|
|
|
|
|
|
|
Home > Archive > Article: Hans Verghese Mathews |
|
|
Borges and Machine Translation:
'Blue Tigers'
Translation by Hans Verghese Mathews |
|
|
|
|
|
|
|
|
|
|
|
This is a translation of a story ‘Blue Tigers’ by Jorge Luis Borges. As hitherto done with previous translations in Phalanx the actual translation is provided through a link to a PDF. But before the actual story is made available this exercise begins with a note from the translator reflecting on the experience of translating Borges online using the different tools available although the actual translation is done in the normal old-fashioned way.
Apology
A story by Borges in an issue given over to AI seems out of place ―― more so
that it should be Blue Tigers, Tigres Azules, one of the very last and more strange
among the ficciones ―― and that the editors should have indulged themselves so
will seem only laziness, moreover, to aficionados; who, even if they think the
phrase ‘artificial intelligence’ an oxymoron, could point to more pertinent writing
in the oeuvre; to Borges’s divagations on the characteristica universalis projected by
Leibniz, for instance, or to his singular Defense of the Kabbalah even.
Some excuse should be offered; or attempted at least; and an opening of
sorts ― a chance chink, not a “window of opportunity” ― is provided by the ending
of Tigres Azules: the englishing of its last sentence presents a peculiar obstacle, as
it happens, to machine-translation out of Spanish into English. Defending such
a claim would not be easy ―― to advance a comprehensive argument in any
economical way is beyond me ―― and I shall only try to suggest why.
So that there will not be a riot of quote-marks I shall use the courier font from
now on, rather than single-quotes, to refer to words or phrases within sentences:
to “mention” rather than “use” them, as philosophers of language used to say. But
I shall keep on using double-quotes in the usual ways: either to quote verbatim or,
else, to hedge assertion.
No oí los pasos del mendigo ciego ni lo vi perderse en el alba: our ficcion closes just
so, upon a sentence set off by itself. Our translator has hazarded what an algorithm
is likely to discard ―― I did not hear the steps of the blind beggar, nor see him lose himself
in the dawn ―― because perderse is not functioning as a reflexive pronominal
verb in the sentence: which a transcribing algorithm would segment and parse
first, assigning “parts-of-speech” to words and phrases, before proceeding. I apologise
for the terminology of the grammarians; but resorting so is the shortest way to go,
and, besides, the Web will tell curious readers enough about pronominal verbs in
Spanish, and in French; and in Latin too, should they be curious enough.
The empirical man will ask if one could ever see someone lose himself any-
where: but he will have better things to do, doubtless, than read Borges. There are
a good many sites offering automatic translation on the Web, and I thought it prudent
to test my guess by passing our sentence to their algorithms. I tried seven sites; and (with
the names of the sites appended) here are the English sentences that were returned:
A No oí | los pasos | del mendigo ciego | ni lo vi perderse | en el alba
A1 I did not hear | the footsteps | of the blind beggar | nor did I see him get lost | in the dawn
Google Translate
A2 I did not hear | the footsteps | of the blind beggar | nor did I see him lose himself | in the dawn
DeepL
A3 I did not hear | the footsteps | of the blind beggar | nor did I see him get lost | at dawn
Lingvanex
A4 I did not hear | the steps | of the blind beggar | nor did I see him get lost | at dawn
Systran
A5 I did not hear | the footsteps | of the blind beggar | or see him lose himself | in the dawn
Translate.com
A6 I did not hear | the footsteps | of the blind beggar | nor did I see him lost | at dawn
Reverso
A7 I didn't hear | the blind beggar's footsteps | or see him get lost | at dawn
PROMT
I have sectioned A in the way that all but two of the seven transcribing sentences
seem to have segmented the original: A2 and A5 are the evident exceptions.
The algorithms of DeepL and Translate.com were not blind-sided by their
parsers, it looks, in the way the others seem to have been; for while get lost may
be a competent transcription of perderse by itself, paying no heed to context,
and though at dawn for en el alba may be passable elsewhere, the rendering
get lost at dawn dissociates alba from perderse distortingly here: in A
the losing does not happen at dawn merely.
Translate.com employs “machine translation technology” and “artificial
intelligence” the site says ―― distinguishing Iberian from Latin American Spanish
and Standard American from English, reassuringly, in the provided options ―― and
that or rather than the literal nor was put in for ni suggests that its algorithm
is more than ordinarily alert to context and usage (if without conspicuous success
just now); and it is tempting to say the same of the DeepL routine.
The output of PROMT was tweaked to sound conspicuously colloquial;
and, as one might expect, the interface provides a “contexts” button activating an
algorithm that, given a string of words, will retrieve actual samples of putatively
cognate usage from its database. For perderse en it supplied the following
sentence together with a translation:
B En este mundo, es fácil perderse en los engaños y las tentaciones.
B1 In this world, it’s easy to lose oneself in all the deceptions and temptations.
The all interpolated in B1 is surprising (one could excise all the there without
loss) but, all the same, the transcribing routine seems to have sectioned B in a natural
way: linking en to perderse as much as to los engaños y las tentaciones
it looks. One wonders why the PROMT algorithm did not do likewise in transcribing
A: as the routines of both DeepL and Translate.com will now appear to have.
Noticing that the comma in B did no work, really, I thought to experiment a
little more by sending
C en este mundo es fácil perderse en los engaños y las tentaciones
to DeepL: expecting that its routine would have en qualify perderse once more,
as I supposed it to have done in transcribing A with A2. But what it returned was
the less than natural.
C1 in this world it is easy to get lost in deception and temptation.
alas: which indicated, to the contrary, that en had been assimilated into the
noun-phrase los engaños y las tentaciones without reference to the verb
perderse.
The next step was to try Translate.com with C in the hope that it
would do better than DeepL: but
C2 in this world it is easy to get lost in deceptions and temptations.
was the disappointing result. A conscientious man would have repeated the exercise
with the other sites used to translate A, but I did not have patience enough for that
(not just then. But a few days after I did try both DeepL and Translate.com with the
final two sentences of Tigres Azules, to see what difference that might make to their
transcriptions of the last one: there wasn’t any.)
So, leaving things just so, I shall try now to characterise the “peculiar obstacle” that,
as I had asserted, machine-translation from Spanish to English must overcome; and
the complication, very summarily, is that Spanish has much richer ways than English
to lexically register the modulation of agency in daily doing and common experience.
Here are three suggestive examples of verb and cognate reflexive pronominal:
caer caerse
to fall to drop
negar negarse
to deny to refuse
salir salirse
to leave to escape
To pursue the matter one would have to consider generally the use of such and like
“deponent” verbs in Spanish ― inherited from Latin, and preserved there more than
in French maybe ― the complex semantics and pragmatics of which have been studied
quite extensively (so a quick look online suggested) with a view to characterising how
their morphology indexes agency medial to and varying between active doing and
passive enduring. I shall venture, thereupon, that learning the use of deponent verbs
requires a body: somatic proprioception would condition success ―― “inner sensing”
must weigh syntactic proximity properly ―― and the discriminations of inner bodily
affect are precisely what a successful transcribing algorithm would have had to detect,
somehow, sans body, in the disjuncta membra of the corpora it was trained upon.
Readers acquainted with linguistics would have pronounced opinions on
the possibilities here. The partisans of Lakoff, for instance, are likely to agree that
mastering deponents requires a body ―― and would deny, maybe, that an algorithm
could sufficiently detect how proprioception conditions utterance (in poetry parti-
cularly) ―― but disciples of Chomsky might be more circumspect: even though the
master has not been very complimentary about “natural language processing” by
computer, as far as I know (indifferent to literature though he appears to be), and
he has scoffed at the “generative AI” that has birthed the seeming wonder risibly
named “ChatGPT”.
To say anything more without probing the processing algorithms themselves would
be very foolish however ... and an enthusiast for AI would say that my tinkering has
proved nothing at all ...
so I must leave the reader to Borges and his blue tigers
now:
venturing this much more only, and improperly maybe, that the man who
contrived those estranging fictions was a lyric poet fending, with words, within
words, the perditions ― las perdiciónes ― of old age.
Hans Verghese Mathews
Complete Story: Blue Tigers 
courtesy: https:Jorge_Luis_Borges_bibliography
|
|
|
|
|
|
|
|
|
|
|
|
Top |
|
|
|
|
|
