A comment thread on Mike’s “in Ephesus” blog (translating the blog name for any non-linguistic readers) that got rather off topic led me to memories of a lecture I gave on the first Tuesday in June 2002, at the UK campus of the European Training Programme of SIL and Wycliffe Bible translators of which I was at that time a member.
That particular lecture was memorable in a number of ways, standing out from the others I gave as a teacher on the translation course that year. The first was that the Tuesday in question was a special bank holiday (public holiday) here in the UK to mark the Queen’s Golden Jubilee. The Monday had also been a bank holiday and I had gone into central London with some visiting Americans to take part in the celebrations. We couldn’t afford the time to give our students both days off, so on the Tuesday I was scheduled to lecture. But I decided to give them some light relief with a less than serious look at using computers in translation.
I started by using a borrowed projector to show the class the clip about the Babel Fish from the BBC TV version of The Hitchhiker’s Guide to the Galaxy. This is now conveniently available on YouTube, but at that time was accessible only through a BBC streaming service which didn’t allow me to download or even buffer in advance. So showing this was something I dared to risk in a live lecture only because the offices were closed for the bank holiday, so I had the site’s Internet connection largely to myself. Of course now to show such a clip would be commonplace. This is how (according to my lesson plan) I followed up the video clip:
Unfortunately, the argument: “Since the Babel Fish exists, God does not” logically implies: “Since God exists, the Babel Fish does not”. That hasn’t stopped humans trying to create both of them in their own image. But human efforts to create the Babel Fish have not been very successful. (I won’t comment on human efforts to create God.)
I then went on to look at the Babelfish program, offered then by AltaVista and now by Yahoo. I was shocked at the time at what a bad job it did, and it has not improved since. In fact it gave almost identical results in 2002 and yesterday, when asked to translate Psalm 23 from German to English – a task which should not be too difficult because the languages are quite closely related:
The gentleman is not my Hirte, me anything will lack. It feasts me on a green Aue and leads me to the fresh water. It refresh my soul. It lead me on right road around its name sake. And whether I already walked, am afraid in finstern the valley I no misfortune; because you are with me, your putting and staff comfort me. You prepare a table in the face of my enemies before me. You salbest my head with oil and gives me fully. Good and mercy will follow me my life long, and I will remain in the house of the gentleman always.
This is the German text I started with, from the 1984 Luther Bible:
Der Herr ist mein Hirte, mir wird nichts mangeln. Er weidet mich auf einer grünen Aue und führet mich zum frischen Wasser. Er erquicket meine Seele. Er führet mich auf rechter Straße um seines Namens willen. Und ob ich schon wanderte im finstern Tal, fürchte ich kein Unglück; denn du bist bei mir, dein Stecken und Stab trösten mich. Du bereitest vor mir einen Tisch im Angesicht meiner Feinde. Du salbest mein Haupt mit Öl und schenkest mir voll ein. Gutes und Barmherzigkeit werden mir folgen mein Leben lang, und ich werde bleiben im Hause des Herrn immerdar.
Today I tried again with Google Translate, probably a newer program, and this did a better job, at least putting the “not” in the first verse in the right place, but oddly enough got confused by “führet”:
The Lord is my shepherd, I shall not want. He graze me on a green and Aue führet me to fresh water. He erquicket my soul. He führet me on the right road to his name’s sake. And whether I have wandered in the dark valley, I fear no accident, because thou art with me thy rod mating and comfort me. You already have a table before me in the face of my enemies. You anoint my head with oil and give me fully. Good and mercy will follow me my life, and I will remain in the house of the Lord forever.
One thing that did impress me with Google Translate was the range of languages offered. This included Greek and Hebrew. But the Greek it expects is obviously modern Greek, as it made a complete mess of a sample of the New Testament. A similar excuse would probably be offered for what it did with the biblical Hebrew of Psalm 23:
1 song Uncle Jehovah Aray not Ahsser:
2 Irvicni the veld – who Mngeota Inalane:
3 mental Ishobev Engeni Evmogali – justice for his name:
4 and that – not go Evgye shadow of death – Ira that bad – you Amdy שבטך ומשענתך Hama Engemni:
5 תערך ago | שלחן against Carary Esnat Kossy head with oil saturation:
But 6 | Good benevolence Eradpuni all – days of my life in Oshevti – Jehovah longevity:
Is this perverse mixture of translation, transliteration and simply copying the original text (minus pointing) the best we can hope for from automatic translation of the Bible? Well, I guess that an improved Google Translate fine tuned to biblical Hebrew would do rather better than this, but I doubt if it could produce a better result than Google’s translation from German to English, surely one of the tasks it is best tuned for and between related languages. Now this translation, if tidied up with proper renderings of the few words it missed, is largely accurate and comprehensible. But no one would suggest that it is anything like clear and natural enough for a published Bible translation.
The best that computers can do in Bible translation is the easy part, producing a just about comprehensible first draft. The hard work in translation is turning that draft into a text which is not just accurate but also clear and natural in the target language, and acceptable to its audience.
So, as I concluded my lecture to trainee translators:
We shouldn’t think that we can come in and provide nice computer generated translations … and expect these translations (although they may be accurate) to be clear and natural, or to be acceptable to the target audiences. … We can and should offer our technology as a service to the people we are translating for, for them to use if they want to; but we shouldn’t expect to be able to offer anything better than a useful first draft.
7 Comments
hilarious as usual! In the world of engineering and language research, the trick is experimentation and refinement – that means publishing alpha and beta versions and having time to study the limited results.
Even assuming that one day computational linguistics gets to a point where it can accurately translation, it will never be viable for translation (even for a very rough draft). To put it simply, sophisticated machine translation (even of Google’s level) requires an obnoxiously high level of understand of both the grammar and lexicon of both the source language and the target language. The amount of work and time that would be required to develop such a grammar and lexicon would be equal if not exceed the amount of time and work required to simply do the translation.
Unfortunately, Bible translators who follow word-by-word translation principles often distort a translation in the same way. I see so many examples of biblical idioms translated to English, making no sense. I see so many examples of syntactic transliteration, where, for example, Greek genitives are pretty much consistently translated with an English “of” even though the result is unnatural, and sometimes incomprehensible, English.
More English Bible translators need to take translation courses such as the one you taught, Peter. I don’t know how we can get this word out to seminaries and other places where the exegetes who work on English Bible translation teams work. But I hope that this blog might help in some small way. And I keep trying to nudge translation teams themselves, as I get to know some of their members.
I agree with Mike (though I might not say ‘never’)
.
Notice that the grammars and lexicons Mike refers to are never complete. They’re more like models of the real thing. Furthermore, the model of the one language is significantly different than the model of the other language. They work differently–it’s like modeling a rocket engine versus a Wankel engine.
What translation software would then have to do is capture a fairly complete mapping between the different models. And that mapping would only get close since the abstractions have only modeled the real thing, they are not the real thing. Concordant translation is an excellent example of how this mapping fails, since it naively assumes the mapping is fairly simple–word to word.
To further appreciate how complex this can get, the software would also have to capture the socio-linguistic features of culture and what are called frames or scripts. And this is not a complete list. But to illustrate: when I say restaurant various objects, events, etc. come to your mind. Another language might not even have the concept of restaurant. How do you map that? The software would have to map these linguistic features between the languages. Well, no one (that I know of) has even begun to think about how to map these.
Also, if one tries to abstract the models (ie, create a general purpose language processor), it seems to me, one would invariably abstract away the very details one needs from the language to make the translation clear and natural. So, to do it well would require different software written for each of the required mappings.
It’s easier to believe in the Babel Fish and not believe in God.
Thanks for the comments. It is sad that so many English translations read a bit as if they have been translated by computer, concordantly and following the source language syntax rather than in the natural forms of the target language. It is certainly much easier for a human to produce this kind of translation, with computers to help check concordance, than to do the hard work of making sure that the result is clear and natural.
As I also mentioned in that lecture and in a comment on Mike’s blog, a few years ago a linguist called Tom Pittman tried to develop a translation system based on a one-off analysis of the source text into an abstract representation of its meaning, to be followed by individualised synthesis of this abstract representation into the forms of each target language. He made bold claims about how much this would speed up the translation process for thousands of target languages. But his experiment was a failure and had to be abandoned, because from his abstract representation he couldn’t even produce an adequate rendering in English – at best first draft quality. So even if one does have “an obnoxiously high level of understand of both the grammar and lexicon of … the target language” that is still not enough to produce a useful result. Nothing beats the actual language intuition of mother tongue translators.
Of course, Peter & Mike, you both are completely right.
Peter,
Joel Nothman has some good observations on how Google deals with Hebrew with its “translation”.
http://www.joelnothman.com/blog/2008/09/28/hebrew-english-online-translation/
or http://tinyurl.com/5mgweq
One Trackback
[...] have just posted at the new Better Bibles Blog a post with this same title. Please go there to read and comment – and to see how much better a Bible blog that is now it has [...]