Biblical Hebrew Forum

[Yitzhak Sapir <yitzhaksapir AT gmail.com>]

On Thu, May 27, 2010 at 2:55 AM, James Christian wrote:
> 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.

James, the issue is not whether a shallow algorithm is better than a deep
algorithm.  The question is how does any algorithm compare to a linguist
who has linguistic training.  A linguist who is required to translate a new
language, can do with a grammar and a dictionary with some few thousand
entries.  We can be generous and allow him to hear sentences in the
language for an entire year.  At a sentence per minute, that's 60 x 24 x 365,
which comes to about half a million.  The output of this human linguist will
be much better than an algorithm fed with millions of sentences.

Now, it may be that with sufficient research we can find how to integrate
linguistic knowledge with such effectiveness so as to replicate these results
in the algorithm.  But the fact that we don't know how to construct a deep
algorithm that replicates the human linguist's performance does not mean a
deep method is impossible.  The human linguist is using a deep method and
does extremely well.  In fact, the machine algorithm is at this point of
technology generally be so poor that companies needing translations of
books, documentation, etc, will generally not consider such solutions at
all, and use only the human linguist!

Naturally, we may similarly not have the best shallow method -- the best
method that removes all linguistic knowledge.  But it is a basic assumption
that the more you know about the language, grammatically, and otherwise,
the better you should be able to translate it.

The Bible itself is only some 20,000 verses.  Yet you and Karl have suggested
that useful linguistic information can be extracted from the Bible, without 
any
other input.

Again, I'm not saying that we can tell everything about the underlying 
language.
Today, we still don't know what זדה of the Siloam means, exactly.  But we can
discern that the plural marker is always -m.  Because the data is varied 
enough,
the data must be taken as representative.  That is, we didn't just by accident
dig up inscriptions all over the place who just by accident
represented the -m and
not the -ym scribal school.  (Whereas differences in languages show up as
dialects, differences in orthography at the same period would show up
as different
scribal schools).  In fact, the Siloam inscription suggests that this is the
orthography used by the royal scribes of Hezekiah.  Who else do you think
copied the Biblical texts at the time if not these scribes?

So to summarize
- Your data and experience in machine translation does not help in comparison
with human linguists given the task of translating.  Human linguists with 
their
linguistic reasoning will do much better than machine translation with
a lot less
input.  This only indicates a failure of the algorithm to effectively
incorporate
linguistic knowledge.
- The inscriptional data is a small data set.  Some linguistic conclusions can
be made, whereas others cannot because the data set is too small.  For the
issue of the orthography of plurals and some aspects of orthography in 
general,
the data set is sufficiently varied and large to make conclusions.

Yitzhak Sapir