[logon] end-to-end reranking

[Francis Bond]

G'day,

Thanks for the swift reply.

>> If we train an end-to-end reranking model (as Eric has: jaen.r.mem),
>> where do we actually read it?  I couldn't seem to find any such
>> setting for noen.  Or do we just set the individual parameters
>> somewhere?
>
> the use of a (proper) MaxEnt re-ranker is completely separate from the
> original, naive re-ranker.  for NoEn, the MaxEnt model is activated in
> `logon.lisp' (aka `no2en.lisp'):
>
>  ;;
>  ;; as of the HandOn release, the re-ranker code is fully integrated, so load
>  ;; the MaxEnt re-ranking model and set up feature selection appropriately.
>  ;;
>  (funcall (symbol-function (find-symbol "READ-MODEL" :tsdb))
>           (format nil "~a/lingo/lkb/src/mt/jh.r.mem" logon))
>  (set (intern "*FEATURE-FLAGS*" :tsdb)
>       '((0 :ascore) (1 :tscore) (2 :rscore)
>         (3 :lm)
>         (5 :lfn) (6 :lnf)
>         (10 :distortion)
>         (11 :distance)
>         (12 :nmtrs) (13 :tratio)
>         (14 :ratio)
>         (15 :nfragments)))
>  (set (intern "*FEATURE-GRANDPARENTING*" :tsdb) -1)
>  (set (intern "*FEATURE-NGRAM-SIZE*" :tsdb) 0)
>  (set (intern "*FEATURE-LM-P*" :tsdb) nil)
>
> as soon as the global %model% is set (the effect of read-model()), the
> code will compute the re-ranking score using the MaxEnt model instead
> of the clunky *mt-analysis-weight* et al.  this setting goes into the
> :translate Lisp process, the parent process to analysis, transfer, and
> generation.
>
> just now, i think the model files should probably go into the transfer
> grammar directory, i.e. `uio/noen/jh.r.mem' in our case.  once i copy
> the HandOn release into the SVN trunk (any day now), i will apply that
> change to NoEn.
>
> --- out of curiosity, are you getting tangible improvements out of the
> MaxEnt re-ranker?  for NoEn, several of us feel that it clearly helped
> with BLEU, but in terms of subjective quality we were a bit uncertain.

We haven't actually been able to turn it on yet, so can't say :-).

I will let you know when we have some results.

-- 
Francis Bond <http://www2.nict.go.jp/x/x161/en/member/bond/>
NICT Language Infrastructure Group