Don't Until the Final Verb Wait:
Reinforcement Learning For Simultaneous Machine Translation
Alvin Grissom II, He He, Jordan Boyd-Graber, John Morgan, and Hal Daumé III
Paper at EMNLP 2014
[PDF] [BibTeX]
[Video]

translation process


 
Computers h
ave been teaching themselves to translate text for some time now,
but most methods
are concerned with translations on entire sentences.
We address
the problem of simultaneous machine translation for distant
language pairs: from verb-final (SOV) to verb-medial (SVO) languages.

Simultaneous translation is the process of translating before a sentence
is complete.  When humans do this, it is called simultaneous interpretation.
Much of the prior work in this area has focused on using rule-based approaches.
We use machine learning to allow the system to teach itself how to
create better simultaneous translations.



 
Learn to Translate Incrementally

We do not program translation strategies into the system. 
Instead, we provide our imitation learning system with features.
On the basis of these features, the system decides when and if
to translate sections of a sentence.  Does it WAIT for more words
or COMMIT to a partial translation?

 
We use imitation learning to allow the system to produce better
simultaneous translations on the basis of its features.  The focus
is then shifted to better features, not programming complex strategies.
Every time the system makes a decision, it compares it to the optimal
decision that it could have made.  In this way, the system
learns from its mistakes and its successes, much in the same way
a professional simultaneous interpreter would.

Predict the Future

Some languages, such as German and Japanese, have verb-final
constructions, in which the main verb appears at the end of the sentence.
How does one translate something that hasn't yet been spoken?
One way is to guess what the speaker will say based on what the
speaker has already said.  We integrate a verb prediction component
to allow for translation from SOV to SVO languages.

Through reinforcement learning, the system learns when to
trust these predictions.


A Cumulative Metric


BLEU has been the standard metric for most languages in machine
translation.  We introduce latency BLEU (LBLEU), which takes into
account the expeditiousness of the simultaneous translation system.

LBLEU sums the BLEU scores of the incremental translations, while
weighting the final translation in proportion to the sentence size.
As a result, translations that are fast and accurate, as
opposed to merely fast or accurate, achieve higher scores.

[PDF] [Video] [BibTeX]