Translation: Human and Machine

From the earliest days of mankind
      From the earliest civilizations
If you study the records, you'll find
      That the world is in need of translations.
The translation of prose piece and poem
      Has engaged many writers of note
George Eliot, Luther, Jerome,
      Seamus Heaney, Saul Bellow, and Pope.

If you hail, let us say, from Hong Kong
      And you want to translate from Hungarian
You must master the strange Magyar tongue
      With the aid of an expert grammarian.
You must learn all the words and their forms
      Conjugation, declension, and cases
The idioms, grammar, and terms,
      How vowels transform, in which places.

That will do for a start. For the rest,
      If you want your translations to shine
A semester in fair Budapest
      Will do wonders your skill to refine
What's more, you must master the topic
      Of the writing you plan to address.
If Bartok, then learn about music
      If Polgar, then learn about chess.

And now when a text you peruse
      In Hungarian, you know what you've seen.
You can grasp why the authors would choose
      Those words to express what they mean.
And while you're translating the piece
      You must bring to bear all that you've known.
You must somehow convey in Chinese
      The content and style and tone.

If to build electronic compu-
      ters to translate the text is your aim
The approach that you need to pursue
      Is not even remotely the same.
Forget what the schoolteachers say.
      It's a pure waste of time, that's the truth.
If you fire a linguist a day,
      Your BLEU score will go through the roof!

To begin, you must manage to gather
      A collection of parallel texts.
The bigger the better. (Don't bother
      To enforce any quality checks.
Mistakes will come out in the wash.)
      Not less than a whole petabyte.
It certainly helps to be Google,
      Or to know Peter Norvig by sight.

If you follow the Rev. Thomas Bayes
      And desire a model stochastic,
A handful of PCFGs
      Yields results little short of fantastic.
You use Bayes' Law to find the posterior.
      Don't bother with normalization.
You reject all the options inferior,
      And produce the most likely translation!

Or if neural networks you'd use ---
      You're a fan of the learning that's deep ---
Get a thousand or more GPU's.
      No one said it was gonna be cheap.
Then you search in parameter space
      Using gradient descent from the start
And thus you converge to a place
      Where you're beating the state of the art!

If you translate a story or song
      Using Google or Systran or Bing
The result's often right --- sometimes wrong ---
      All too often, it don't mean a thing.
If you hire a fellow you know,
      Instead of those wondrous machines
He's erratic, expensive, and slow,
      But he might understand what it means.


This is part of the collection Verses for the Information Age by Ernest Davis