[b-hebrew] Computer Analysis

Yaakov Stein yaakov_s at rad.com
Thu Jan 10 12:21:21 EST 2008


Jim and Michael, 

Although I have never been overly involved in Computational Linguistics,

for many years I was an active researcher in Natural Language Processing
and Statistical Pattern Recognition, and so I allow myself to comment on
this subject.

Please note that I have not read Radday nor Houk's original works,
and so am commenting solely on what I have seen in third party accounts.

I strongly disagree with Jim's remark :
   If the computer cannot distinguish the first 10 chapters of Genesis
from the 
   Patriarchal narratives in the last 40 chapters of Genesis, or if the
computer 
   finds 11 different authors of Genesis, who on earth would trust the
computer?
and even more so with his conclusion that:
   What is needed is a human being who is generally familiar with the 
   well-documented secular history of the mid-14th century BCE, and who
is brave enough to 
   compare that secular history, no holds barred, to the received text
of the 
   Patriarchal narratives. 

We don't need further biased results.
What we really need is a strong "blind test".

A blind test means that the material should be given to someone
who has never heard of the documentary hypothesis,
preferably one who has never even heard of the bible.

A blind test, as routinely used in statistical pattern recognition,
must first show that the program is able to
1) consistently show high probability for common authorship
    when presented with test texts written by a single author,
    even when these texts were written years apart and
    in different styles (Shakespeare sounds like a good candidate).
2) consistently differentiate between artificially mixed test text
segments 
    of different authorship, even when these authors worked about the
same time
    and in similar styles
and all these tests must be carried out by someone who does not know
what the results should be.

Only thereafter can any credence be given to the results.

>From what I have seen quoted, Radday's software gave low probabilities
that Goethe's and Kant's works were written by a single author.
This is used to reinforce his conclusion that a "82 percent probability
of identity between the J and E definitions of Genesis make their unity 
very highly probable". 
Quite the contrary. Before we can trust the results of such software,
it has to reliably predict common authorship for works known to be
authored 
by a single person.

Although Houk does use standard statistical measures, 
he is not only aware of the theories of J and E authors,
he seems to have been interested in further tracking the 
development over time of the J sections.
I have no problem with "annealing" techniques that,
after showing major subdivisions, can further subdivide these.
But first we must see that these divisions are not necessary results
of the methodology used.

This type of bias is known to be very hard to avoid in non-blind tests.
Even an eminent mathematician such as Eli Ripps can fall into that trap
and miss the bias he introduced in the "bible codes".

>From what I have read, neither Radday nor Houk attempted any blind
testing,
and so from a technical standpoint both results are completely
meaningless.
    
Yaakov (J) Stein



More information about the b-hebrew mailing list