[developers] Re: Transfer module in LKB
oe at csli.Stanford.EDU
Sun Mar 13 23:37:22 CET 2005
hi jostein, petter, and anette,
finally, here are some thoughts on the status of transfer in the LKB.
again, my apologies for the tardy reply (end of quarter here now :-)!
the code, actually, is part of the standard LKB distribution by now; it
resides in a separate :mt package. i am still working on it. usually,
i develop in a separate source code branch (LOGON) and every few months
propagate the transfer code back into the LinGO CVS repository. there
is absolutely no documentation, currently. frankly, i could not decide
whether to encourage or discourage users at this point. the kernel of
the transfer rewrite engine seems reasonably stable. in LOGON, we use
some 1000 transfer rules these days, expecting to greatly increase this
number until this summer. however, i plan to make a few modifications
to the formalism this spring (specifically the treatment of coreference
in transfer rules and the facilities for non-monotonic outputs). right
now, it can be difficult to predict which forms of re-entracy in types
for transfer rules will have the intended effect. another modification
expected for later this year is a move towards `packing' of MRSs, i.e.
a facility that allowed local disjunction, so as to not have to explode
the number of independent transfer paths for each lexical ambiguity.
from this point of view, it could be risky investing a lot of work on
the current formalism. also, lacking documentation (and good support)
for what is there, i imagine there could be a high start-up cost. we
have two and a half experience reports so far: dorothee, in LOGON, had
a starting transfer grammar (that i had created) and tried extending it
with additional types and rules. i suspect she would characterize her
experience as frustrating. dan is using the transfer machinery in the
ERG for the (very experimental) paraphraser and, in the next version,
some idiom machinery. again, i set him up initially, and he seemed to
then be able to work things out (down to making code fixes :-). emily,
finally, has experimented with baby transfer grammars in teaching; she
seemed to get out what she wanted, but would likely attest to start-up
in our latest grammar writing course at stanford,
we included one exercise involving transfer. that seemed to work okay
for the class, and i am currently preparing our model solution for the
MT assignment. i was also expecting to make that publicly available,
ideally even starting some documentation on the LKB wiki. however, my
concerns about formalism changes down the road remain: i would surely
not want to have to support legacy transfer grammars as the formalism
evolves. for the time being, in my view, it is a moving target.
--- i fear i ended up being more discouraging than encouraging, but at
least you should now know enough to draw your own conclusions :-).
all the best - oe
+++ Universitetet i Oslo (ILF); Boks 1102 Blindern; 0317 Oslo; (+47) 2285 7989
+++ CSLI Stanford; Ventura Hall; Stanford, CA 94305; (+1 650) 723 0515
+++ --- oe at csli.stanford.edu; oe at hf.uio.no; stephan at oepen.net ---
More information about the developers