Does Machine Translation
have a role in language
Harold Somers
Centre for Computational Linguistics
UMIST, Manchester
MT: state of the art
Language learners – trainee translators?
Previous suggestions for MT as CALL
Some suggestions/open questions
 Using MT as a bad model
 The place of translation in language teaching –
can MT play a part?
 Final thoughts – a bit disorganised: what do you
1. MT: the state of the art
 50+ years old: the original
application of “language engineering”
 FAHQT not achieved …
 … and no longer sought
 Stability and maturity with users who
(mostly) understand its pros and cons
 Available for ~12 major languages,
often free (WWW)
1. MT: the state of the art
 Experts stress the distinction between
translation for assimilation and translation
for dissemination
 MT adequate for former
 Text must be highly constrained for latter
 Why?
 Linguistic knowledge can be quite sophisticated
 But Bar Hillel’s “semantic barrier” (1959) still
there: real world knowledge, common sense
reasoning, “understanding”
1. MT: the state of the art
 Most MT systems translate rather
 “structure preserving” if not word-forword
 Therefore most CALL uses of MT
exploit MT’s weaknesses rather than
its strengths
2. Language learners as trainee
 To what extent is translation a legitimate
activity for language learners?
 As an exercise?
 As a vocational activity?
 We will return to this question
 Inasmuch as language learners may
become translators, they should be made
aware of translation technology in all its
 But this is not what I understand as “CALL”
3. Previous studies
 Loffler-Laurian (1983, 1985, 1987)
 rather general comments
 need for trainee translators to be aware of
technological advances
 role of post-editing and revision
 question of style and “l’adaptation du style aux
besoins spécifiques de la communication”
 MT output can be useful in reconsidering the
traditional notions of “mistake” and “error”.
Corness (1985, 1988)
 uses ALPS’s interactive MT system with
advanced learners of German
 “interactive translation” mode: user
chooses among alternative interpretations
of an ambiguous phrase
 e.g. a big computer user
 heightening awareness of varieties of
possible translation due to differences in
interpretation, or stylistic difference
MT as a bad model
 Typical view in 1980s:
“As language learning aids they are woefully
inadequate, but … might provide a teacher with
an interesting peg on which to hang a discussion
of grammar, asking the students to spot the
machine’s howlers and account for them.”
(Higgins and Johns, 1984)
 “… withholding the source text and inviting the
student to reconstruct it from the raw
translation. … This can be quite useful for
drawing attention to half-forgotten points of
grammar and usage.” (Ball, 1989)
Anderson (1995)
 bidirectional English–Hebrew MT system
 Students manually enter sentences one by one
from a corpus provided to them, note results,
and then use native-speaker intuition and/or
L2 reference works (depending on the
translation direction) to identify and correct the
 If into L1, can reinforce students’ awareness of
differences between the languages by showing
them a bad translation into their own language
“Doing it backwards” (Richmond
But if into L2, danger of showing learners examples of bad
Overcome by providing a model translation
Students asked to type in L1 sentence, note that system gets
it wrong.
Modify the L1 sentence until appropriate target text is obtained.
In order to get desired output, L1 text has to be modified to
make it more like the L2 target text!
“This is, of course, the reverse of normal student behaviour,
which so often consists of producing incorrect French that
sounds like English.”
No danger of reinforcing L2 errors, nor of introducing L1
“… by increasing the students’ awareness of the differen[c]es
between their first language and the target-language, the
backwards translation method places the emphasis on
linguistic processes and linguistic input rather than on
linguistic forms and output.”
Pre-editing (Shei 2002)
 both L1 and L2 text, either the student’s
own, or a given (native speaker’s) text
 Chinese-English
 L1 pre-editing encourages students to “reflect[…]
on their knowledge in the target language”
 Editing native quality L2 to coerce a better
translation is controversial approach. Mixed
reactions: some say it only reveals how
translating is not a good way to learn L2, or how
poor student’s L2 grammatical competence was
Evaluation (Belam 2002)
 compare alternative translations both
human and MT systems
 focus on question of wider context in
which translations are done
 question of exposing students to L2
texts of varying quality
 assumes they are competent to give
a relative judgment about L2
translation quality
Post-editing (Niño 2003)
 Work in progress
 Post-editing (revising) L2 MT output
to develop L2 writing skills
 Advanced students
Previous studies: summary
Focus on translation
MT as a bad model
Danger of exposure to bad L2
(Mainly) advanced students
Heightens awareness of contrastive
4. Some suggestions
 Using MT as a bad model
 Agree that showing ill-formed L2 may be
 Useful to link “bad model” activity with
understanding of how MT works
 Which in turn can focus attention on how
languages differ
Example (1)
On a donné le livre à Paul.
On a dormi dans ce lit.
One gave the book to Paul.
One slept in this bed.
Paul was given the book.
The bed was slept in.
Example (2)
Mon cousin est beau. Ma cousine
est belle. Ma cousine est riche.
My cousin is beautiful. My cousin is
beautiful. My cousin is rich.
My cousin is handsome. My cousin is
beautiful. My cousin is a rich
Translation in language teaching
 Classical “Grammar–Translation”
method – much derided
 But mainly for
 Dullness of “grammar” bit
 Choice of texts
 Note also, original model involves
only translation L2-L1
 aim is to ensure comprehension and,
perhaps, to improve L1 writing skills
More reasons not to teach
 It is independent of the four skills
which define language competence
(reading, writing, speaking, listening)
 It is radically different from these
 It takes up valuable time which could
be better used teaching these
 It is unnatural
More reasons not to teach
translation (cont.)
 It misleads students into believing in
1:1 correspondences between
 It prevents students from thinking in
the L2
 It produces interference
 It is a bad test of L2 skills
 It is only appropriate for trainee
 Many of these can be refuted
 Indeed translation persists as a
classroom activity
 Both formally as an exercise
 And informally, as a quick means to
explain things
Contemporary use of translation in
language classroom
 French: thème (into L2) vs. version (into L1)
 Different uses of translation at different stages of
 “In the elementary stages, translation from L1 to L2
may be useful as a form of control and consolidation
of basic grammar and vocabulary. […]
 “In the middle stages, translation from L2 to L1 of
words and clauses may be useful in dealing with
errors; therefore interference, interlanguage or
unconscious translationese can be illuminated by
back-translation. […]
 “In the advanced … stage of language teaching,
translation form L1 to L2 and L2 to L1 is recognised
as a fifth skill…” (Newmark 1991).
Translation into L2
Generally not seen as something translators do
Useful as a measure of L2 acquisition
It’s what people think linguists do
Exercises can give some insight into interlanguage
 Little or no attention paid to concept of interlanguage
in TS community, though translators often produce a
style which is neither L1 nor L2 – Campbell 1992
 Easier to control vocabulary and structures to be
 Easier to assess, thanks to model answers (albeit
 Are these pedagogically sound reasons?
How can MT help?
 Keep coming back to “MT as bad
model” to reinforce awareness of
 Alternative to full MT: some CAT tools
 Translation memory: can provide some
nice exercises (alternative to gap-filling)
 Interactive translation: pinpoints areas of
Translation memory
Translation into L1
 Seems a natural thing for learners to
 But it develops L1 skills, not L2 skills
 Brings us back to recurrent themes:
 Grammar-Translation model
 language learner as trainee translator
Regardless of direction …
 Translation in either direction an
important element in lexical
 Highlighting general differences, e.g.
motion verbs in E and F
 Focussing on shared cognates and false
(Anderman 1998)
How to conclude?
“It would seem that very many lovers of
languages love to translate, it is a
very motivating activity, more so
perhaps than some other language
learning activities conducted
exclusively in the target language.
This feature is perhaps something
teachers can capitalize on.” (Sewell
How to conclude?
Anecdotal evidence is that using MT is
an enjoyable exercise which “makes a
change” for some students. The
“strange and often humorous” L1
constructions produced by the
students help to fix the correct L2
constructions in their minds.
(Richmond 1994)
How to conclude?
“It would, of course, be foolish to claim that a
study of MT should be part of the standard
repertoire of language-learning activities.
However, many students expressed the
view that they have increased their
cognitive knowledge of [L2] grammar
through having to enter information in the
system’s dictionaries; for those students
whose command of formal grammar is
weak, the MT dictionaries appear to provide
a stimulus for researching areas of basic
grammatical structure ....” (Lewis 1997)
Tentative (surprise?) conclusion
 MT may be a nice toy – a novelty –
but it’s not designed as a languageteaching tool, so you shouldn’t use it
as one:
 If you want to haul hay, get a tractor

Does Machine Translation have a role in language learning?