Rico Sennrich
908c006e32
online combination of multiple phrase tables
...
- creates a virtual phrase table at decoding time based on a vector of component models and a combination algorithm
- linear interpolation or instance weighting
- two possible component model types supported so far: 0 (in-memory) or 12 (compact)
- weights can be set in config, and overriden on a sentence-level through mosesserver API
- online optimization (perplexity minimization) using dlib and xmlrpc-c call
2013-04-22 13:21:59 +02:00
Barry Haddow
9d42c7f6f7
Merge branch 'master' of github.com:moses-smt/mosesdecoder
2013-04-12 16:07:26 +01:00
Hieu Hoang
44a0e52e30
fixed ShowWeights() for confusion networks. This is a reason why we should get rid of ShortNames and move to refactored moses pdq
2013-04-09 14:44:32 +01:00
phikoehn
ac82be3120
Hal moved. We follow.
2013-04-03 21:59:03 +01:00
Ondrej Bojar
93433cf015
support --translation-details OUTFILE in moses-parallel
2013-04-03 18:10:44 +02:00
phikoehn
0a978e9f01
bug fixes
2013-04-01 14:31:32 +01:00
Barry Haddow
42526b5b6e
Merge branch 'master' of github.com:moses-smt/mosesdecoder
2013-03-18 21:50:11 +00:00
Barry Haddow
8efeb59228
don't lowercase reference if there's a recaser
2013-03-18 21:29:17 +00:00
Achim
038871fdb3
Hungarian and Latvian non-breaking prefix files
2013-03-18 17:17:35 -04:00
Hieu Hoang
1b83b85f44
debug info from sort command
2013-03-18 16:48:40 +00:00
Barry Haddow
4c2e2d768b
Update mert training to use interpolated ttable
2013-03-15 16:13:33 +00:00
phikoehn
3a7f4f776a
minor
2013-03-13 17:54:29 +00:00
Phil Williams
6fa279fadb
filter-rule-table.py: change default pruning count from 1 to 0
...
Change the default pruning threshold from 1 to 0 to allow for
Hiero-style fractional counts.
2013-03-04 21:02:50 +00:00
Christian Buck
26bf04df5d
added unbuffered mode for casers (using -b)
2013-03-04 15:29:13 +00:00
amittai
7ca271b200
fixed typo
2013-02-26 19:47:44 -08:00
amittai
1f82a43837
where'd the edit go?
2013-02-26 11:37:31 -08:00
amittai
5cdf65ba33
Revert "Revert "let's be consistently case-insensitive with respect to the xml tags""
...
This reverts commit 8b6e98c633
2013-02-26 11:32:29 -08:00
amittai
1fb51dc674
use 'gunzip -c' instead of 'zcat' for better cross-platform compatibility
...
zcat is identical to "gunzip -c", but Mac OS X doesn't ship with zcat.
2013-02-26 11:19:33 -08:00
Barry Haddow
ed117f55c9
Timing info. Command line args.
2013-02-25 09:36:58 +00:00
amittai
8b6e98c633
Revert "let's be consistently case-insensitive with respect to the xml tags"
...
This reverts commit 2eb0c5e11d
2013-02-24 18:10:19 -08:00
amittai
2eb0c5e11d
let's be consistently case-insensitive with respect to the xml tags
2013-02-24 18:07:11 -08:00
Barry Haddow
2f221473f0
Change from phrase-weighting to promix
2013-02-21 21:40:01 +00:00
Barry Haddow
51ab9aa19d
Merge remote branch 'origin/master' into phrase-weighting
2013-02-21 17:34:59 +00:00
Barry Haddow
9ca364fb22
Implement brevity penalty smoothing for PRO
...
As in Nakov et al (Coling 2012)
2013-02-18 11:11:20 +00:00
Ales Tamchyna
2b7725db34
support LM OOV feature in train-model.perl
2013-02-01 15:47:05 +01:00
Tetsuo Kiso
8b4a1fa2b8
Fix MegaM's URL.
...
Because I got a 404 error.
2013-01-20 02:30:20 +09:00
amittai
176647e342
accept either "mgiza" or "mgizapp" and either "snt2cooc.out" or "snt2cooc"
...
Fixed a mismatch between the wiki and mgiza.
Installing mgiza produces a file called "mgiza".
However, the Moses instructions on the wiki here
http://www.statmt.org/moses/?n=Moses.ExternalTools#mgiza
insist that the "mgiza" binary be renamed "mgizapp", but then
train-experiment.perl only accepts the binary called "mgiza".
2013-01-15 19:11:49 -08:00
phikoehn
124c36a837
bug fix with MML settings
2013-01-14 19:39:26 +00:00
phikoehn
a7f7379fa4
fixed bug in detruecaser / interaction with esacping
2013-01-14 19:25:43 +00:00
phikoehn
d5cf38cab2
Merge branch 'master' of git://github.com/moses-smt/mosesdecoder
2013-01-14 19:23:02 +00:00
phikoehn
344b150372
bug fixes with escaping / truecasing interactions
2013-01-14 19:22:29 +00:00
Kenneth Heafield
c9687e3b50
Fix longstanding bug in sentence splitter spacing.
...
"Foo Bar. Baz Quux." is two sentences even though there are two spaces instead of one.
2013-01-11 13:32:24 +00:00
Graham Neubig
c55a1474df
Updated experiment.meta
2013-01-10 16:16:23 +09:00
Barry Haddow
936dbf6516
Instance weighting
2013-01-08 16:40:00 +00:00
Barry Haddow
c86c11abbe
instance weighting of lex weights
2013-01-08 15:34:29 +00:00
Barry Haddow
a55a936182
remove warning
2013-01-08 14:28:16 +00:00
Barry Haddow
459acf87b1
Add support for instance weights file
2013-01-04 14:55:24 +00:00
Phil Williams
06081f7ddb
extract-target-trees.py: minor fixes, code style
2012-12-17 18:49:50 +00:00
phikoehn
b275c94dbf
allow for inclusion of extract from previous run
2012-12-12 07:02:59 +00:00
phikoehn
24e1df7520
support for use of baseline alignment model
2012-12-12 03:59:14 +00:00
phikoehn
438dcb1a34
bug fix in experiment.perl wrt. get-corpus-script
2012-12-10 23:50:14 +00:00
Barry Haddow
16ea68f55f
Fix bug in mml scoring
...
Line length calculation was out of step with LM scoring.
2012-12-10 15:54:24 +00:00
phikoehn
ed2d191821
allow specification of end point for experiment.perl
2012-12-10 05:56:51 +00:00
phikoehn
ccf9e13d8e
bug fix with multicore parallelizer
2012-12-09 22:27:02 +00:00
phikoehn
466b502ae0
minor bug fixes with MML
2012-12-09 20:31:20 +00:00
phikoehn
ab2effb6fe
train MML in-/out-of-domain language models with same vocabulary
2012-12-01 13:46:59 +00:00
Hieu Hoang
5fd9cbb529
delete reference to numpy. Doesn't need it
2012-11-30 10:28:51 +00:00
phikoehn
338b7656a6
ooops
2012-11-30 07:36:59 +00:00
phikoehn
84cb04c05a
fixes and extensions to modified Moore-Lewis filtering, now works with domain features
2012-11-30 07:28:31 +00:00
phikoehn
1f7ee0e6c5
change of settings for sigtest filtering
2012-11-29 23:44:10 +00:00
