summaryrefslogtreecommitdiff
path: root/README
diff options
context:
space:
mode:
authorredpony <redpony@ec762483-ff6d-05da-a07a-a48fb63a330f>2010-06-22 05:12:27 +0000
committerredpony <redpony@ec762483-ff6d-05da-a07a-a48fb63a330f>2010-06-22 05:12:27 +0000
commit0172721855098ca02b207231a654dffa5e4eb1c9 (patch)
tree8069c3a62e2d72bd64a2cdeee9724b2679c8a56b /README
parent37728b8be4d0b3df9da81fdda2198ff55b4b2d91 (diff)
initial checkin
git-svn-id: https://ws10smt.googlecode.com/svn/trunk@2 ec762483-ff6d-05da-a07a-a48fb63a330f
Diffstat (limited to 'README')
-rw-r--r--README57
1 files changed, 57 insertions, 0 deletions
diff --git a/README b/README
new file mode 100644
index 00000000..47b52355
--- /dev/null
+++ b/README
@@ -0,0 +1,57 @@
+cdec is a fast decoder.
+
+SPEED COMPARISON
+------------------------------------------------------------------------------
+
+Here is a comparison with a couple of other decoders doing SCFG decoding:
+
+ Decoder Lang. BLEU Run-Time Memory
+ cdec c++ 31.47 0.37 sec/sent 1.0-1.1GB
+ Joshua Java 31.55 2.34 sec/sent 4.0-4.8GB
+ Hiero Python 31.22 27.2 sec/sent 1.7-1.9GB
+
+The maximum number of pops from candidate heap at each node is k=30, no other
+pruning, 3gm LM, Chinese-English translation task.
+
+
+GETTING STARTED
+------------------------------------------------------------------------------
+
+See the BUILDING file for instructions on how to build the software. To
+explore the decoder's features, the best way to get started is to look
+at cdec's command line options or to have a look at the test cases in
+the tests/system_tests/ directory. Each of these can be run with a command
+like ./cdec -c cdec.ini -i input.txt -w weights . The files should be
+self explanatory.
+
+
+EXTRACTING A SYNCHRONOUS GRAMMAR / PHRASE TABLE
+------------------------------------------------------------------------------
+cdec does not include code for generating grammars. To build these, you will
+need to write your own software or use an existing package like Joshua, Hiero,
+or Moses.
+
+
+OPTIMIZING / TRAINING MODELS
+------------------------------------------------------------------------------
+cdec does include code for optimizing models, according to a number of
+training criteria, including training models as CRFs (with latent derivation
+variables), MERT (over hypergraphs) to opimize BLEU, TER, etc.
+
+Eventually, I will provide documentation for this.
+
+
+ALIGNMENT / SYNCHRONOUS PARSING / CONSTRAINED DECODING
+------------------------------------------------------------------------------
+cdec can be used as an aligner. For examples, see the test cases.
+
+
+COPYRIGHT AND LICENSE
+------------------------------------------------------------------------------
+Copyright (c) 2009 by Chris Dyer <redpony@gmail.com>
+
+See the file LICENSE.txt for the licensing terms that this software is
+released under. This software also includes the file m4/boost.m4 which is
+licensed under the LGPL v3, for more information refer to the comments
+in that file.
+