Biblical Hebrew Forum

[James Christian <jc.bhebrew AT googlemail.com>]

Hi Yitzhak,

this is a partial reply to your email. I haven't read it all yet because you
have already made some basic misconceptions which I will clarify for you.

You are right that SMT makes use of no linguistic knowledge. In
computational linguistics (CL) speak we call these methods shallow methods
as opposed to deep methods which attempt to form formal definitions of
grammatical constructions. Read up on LFG and HPSG for attempts at deep
methods. Currently the shallow methods are beating the deep methods not only
in machine translation but in just about every basic CL technology.

Google search is based on a shallow search method. A bag of words. Their
search algorithm is also influenced by page rank. Using measures of
precision and recall and the combined f-measure these shallow search methods
work better than any deep attempts at search. Microsoft's Bing search is
based on powerset, an attempt at deep search methods, and is heavily based
on XLE which is based on LFG research. Even so, their search results are
probably largely shallow method driven for them to stand a chance at
competing with google.

Automatic sentence alignment of text works best a la Gale and Church by
counting the number of letters in sentences and making probabilistic gambits
of the likelihood of sentences aligning 1:1, 2:1 or 1:2. This is pure
shallow methods. It works even better than counting words which are
translation of each other to work out which sentences align to each other.
You can get about 98% accuracy using the Gale and Church method.

Part of speech tagging works best using hidden markov models. Again no
linguistic knowledge used. Just pure shallow methods.

Google translate is based on the paper you read. It is largely trained using
Europarl data. Google regularly enter SMT competitions and come in about
second place to Koehn et al who wrote the Europarl paper. They guys are who
I will be competing with in my MT based PhD when I launch my pioneering MT
system that will beat the pants out of Google translate by a long shot.

Your intimation that lack of linguistic knowledge holds back the system. In
fact SMT research is largely characterised by any attempt to introduce
explicit linguistic rules reduces the performance of the system. It is
exactly the machine learned nature of the rule system that gives the systems
such good performance. In fact the *only* attempt I know of that involves
adding linguistic knowledge to phrase based SMT systems that actually
improves performance is the synchronous CFG work of Chiang:

http://www.isi.edu/~chiang/papers/chiang-acl05.pdf

But this only increases performance for the very reason that no hand crafted
assumptions about so called linguistic knowledge are implemented in the
system. The SCFG's are learned from the data. Again, the more data the
merrier the rule system that is learned.

Chiang is going in the right direction and his ideas are similar to my own
but not quite there. Another good approach which is similar in principle to
my own but not quite there is U-DOT based on data oriented processing.
Again, the improvement in performacne of the U-DOT system over DOT systems
was letting the data speak for itself without imposing so called linguistic
knowledge on the data.

I'll answer the rest of your email later. I just had to clear that up for
you.

Key points:

*shallow methods work better than deep methods presently in ever CL task
*Google is not doing any better than Koehn et al using Europarl data
*letting the data speak for itself with automated learning algorithms works
best
*the more data the better the patterns learned

James Christian