Alan K. Melby 
Brigham Young University 
Dept. of Linguistics 
Provo, Utah 84602 
USA 
ABSTRACT 
One of the necessary tasks of a machine 
translation system is lexical transfer. In 
some cases there is a one-to-one mapping from 
source language word to target language word. 
What theoretical model is followed when there 
is a one-to-many mapping? Unfortunately, 
none of the linguistic models that have been 
used in machine translation include a lexical 
transfer component. In the absence of a 
theoretical model, this paper will suggest a 
new way to test lexical transfer systems. 
This test is being applied to an MT system 
under development. One possible conclusion 
may be that further effort should be expended 
developing models of lexical transfer. 
i. An Early Approach to Lexical Transfer 
Years before the machine translation 
community was burdened with guilt by the 
ALPAC report, David Hays, former chairman of 
the International Committee on Computational 
Linguistics, proposed a procedure for lexical 
transfer (Hays, 1963, pp 205-208). We will 
describe it, quoting pieces to preserve the 
original flavor. 
i. a. File of Equivalent-Choice Data 
"Most words...have uniform translations, but 
not all." "These exceptions to the general 
rule must be discovered and taken into 
account. The procedure is simple and 
straightforward. A file of equivalent-choice 
data...is required." This file is prepared 
usina real text. When a word is encountered 
for the first time, one translation is 
selected and entered into a bilingual 
glossary. When the same word is encountered 
again, the human translator/editor attempts 
to use the translation already in the 
glossary. Additional translations are added 
only when the one(s) in the glossary are not 
acceptable. This procedure is supposed to 
avoid entering interchangeable alternatives 
that are only stylistic variations. The file 
of equivalent-choice data mentioned above is 
a record of how many times each translation 
was used. 
i. b. Category 
Once the equivalent-choice file has been 
compiled, the first step in analyzing it is 
to identify those words with two or more 
translations (i.e. equivalents). The next 
step is to identify whether the translation 
is governed by "grammatic category". 
i. c. Function 
If there are two or more translations within 
the same category, then the analyst looks at 
"grammatic function" (e.g. subject, object of 
preposition, etc). 
104 
I. d. Features 
"If there is any kind of relation in which 
the word has two equivalents, the analyst 
continues by examining each word that governs 
the multiple-equivalent word." From the 
words that govern the word in question, the 
analyst derives "criterial classes" (i.e. 
features to mark on the governing words in 
the dictionary). 
i. e. End of Procedure 
But what if testing features on other words 
is not sufficient to determine the 
translation? Hays recognizes this 
possibility by noting "There is no certainty, 
of course, that the governors and dependents 
of an occurrence determine its translation, 
but is seems plausible that they will often 
do so." Even following this procedure of 
assigning features may be anything but 
"simple and straightforward", and (even 
worse) it is not sufficient. 
2. Some Counterexamples 
Obvious counter-examples to Hays' hope come 
to mind, such as "chip" in micro-electronics 
(=integrated circuit) and in gambling (=token 
to represent money), or "buck" in hunting 
(=male deer) and in slang (=dollar). Even 
older systems like SYSTRAN provide for this 
by prioritizing various domain-specific 
dictionaries. 
The other situation where the Hays approach 
clearly breaks down is when the word is part 
of a fixed expression, such as "chip off the 
old block" (=like his father) or "pass the 
buck" (=avoid responsibility). All machine 
translation systems provide for this by 
including expression dictionaries that 
override word-for-word lexical transfer. 
A variation on the expression dictionary is 
to key a lexical-structural transfer from a 
single word that affects surrounding words. 
For example, the adjective "hungry" 
stimulates a transfer going into French that 
changes "x is hungry" to "x has hunger',. 
Unfortunately, sometimes all the tricks 
listed above combined still do not suffice to 
identify an acceptable translation. The 
following two examples and some of the 
subsequent discussion are adapted from Melby 
(1985). 
2. a. The "plate" Example 
For example, consider the word "plate" (which 
can mean a household item on which food is 
placed to be eaten or a marker on the ground 
in the game of baseball) in the sentence 
There was an egg on the plate 
in the context of a discussion of a baseball 
game in which an angry fan threw a raw eqq 
and it landed on home plate. 
Then consider the sentence 
She threw the food on the plate 
in the discussion of what a teenager did 
because she was in a hurry to fix her 
breakfast and get to school. 
Either of these sentences could occur in a 
narrative of the life of a young person (so a 
special dictionary Will not help), and there 
is apparently nothing in the syntactic or 
lexical environment of "plate" that 
determines the translation, and no idiomatic 
expressions are involved. 
2. b. The "rock" example 
Or consider the problem of "rock", which is a 
single lexical item in English with several 
specific translations in French ("pierre, 
roc, caillou", etc.), no one of which is as 
general as the English word "rock", which 
would be translated differently in each of 
the following sentences: 
She found the rock on the beach and 
placed it in her pocket. 
He climbed up and sat on the rock to get 
a better view. 
While watching the parade, she got tired 
and sat down, not seeing the sharp rock, and 
screamed from the pain. 
2. c. Opinions 
Some claim that the above examples are far 
fetched and that the identification of 
lexical category, grammatical function, 
general subject matter, and fixed expressions 
is sufficient to develop lexical transfer 
algorithms that produce acceptable 
translations. 
Others say that the dynamic nature of natural 
languages will often present a significant 
number of cases where lexical transfer is not 
handled adequatelyby standard techniques. 
Consider, for example, the search for lexical 
transfer criteria for the English preposition 
"through" going into French found in 
Bourquin-Launey (1984). 
The facts are (i) that raw machine 
translation output, even after these many 
years of development, is seldom up to 
publication standards without post-editing; 
and (2) there has been little development of 
the models of lexical transfer beyond the 
stage described in section one (which was in 
place already in the mid 1960's). One reason 
is that the linguistic models that have been 
used do not include a lexical transfer 
component. This can be verified by looking at 
Hutchin's updated survey of linguistic models 
in machine translation (Hutchins, 1984). 
To summarize the discussion to this point, 
the fact is that work in lexical models has 
been neglected for the past twenty years; the 
question is whether that neglect is 
justified. 
It seems that it would be worthwhile to at 
least examine the nature of the problems in 
machine translation output. If it turns out 
that a significant portion of the problems 
are actually failures in lexical transfer, 
then further studies of lexical transfer in 
the computational linguistics community 
should be encouraged. 
The following section describes an on-going 
effort to test the lexical transfer component 
of a machine translation system currently 
under development. The method allows a test 
of lexical transfer even before the entire 
system is operational, thus providing 
feedback as early in the life of the project 
as possible, hopefully allowing design 
changes to be made, J f they are needed, 
before they become too costly. The method 
could also be adapted to studies in lexical 
transfer somewhat independent of a particular 
machine translation system. 
3. A Method of Testing Lexical Transfer 
3. a. Origin of the Testing Project 
The BSO Company is a systems house in 
Utrecht, The Netherlands. The BYU-HRC is a 
center which promotes research in the College 
of Humanities at Brigham Young University 
(BYU), in particular providing support for 
research involving language and computers. 
In 1984, the author replied to a request for 
comments on the then proposed BSO machine 
translation project. That reply led to 
discussions which resulted in an agreement 
between BSO and the BYU-HRC to create a text 
and translation data base as a joint 
venture. 
The specifications for the data base were 
that it would consist:: of paired texts in 
French and English, that it would include at 
least 500,000 words in each language, that 
the translations would be done by qualified 
professional translators, and that the text 
type would be straightforward modern English 
and French avoiding texts that are literary 
or intentionally ambiguous. 
BSO supplied the source documents, which were 
mostly public reports on agriculture, social 
conditions, etc., published by the 
European Economic Community (EEC) or the 
United Nations (UN). 
The documents were placed in machine readable 
form using the Kurzweil OCR device and 
transferred to a disk pack on an IBM 
370/138. Preliminary to the test, the data base was 
split into two parts. A smaller part (about 
200,000 words in each language) is accessible 
to BSO for syntactic studies. The larger 
part (about 300,000 words in each language) 
will be used for the lexical test, under 
105 
control of BYU, and will not be accessible to 
BSO. The test will be executed in five 
steps. 
3. b. Secret selection of sample texts 
From the lexical part of the data base 
(300,000 words of English and 300,000 words 
of French), BYU has selected 4 sample texts. 
Each of these samples is a coherent stretch 
of text of approximately 300 word tokens. 
Unless otherwise indicated, we now refer to 
the English version of each sample text. 
From the 4 texts there were about 600 non- 
function words. Adding some "misleaders" 
from other sections of the data base, a list 
of 800 words was sent to BSO. 
3. c. Construction of trial lexicons 
BSO has received the alphabetically sorted 
list of 800 English words. They do not know 
the context of any of the words, nor will 
they know whether a word is part of one of 
the secret sample texts or merely a 
"misleader". 
BSO is now building an English to 
intermediate language (IL) lexicon and an IL 
to French lexicon. In the case of the BSO 
project, the intermediate representation is a 
formalized Esperanto. The English-IL lexicon 
will have more "exits" than input words 
(estimate -- 1200). So the IL-French lexicon 
will have about 1200 entries. 
The English-IL and IL-French lexicons will 
each be separately tested at BSO. For this 
purpose, BSO will use English and IL trial 
input texts, especially written for this 
purpose by external consultants based on the 
800 and 1200 word lists. In addition to the 
800 words, BSO will add a few hundred 
function words (articles, pronouns, 
prepositions, etc.). 
3. d. Translation using the trial lexicons. 
After completion of both trial lexicons by 
BSO, they will be sent to the USA for overall 
testing by BYU. Two translation phases will 
be distinguished as part of the testing 
procedure. 
First, monolingual students of English will 
translate the four sample texts (see A) into 
IL by mechanically following the rules 
contained in the entries of the English-IL 
trial lexicon. We emphasize that these 
students will have no knowledge of IL, which 
will help them apply the rules as 
mechanically as possible. 
Second, a different group of students -- 
again English monolinguals -- will translate 
the IL-output of phase I to French, by 
mechanically applying the rules of the IL- 
French lexicon (without knowing IL or 
French). Since the test is directed at 
lexical word-choice, no complete sentences 
will be required for the French output. 
The presence of an IL-version as intermediary 
between the two translation phases will 
require that the IL-output of phase I be 
converted into complete sentences before 
serving as input to phase If. 
3. e. Evaluating the French output 
As a final part of the whole procedure, BYU 
will carefully compare the French output of 
the above described DLT trial translation 
with the high-quality human translation of 
the text samples in the database. Of course, 
this comparison will concern only lexical 
elements, ignoring case endings, word order, 
etc. 
Where the output differs from the translation 
in the data base, a French-English bilingual 
will decide whether the output is an 
acceptable alternative and if not will note 
the discrepancy for further study by BSO. 
4. Conclusion 
Once the study is completed (which will be 
years since the above procedure is only the 
first major phase), we should know more about 
the types of problems to be encountered in 
lexical transfer at later stages when the 
entire machine translation system is in 
place. The results should identify general 
problems in lexical transfer that are not 
specific to the BSO project. Based on the 
findings, we may recommend that further work 
in lexical transfer be pursued by the 
computational linguistics community. 
REFERENCES 

Bourquin-Launey, M-C. (1984) Traduction 
Automatique -- Aspects Europ/eens. 
Paris: ADEC, 99, boulevard Saint-Michel 
(5th arrondissement); January 1984. 

Hays, David G. (1963) "Research Procedures in 
Machine Translation" in Natural Language 
and the Computer, edited by Paul L. 
Garvin, pp. 183-214. New York: McGraw- 
Hill; 1963. 

Hutchins, J. (1984) "Methods of Linguistic 
Analysis in Machine Translation" in the 
proceedings of the International 
Conference on Machine Translation, 
Cranfield, England, February 1984. 

Melby, Alan K. (1985) "A Text and Translation 
Data Base", a paper presented at the 
International Conference on Data Bases 
in the Humanities and Social Sciences, 
Grinnell College, 1985. (Submitted for 
publication in the proceedings.) 
