Biblical Hebrew Forum

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

Hi Yitzhak,

you raise many points here which we could comment further on. Firstly, you
suggest that humans are able to recognise patterns more efficiently with
less data than machine learning algorithms. You may have some good empirical
basis for this assumption but I'm not sure I entirely agree with your
conclusions. I would suggest that this means that humans have more efficient
learning algorithms and don't really see any concrete evidence that supports
your conclusion that humans are using deep linguistic methods.

Further, you claim that the act of translation as conducted by a human
translator is largely guided by some form of rule based deep linguistic
process. Proponents of data oriented processing have good psycholinguistic
and cognitive psychological basis as well as experimental computational
experience to suggest that this is not the case and that humans parse
sentences mostly by repiecing together fragments of language. Perhaps one of
the most active proponents of DOP driven linguistic processes is Rene Bod:

http://staff.science.uva.nl/~rens/

If you would like to read up more on DOP used for linguistic tasks including
machine translation see his publications here:

http://staff.science.uva.nl/~rens/publications.html

My favourite paper is perhaps:

http://staff.science.uva.nl/~rens/mtsummit2007.pdf

This paper demonstrates how unsupervised machine learning methods lead to
better performance of the system. The same point demonstrated in the Chiang
paper of SCFG I sent to you.

If you want to read the first paper that started the whole DOP thing it is
online at:

http://www.iaaa.nl/rs/LeerdamE.html

Anyway, you suggested that translators can do well with about half a million
input source language sentences. What you forget to mention is that
translators are typically native in the target language. Their input for the
target language (i.e. the target language model LM) is considerably much
higher. Further, you forget to mention that a particularly good translator
will know that a lot of translation is not the simple transfer of
grammatical constructions from language to the other. In fact, on the
contrary a lot of the translation task is selecting the best corresponding
phrase in the target language e.g. Italian 'In bocca al lupo' means 'In the
mouth of the wolf' with the basic sense of 'Good luck!'. Deep linguistic
processes are of no help here.

In any case, you raise the point of what is missing from these learning
algorithms. Why can a human do better with less data. You suggest this is
something to do with deep linguistic processes. In fact, a lot of it is to
do with cognition and not to do with any kind of linguistic processes. A lot
of it to do with knowledge basis of the world. A lot of it is to do with the
human ability to perform anaphoric resolution by connecting the meaning of
one sentence with another. In any case, we only seem to be going further
from the discussion while this is all still very interesting.

Anyway, getting back to the point I was a little disappointed by your
response. My comments on frequencies were not targeted at the -m versus -ym
phenomenon. I clearly suggested to you this could have been common practise
as there really is no ambiguity and it is much quicker to carve -m into a
rock that it is to carve -ym. We see the same phenomenon in modern text
messaging. It is much quicker to type '2' than 'to'.

My comments on frequencies were aimed at your completely unrelated
observation that some other phenomenon of alternative spellings is testified
in the inscriptional corpus with anti polaric frequency with respect to the
frequencies in the b-hebrew corpus. I pointed out to you that your data set
(the data set you rely on) is too small to be truly indicative of this
phenomenon. In fact, as you quite clearly point the data set in the b-hebrew
corpus is also too small to be indicative of this phenomenon.

In conclusion, if you are going to continue the discussion please understand
the different points are respond to them accordingly without mixing to two
unrelated issues.

James Christian