SESSION 8A: MACHINE TRANSLATION 
Jerry R. Hobbs, Chair 
Artificial Intelligence Center 
SRI International 
Menlo Park, California 94025 
The three papers in this session exemplify three different 
approaches to machine translation. Each addresses a 
significant problem encountered within the approach. 
In the dark ages of machine translation, it is said that re- 
searchers attempted a Direct Approach--word-for-word 
substitution. But this could not have survived much be- 
yond the first inspection of the output. The most com- 
mon approach adopted in working machine translation 
systems is the Transfer Approach illustrated in Figure 1. 
A text in a source language is analyzed to some depth, 
producing, for example, a parse tree or a logical form for 
the sentences in the text. Then transfer rules are applied 
to this representation to produce a representation at the 
corresponding level for the target language. The text is 
then generated in the target language. 
Source Direct Target 
tt Language Language 
Analysis 
of 
Source 
Transfer 
Generation 
of 
Target 
Figure 1: Transfer Approach to Machine Translation. 
A French-to-English translation system might have a 
transfer rule like the following. The French phrase 
la r@mun@ration du temps suppl6mentaire 
has the following syntactic structure: 
\[NP la r@mun@ration \[pp du 
\[NP temps suppl@mentaire\]\]\] 
The corresponding English phrase 
overtime pay 
has the syntactic structure 
\[Np \[Novertime\] pay\] 
The transfer rule would specify how fragments of the 
French parse tree mapped into the corresponding frag- 
ment of the English parse tree, for example, how the PP 
in French maps into the prenominal noun in English. 
The rule would be stated with whatever lexical general- 
ity is appropriate. 
There are a number of classical problems with the Trans- 
fer Approach, principally arising when the two languages 
express the same concept in very different ways syntac- 
tically. For example, what is expressed by the main verb 
in one language may be expressed adverbially in another. 
Conjunction reduction may be possible in one language, 
while lexical factors make it impossible in another. The 
paper by Kinoshita et al. addresses many of these prob- 
lems and describes how they can be handled within the 
Transfer Approach. 
Another approach, long advocated but rarely given ex- 
tensive implementation, is the Interlingua Approach, il- 
lustrated in Figure 2. Here one does a much deeper 
analysis on the source-language text, to the level of a 
language-independent conceptual representation called 
Interlingua. There are two advantages to this approach. 
First, a text must often be analyzed to a conceptual level 
in any case to achieve an adequate translation. The text 
must be understood before it can be translated. Sec- 
ond, when there are many languages one must translate 
among, we need only define the mapping between each 
language and the Interlingua, rather than specifying the 
transfer rules for every pair of languages. 
One problem with the Interlingua Approach is the dif- 
ficulty, if not impossibility, of devising an adequate In- 
terlingua. The English word "wall" seems like a per- 
fectly straightforward primitive concept to an English- 
only speaker. But when translating into French, we must 
distinguish between walls seen from the inside and walls 
seen from the outside. Will every new language we add 
to an Interlingua-based machine translation system force 
253 
Source Target 
Language Language 
Analysis Generation 
of of 
Source <ans~ Target 
Interlingua 
Figure 2: Interlingua Approach to Machine Translation. 
on us further similar fragmentation of what we had be- 
lieved were primitive concepts? The issue of how to con- 
struct an Interlingua incrementally is the topic of the 
paper by Hovy and Nirenburg. 
Recently a new approach to machine translation--the 
Statistical Approach--has been attempted, principally 
at IBM Yorktown. In order for a statistical approach to 
work, there must be enough data available encoding the 
relevant information. This forces one to simplify the un- 
derlying model of the languages and the transfer between 
them to the point where statistical analysis becomes fea- 
sible. 
In earlier work by Brown and his colleagues, the fol- 
lowing simplifications were used. Instead of having a 
grammar of the source language, no analysis was done 
of the source text. Instead of having a grammar of the 
target language, a trigram model of the target language 
was used, capturing some but by no means all the struc- 
ture of the target language. In place of transfer rules 
there were, first, a model of lexical correspondences be- 
tween the two languages, and, second, a model of the 
alignment of corresponding sentences. The alignment 
model, rather than using the richer information about 
correspondences between parse trees that transfer rules 
seek to capture, encodes the simpler structural relations 
illustrated in Figure 3. 
This approach is then illustrated in Figure 4. 
There is no reason in principle that a Statistical Ap- 
proach need be this shallow, and in fact in more recent 
work, the IBM Yorktown team has attempted to incorpo- 
rate more linguistic structure into their efforts. As large 
la rdmundration du temps suppl~mentaire 
overtime pay 
Figure 3: Alignment Information. 
bracketed corpora become available, it may be possible 
to align tree fragments as words are aligned today and 
derive transfer rules statistically. 
A problem that everyone who deals with real-world text 
eventually encounters is the problem of how to analyze 
very long sentences. This has especially engaged the 
attention of researchers in text understanding in the last 
several years. The solution is to break the sentences into 
phrases in the "right" way, whatever that is. 
Statistical approaches to translation are computation- 
intensive, and hence sentences become very long a lot 
sooner than in other approaches. The problem of break- 
ing sentences into phrases that can he translated inde- 
pendently must be faced a lot earlier. The paper by 
Brown et al. in this volume addresses this problem in 
the statistical translation framework. 
Source Target 
Language Transfer Lanl :uage 
Trigram 
Model 
I 
Interlingua 
Figure 4: Statistical Approach to Machine Translation. 
254 
