results of the six 
events. The average doculnent reduction rate is 
43.79%. On the average, the summary saves 
30.86% of reading time. While reading the 
summary only, the correct rate of question- 
answering task is 88.46%. 
Conclusion 
This paper sketches architecture for 
multilingual news summarizer. In multilingual 
clustering, lnatching all pairs of news clusters in 
all s is time-exhaustive. Because only 
English and Chinese news articles are considered 
in this paper, it is not a problem. In general, an 
164 
effective way is to predefine a sequence of 
 pairs according to the degree of 
translation ambiguity. The hmguage pair of less 
ambiguity is tried first. 
To discuss which fi'agments of multilingual 
news stories denote the salne things, this paper 
defines the concept of MUs. Punctuation marks, 
linking elements and topic chains are cues to 
identify MUs for Chinese. Select-high-frequent 
English translation and name transliteration are 
adopted to transhtte Chinese MUs into L;nglish. 
Five models are proposed to link the similar MUs 
together. Different formats used in time, date 
and monetary expressions, e.g., implicit time zone, 
affect the performance of linking. It should be 
studied in the fllture. 
In presentation o1' summarization results, the 
information decay strategy helps reduce the 
redundancy, and the user can get all the 
information provided by the news sites. 
However, the news sequence is not presented 
according to the importance. The user may quit 
reading and miss the information not shown yet. 
The voting strategy from reporters gives a shorter 
summarization in terlnS of user-preferred 
s. However, it also misses some unique 
information reported only by one site. A hybrid 
strategy should be developed in the future to meet 
all the requirements. 

References 

Chen, H.H. (1994) "The Contextual Analysis of 
Chinese Sentences with Punctuation Marks," Litelwl 
aud Linguistic Computing, Oxford University Press, 
9(4), 1994, pp. 281-289. 

Chen, H.H; et al. (1998a) "Description of the NTU 
System Used for MET2." Proceedings of 7th 
Message Undertanding Conference, 1998. 

Chen, H.H.; et al. (1998b) "Proper Name Translation in 
Cross-Language Information Retrieval," Proceedings 
of COLING-ACL 98, 1998, pp. 232-236. 

Chen, H.H.; Bian, G.W. and Lin, W.C. (1999) 
"Resolving Translation Ambiguily and Target 
Polysemy in Cross-Language Information Retrieval," 
Proceedings of 37th Annual Meeting of the 
Association for Computational Linguistics, 1999, pp. 215-222. 

Chert, K.H. and Chert, H.H. (1995)"A Corpus-Based 
Approach to Text Partition," Pivceedings of 
International Col!/'erenee of Recent Advances on 
Natural Language Processing, Tzigov Chark, 
Bulgaria, 1995, pp. 152-160. 

Chen, ILH. and Huang, S.J. (1999) "A Sunnnarization 
System for Chinese News from Multiple Sources," 
Proceedings of 4 't' International Workshop on 
lqformation Retrieval with Asia l~nguages, 1999, pp. 
1-7. 

Chen, H.H. and Lee, J.C. (1996) "Identification and 
Classification of Proper Nouns in Chinese Texts," 
Proceedings of 16th International Conference on 
Computational Linguistics, 1996, pp. 222-229. 

Hovy, E. and Marcu, D. (1998a) Automated Text 
Summarization Tutorial in 17th ACL and 36th COLING, Montreal, Quebec, Canada, 1998. 

Hovy, E. and Marcu, D. (1998b) Multilingual Text 
Summarization, Tutorial in AMTA-98, 1998. 

IA, C.N. and Thompson, S.A. (1981) Mandarin 
Chinese: A Functional Re\[erence Giwmmar, 
University of California Press, 1981. 

Mani, I. and Bloedorn, E. (1997) "Multi-documenl 
Summarizalion by Graph Search and Matching," 
Proceedings of the Fourteenth National Con.lisrence 
oil Arti/icial Intelligence, Providence, RI, pp. 622- 
628. 

MUC (1998) Ptvceedings of 7 ~1' Message 
{hMet:s'tanding Cot!ferenc.e, http://www.muc.saic. 
corn/proceedings/proceedings index.broil. 

Radev, D. R. and McKeown, K.R. (1998) "Generating 
Natural Language Summaries from Multiple On-Line 
Sources," Computational Linguistics, Vol. 24, No. 3, 
pp. 469-500. 

Yang, Y. (1981) The Research on lhmetuation Marks, 
Tian-iian Publishing Company, ltong Kong, 1981. 
