Implemented -r (-recover-input-path) to recover the actual path through the CN/word lattice that was used in the best translation. Also cleaned up handling of source Phrases in the context of a particular hypothesis, which may improve performance translating with lexicalized reordering models.

Removed PhraseReference since the concept was broken for CN/lattice inputs.



git-svn-id: https://mosesdecoder.svn.sourceforge.net/svnroot/mosesdecoder/trunk@1468 1f5c12ca-751b-0410-a591-d2e778427230

2 files changed, 9 insertions, 6 deletions

diff --git a/regression-testing/tests/lattice-distortion/moses.ini b/regression-testing/tests/lattice-distortion/moses.ini
index 0aabf01ba..1a565d4e3 100644
--- a/regression-testing/tests/lattice-distortion/moses.ini
+++ b/regression-testing/tests/lattice-distortion/moses.ini
@@ -55,6 +55,9 @@ T 0
 [verbose]
 2
 
+[recover-input-path]
+1
+
 [n-best-list]
 nbest
 10
diff --git a/regression-testing/tests/lattice-distortion/truth/results.txt b/regression-testing/tests/lattice-distortion/truth/results.txt
index 0d47a7393..e3da5a06e 100644
--- a/regression-testing/tests/lattice-distortion/truth/results.txt
+++ b/regression-testing/tests/lattice-distortion/truth/results.txt
@@ -1,9 +1,9 @@
-TRANSLATION_1=1 2 
-TRANSLATION_2=1 2 3 
-TRANSLATION_3=1 2 3 
-TRANSLATION_4=1 2 3 
-TRANSLATION_5=4 1234 5 
-TRANSLATION_6=1 
+TRANSLATION_1=A B ||| 1 2 
+TRANSLATION_2=A B C ||| 1 2 3 
+TRANSLATION_3=A B C ||| 1 2 3 
+TRANSLATION_4=A B C ||| 1 2 3 
+TRANSLATION_5=D E X ||| 4 1234 5 
+TRANSLATION_6=A ||| 1 
 LMLOAD_TIME ~ 0.000
 PTLOAD_TIME ~ 0.000
 SCORE_1 = 1.090