SESSION 3: MACHINE TRANSLATION 
Jaime Carbonell 
Carnegie Mellon University 
Center for Machine Translation 
Pittsburgh, PA 15213 
Machine Translation (MT) technology has progressed 
significantly since the days of the ALPAC* report. In 
particular, multiple paradigms are being investigated ranging 
from statistical methods to full knowledge-based interlingual 
MT systems. Much of the recent work is based on advances in 
natural language processing since ALPAC in the 1960s, 
including: 
• Semantic analysis to resolve lexical and syntactic 
ambiguities during parsing, and thus reduce translation 
errors very significantly. 
• Unification grammars allowing syntactic and 
semantic constraints to be checked in a unified manner 
while parsing, and permitting reversible grammars--i.e., 
the same grammars to be used for generation as well as for 
analysis. 
• Advanced parsing methodologies, including 
augmented-LR compilation where knowledge sources 
(syntactic grammars, lexicons, and semantic ontologies) 
can be defined and maintained separately but are jointly 
compiled to apply simultaneously at run time, both in 
parsing and in generation. 
• Natural language generation, focusing on how to 
structure fluent target-language output, an activity not 
truly investigated in the pre-ALPAC days. 
• Automated corpus analysis tools, statistical and 
other means of extracting useful information from large 
bi- or multi-lingual corpora, including collocations, 
transfers, and contextual cues for disambiguation. 
• MRDs => MTDs, use of electronic machine-readable 
dictionaries (MRDs) to partially automate the creation of 
machine-tractable dictionaries (MTDs) in processable 
internal form for parsers and generators, permitting 
principled sealing up in MT configurations. 
APPROACHES TO MODERN MT 
In light of these advances, several major MT paradigms have 
evolved to supplant the early hand-coded direct-transfer 
methods. One approach is purely statistical, as practiced at 
In 1965 the United States Academy of Science cemmissoned a study of he 
state of the art in Machine Translation, whose findings were published the 
following year and become popularly known as the ALPAC report. In 
essence, ALPAC argued that there was insufficient scientific basis in natural 
langauge processing to perform reliable machine t~anslatinn, and the large 
expensive computers of the time would make NIT eeonornically infeasible. 
Both situations have since changed drastically, invaLidating the ALPAC 
conciusions. In fact, DARPA has played a major role in fostering the 
development of the NLP scientific infrsstmcture in the post-ALPAC years. 
IBM, in which the direct-transfer paradigm is still king and 
translation is viewed as transduction between two character (or 
word) streams--essentially two encodings of the same 
message. However, the direct transfer rules are totally learned 
by statistical analysis of large bi-lingual corpora, rather than 
laboriously and incompletely hand-coded. A drawback of the 
statistical approach, of course, is that it carmot guarantee the 
accuracy of any textual passage being translated, but rather 
strives to minimize the total number of errors over time. 
Another MT approach is to provide a measure of analysis of 
the source language prior to transfer. At minimum, 
morphological and syntactic analysis is performed, then the 
transfer component transforms the parse trees into 
corresponding parse trees in the target language with 
appropriate lexical substitution. These transformed parse trees 
are then used to generate the desired target texts. Performing 
analysis and generation reduces the size of the transfer 
component, which is a major benefit, considering that 
translation across N languages requires O(N 2) transfer 
grammars. Transfer at the syntactic level represents the 
classical approach on which most commercial attempts at MT 
are based. The problem with classical transfer is that it too 
makes a significant number of errors in the translation, 
primarily through its inability to reduce much of the lexical and 
syntactic ambiguity of the source language texts. 
A deeper analysis, including semantic restrictions to produce 
case frames (rather than parse trees), reduces both the number of 
errors (as some ambiguity has been resolved) and the size of the 
transfer component (case frame representations in different 
languages will be much more similar to each other than 
syntactic parse trees). Some Japanese firms working in 
machine translation, for instance, have adopted this approach, 
which they call semantic transfer. Since Japanese and English 
are more different than two Indo-European languages, there is 
more justification for the deeper level of analysis and the desire 
to minimize the size of the transfer component. 
The deepest level of analysis produces an underlying non- 
ambiguous semantic representation, independent of both source 
and target language. This is called the interlingua orpivot 
approach, and it trades off much more work at analysis and 
generation for no work at all in the transfer phase. Benefits 
include much lower errors (as ambiguities must be resolved to 
produce interlingua), and no N 2 problem, as there is no transfer 
component. Thus the interlingua approach is particularly well- 
suited for multi-lingual translation. The most serious problem 
with the interlingual approach is its requirements for vast 
knowledge bases if one desires general-purpose, highly 
accurate translation in any domain. Specialized-domain 
interlingual systems are far more practical. 
139 
EVALUATION METHODS 
Yorick Wilks, one of America's foremost MT researchers, 
states that "Machine translation evaluations methods are better 
developed than machine translation itself." In a sense, he is 
right. Since MT is a complete throughput process from source 
to target text that corresponds precisely to the task of human 
translations, it is not difficult to compare the two and provide 
relative assessments. MT has been evaluated with respect to 
semantic accuracy of the translation and intelligibility of the 
final output. But, other factors such as degree of automation are 
equally relevant. The more human intervention (e.g., pre and 
post editing) required to produce a good translation, the less 
useful the MT system. 
MT evaluations must always be made task-relative. On the 
one hand, MT for the sole purpose of scanning translated texts 
in order to establish their relevance to a given topic must be 
fast, with little if any human assistance, but can be rough, 
partially inaccurate and of low-legibility output. On the other 
hand, technical or legal texts translated for publication must be 
accurate and legible, although slower processing and additional 
human assistance may be tolerated. Therefore, one challenge 
for the DARPA MT research community is to develop more 
appropriate, task-sensitive and comprehensive evaluation 
criteria. 
PREVIEW OF MT PAPERS 
The three papers on machine translation in this section cover 
only part of the DARPA MT effort--and this should be 
interpreted as a sign of the breadth of coverage and DARPA 
interest in the field, pursuing different technologies at different 
sites. In particular, the knowledge-based interlingua 
approaches of CMU, CRL and ISI are not represented, but form a 
major component of the overall program. Those areas 
represented in this volume include translation as abduetive 
reasoning at SRI and two papers on the role of statistics in 
machine translation. More precisely, statistical word-sense 
disambiguation at IBM and establishing lexical-transfer 
correspondences for MT at AT&T are discussed in some detail. 
Jerry Hobbs and Megurni Kameyama at SRI extend their 
existing TACITUS architecture for abductive natural language 
interpretation and apply it to the task of translating a few 
examples into Japanese. The work is exciting in the sense of 
showing how an existing system and underlying theory are 
sufficiently general to address the new task: machine 
translation across radically different languages. Other issues, 
such as sealing up, are not yet addressed in this work. 
Peter Brown, Stephen Delia Pietra, Vincent Della Pietra and 
Robert Mercer at IBM developed a statistically-based word- 
sense disambiguation method as an integral component of a 
statistical machine translation system. In fact, statistical help 
in word disambiguation may prove to be of major help in more 
traditional MT approaches (transfer and interlingua) when 
definitive semantic and syntactic knowledge do not narrow 
down word senses to a single candidate. Therefore this work 
should be followed closely by those researchers of the non- 
statistical-MT persuasion as well. 
William Gale and Ken Church develop a mathematical 
infrastructure for determining lexical correspondences across 
words in parallel ~,-~s. Parallel text means that the same text is 
available in two (or more) languages. Establishing lexical 
correspondences is crucial for building knowledge bases for 
symbolic translation methods (whether transfer or interlingual) 
and for automated training of statistical translation methods 
such as that advocated at IBM by Brown, Mercer, et al. Gale and 
Church argue that their methods are statistically more reliable 
than the earlier IBM methods for establishing word 
correspondences. 
140 
