HOW COULD RHETORICAL RELATIONS BE 
USED IN MACHINE TRANSLATION? 
(AND AT LEAST TWO OPEN QUESTIONS) 
Ruslan Mitkov 
Machine Translation Unit 
University of Science Malaysia 
11800 Penang, Malaysia 
Fax (60-4) 873335, 
Email ruslan@cs.usm.my 
My position paper addresses more or less Workshop question No. 5: "How are 
rhetorical relations used in discourse understanding? How are linguistic clues and 
word knowledge brought to bear?" 
The paper shows how rhetorical relations could be used in Machine Translation 
(MT). It introduces in brief, a discourse-oriented approach for MT which uses 
schemata of rhetorical predicates for describing the structure of a paragraph. At the 
same time, it poses at least two questions (in my opinion practically unsolved 
problems): 
1) How can rhetorical predicates be computationally recognized? 
2) Are the so far defined predicates sufficient and precise enough to describe the 
real world? 
INTRODUCTION: DISCOURSE-ORIENTED MACHINE TRANSLATION 
The discourse-oriented MT should be regarded as a very important research topic, 
since it is expected to make the translation more natural in MT systems. 
Unfortunately, not much attention has been given to this problem yet and the 
availability of a discourse component in a MT system has been reported very briefly 
in \[7\] only. 
Most of the MT systems perform sentence-by-sentence translation. Only a few try to 
translate paragraph-by-paragraph and in these cases, the discourse structure of the 
output language is identical with that of the input language. However, I have shown 
that the discourse structures across the different sublanguages are not always the 
same for any pair of natural languages \[5\]. 
Paragraph-by-paragraph machine translation seems to be for now, an unjustifiably 
complicated task for practical needs. It involves the complete understanding of the 
paragraph, the determination of discourse topic(s), goals, intentions, so that the 
output can be produced in accordance with the respective discourse rules and 
purposes. However, recognizing topic, goal, intention by a computer program seems 
to be a very tough problem. Moreover, analyzing a paragraph is itself a complicated 
task which does not always yield satisfactory results. 
86 
On the other hand, translating sentence-by-sentence with the sequence of the 
original sentences preserved is a general approach, which guarantees in most of the 
cases an understandable output. However, in order for a translated message to sound 
as natural as possible, it should be conveyed in accordance with the discourse 
organization rules of the target language. If we examine more closely the work of a 
professional translator, we shall inevitably note that he/she does not always follow 
the order of sentences in the source text. 
Taking into account the complexity of paragraph understanding and the necessity of 
observing the specific target sublanguage rules, I have been working on a practical 
discourse-oriented MT approach (within an English to Malay MT system) which 
analyzes a source paragraph as a schema of rhetorical predicates and generates the 
target text possibly as another schema of rhetorical predicates. Towards this end, I 
have developed a Text Organization Framework Grammar which maps source 
paragraph structures of rhetorical predicates into the specific target paragraph 
structures of rhetorical predicates \[6\]. 
SELECTION OF TEXT ORGANIZATION APPROACH 
I have been studying different approaches which have been so far used to describe 
the organization of a given text (paragraph). From a practical point of view, I argue 
that the most appropriate approach would be the "schemata-based approach" 
introduced by K. McKeown \[3\] and used by other researchers. 
Though some researchers point out the relatively missing flexibility of this 
approach, I found this approach more suitable for the needs of MT. The plan-based 
approach \[4\] seems to be too complicated and unrealistic to be implemented in an 
MT system because its rhetorical relations are dependent on an expected effect on 
the hearer achieved by their combination. In a MT system, as already mentioned, it 
is very hard, if practically not possible, to recognize automatically in a paragraph the 
goals and intentions of the speaker. 
SUBLANGUAGES AND SCHEMATA 
In the sublanguages I studied, however, I found out that the schemata of rhetorical 
predicates could not be always uniquely defined. There are sublanguages where 
more than one typical schema should be defined and consequently used. I examined 
numerous texts on which basis I defined "stable schemata". The schemata S 1, $2 .... 
SN can be considered "stable" if 1) SI/N~5, VI and 2) ESI/N'~y where N is the 
number of all examined texts, 5, '1' are numbers in the interval (0,1) which we call 
"individual contribution minimum" and "global contribution minimum" 
respectively. The idea behind these mathematics is that schemata can be considered 
as "stable" if they as a whole represent a significant portion of all examined texts 
and yet every "stable" schema should be itself representative. 
For translation from English into Malay, if more than one stable schema is available 
in the respective sublanguage, the stable schema, which is closest to the input of 
English text is chosen. For determining closeness, special metrics has been 
developed which takes into account not only the number of displaced predicates, 
but also the size of the displacement and the maximal length of matched substfings 
from the input and output schemata of rhetorical predicates. 
We have studied the discourse structure of a few sublanguages (for both English and 
Malay), potential candidates for translation domains in our MT system: the 
87 
sublanguages of job vacancies, residential properties for sale, cars for sale and 
education advertisements from different newspapers in English and Malay. 
From our investigations on these sublanguages we have drawn three main 
conclusions: 
1) The stable schemata for English and Malay are not always identical and do not 
occur equally frequent 
2) For some sublanguages there are more than one stable schema 
3) For some sublanguages there exists no stable schema 
These conclusions are important for MT because in the third case there is no need 
for discourse transition rules and the translation should be undertaken sentence-by- 
sentence. 
THE BIG PROBLEM: IDENTIFICATION OF RHETORICAL PREDICATES 
During the analysis, rhetorical predicates should be recognized. In certain 
sublanguages this can be done by means of key words and other clues \[5\]. However, 
in general this seems to be a very complicated problem and extensive world 
knowledge and inferencing mechanisms are needed. How could a program 
recognize a sentence (proposition) as amplification, attributive, etc. rhetorical 
predicate? For our sublanguage-based MT needs, I am considering two approaches 
for the identification of rhetorical predicates. 
One approach would be to define "verb frameworks" characteristic of a verb within 
the sublanguage. Each verb should be associated with possible rhetorical predicates 
and the predicate should be identified on the basis of the logical structure of the 
analysis. However, this approach may not be powerful enough in certain cases. 
Consider the sample text from \[2\] describing Kyushu Daigaku (Kyushu University): 
"A national, coeducational university in the city of Fukuoka. Founded in 1910 as Kyushu 
Imperial University. It maintains faculties of letters, education, law, economics, science, 
medicine, dentistry, pharmacology, engineering, and agriculture. Research institutes include 
the following: the Research Institute of Balneothempeutics, the Research Institute of Applied 
Mechanics, the Research Institute of Industry and labor, and the Research Institute of 
Industrial Science. Enrollment was 9,425 in 1980". 
It will be quite difficult, however, using only verb framework, to recognize the first, 
the third, fourth and the last sentence as rhetorical predicates. An useful approach in 
this case would be to use a domain knowledge which would enable the recognition 
of the rhetorical predicate after a semantic analysis. For instance a proposition 
describing entities which are in 'sub-part' relation should be classified as a 
constituency predicate. This 'sub-part' relation could be easily recognized, provided 
it has been already described in the domain knowledge base. Consider again the 
sample text under the assumption that such a knowledge base exists. In this case, 
from the 'is-a' relation ("Kyushu Daigaku"- "University"), from the respective 'sub- 
part relations' ("university"-"faculty", "research centre") and the 'has' relation 
("university" -"enrollment of students"), the program could assign to the above 
sentences identification (1. sentence), constituency (3., 4. sentences) and attributive 
(last sentence) predicates, respectively. 
Consider, however, the second sentence. Is it "amplification"? If yes, how is the 
program supposed to conclude that this sentence is an elaboration of the first one? 
How feasible is in general the computational recognition of the rhetorical 
88 
predicates? And here comes an important question: how much domain and world 
knowledge, as well as AI inferencing techniques, are needed? 
And if yes, does not it seem that "amplification" is not fine and precise enough (I 
can give many examples of propositions to which the rhetorical predicate 
"amplification" is to be assigned, because they simply do not fit the def'mition of the 
rest of the predicates)? Should not one introduce an additional predicate called e.g. 
"initiation" which would be associated with the act of founding, setting up, opening, 
organizing etc. something? This gives a rise to a second important question. Is the 
set of rhetorical predicates given in \[1\], \[3\], \[8\] or \[9\] sufficient and precise enough 
to describe the real word? But if we propose additional predicates, how far should 
we go? 

REFERENCES 
\[1\] Grimes J. - The thread of discourse. Mouton, The Hague, 1975 

\[2\] Kodansha Encyclopedia of Japan, Vol.4, Kodansha Ltd., Tokyo, 1983 

\[3\] McKeown K. - Text generation: using discourse strategies and focus constraints 
to generate natural language text. Cambridge University Press, 1985 

\[4\] Mann W., Thompson S. - Rhetorical structure theory: description and 
construction of text structures. In Kempen G. (Ed.): "Natural language 
generation: new results in artificial intelligence, psychology and linguistics", 
Dodrecht, Boston, 1987 

\[5\] Mitkov R. - Multilingual generation of public weather forecasts, Proceedings of 
the SPICIS'92 (Singapore International Conference on Intelligent Systems) 
Conference, 28 September-1 October 1992, Singapore 

\[6\] Mitkov R. - Discourse-based approach in machine translation From Proceedings 
of the International Symposium on Natural Language Understanding and 
Artificial Intelligence, Fukuoka, Japan, 13-15 July, 1992 

\[7\] Nirenburg S. - A distributed generation system for Machine Translation: 
Background, Design, Architecture and Knowledge Structures, CMU-CMT-87- 
102, Pittsburg, 1987 

\[8\] Shepherd H. - The fine art of writing. Macmillan Co., New York, 1926 

\[9\] Williams W. - Composition and rhetoric. D.C. Heath and Co., Boston, 1983 
