Comments/Questions about today's discussion paper:

=========================================================================

Kwong Tim Ng: I was thinking that why they did not include the POS
information when they did word-by-word translation. For example, the
English word 'interest' can have several different meanings. If they
didn't consider the POS information, the translation for this word can
be probably wrong.

=========================================================================

From Jeremy Kahn:

The independence assumption that the node operations {n,r,t} are
entirely mutually independent across nodes, and that within each node,
n, r, and t are independent of each other is flawed but not
unreasonable. Example:

Japanese sentence (glossed to English):
  book-OBJ buy-PAST.

English translation might be a passive:
  [A] book [was] buy-PAST.

German (glossed) might be:
  Book-NOM was buy-PASSIVE.
or
  Book-OBJ [I-NOM] buy-PAST

but the following is wrong:
  Book-OBJ was buy-PASSIVE.
or
  Book-NOM [I-OBJ] buy-PAST

(note inserted words are bracketed)

Note that the translation of OBJ to NOM co-varies with the presence of
the PASSIVE morphemes.

However, the language model might be able to identify this problem -- 
maybe not with trigrams, though.

=====
The model initialization is uniform probability. Both Katrin and I
were suggesting that EM's problems with local minima might want to
start at several different points and/or start with a
linguistically-motivated start point -- at least, set up the t-table
based on using words found in the same sentences, or use one of the
IBM models to get an initial alignment.
=====
How does this model vary as different parsers are tried?  What
characteristics of the parser are required -- and what characteristics
would make this model difficult to use?  For example, they chose to
flatten much of the binary structure out of the traditional syntactic
analyses.

=========================================================================