diff options
| author | adam.d.lopez <adam.d.lopez@ec762483-ff6d-05da-a07a-a48fb63a330f> | 2010-07-13 03:27:59 +0000 |
|---|---|---|
| committer | adam.d.lopez <adam.d.lopez@ec762483-ff6d-05da-a07a-a48fb63a330f> | 2010-07-13 03:27:59 +0000 |
| commit | bded9a46cb3a27b8049f74e9948be783ae6ec42a (patch) | |
| tree | 6d161d306ae43db7226dbefc84d5cf58dbe50c43 /gi/posterior-regularisation/prjava/src/test | |
| parent | 2530464e1c2cc7eec9445b2f9b0ae90c59265f57 (diff) | |
cdec now supports coarse-to-fine decoding (for SCFG only).
CTF has several options:
-coarse_to_fine_beam_prune=<double> (required to activate CTF)
assign an alpha parameter for pruning the coarse foreast
-ctf_beam_widen=<double> (optional, defaults to 2.0):
ratio to widen coarse pruning beam if fine parse fails.
-ctf_num_widenings=<int> (optional, defaults to 2):
number of times to widen coarse beam before defaulting to exhaustive
source parsing
-ctf_no_exhaustive (optional)
do not attempt exhaustive parse if CTF fails to find a parse.
Additionally, script extools/coarsen_grammar.pl will create a
coarse-to-fine grammar (for X?? categories *only*). cdec will
read CTF grammars in a format identical to the original, in which
refinements of a rule immediately follow the coarse projection,
preceded by an additional whitespace character.
Not fully tested, but should be backwards compatible. Also not
yet integrated into pipelines, but should work on the command line.
git-svn-id: https://ws10smt.googlecode.com/svn/trunk@231 ec762483-ff6d-05da-a07a-a48fb63a330f
Diffstat (limited to 'gi/posterior-regularisation/prjava/src/test')
0 files changed, 0 insertions, 0 deletions