From 76272dba984264776ca4d68d3a50a032c48aa606 Mon Sep 17 00:00:00 2001 From: graehl Date: Wed, 28 Jul 2010 08:26:10 +0000 Subject: oracle prune graph data under 3 1st-pass tuning conditions git-svn-id: https://ws10smt.googlecode.com/svn/trunk@449 ec762483-ff6d-05da-a07a-a48fb63a330f --- report/prune_results/README | 21 +++++++++++++++++++++ 1 file changed, 21 insertions(+) create mode 100644 report/prune_results/README (limited to 'report/prune_results/README') diff --git a/report/prune_results/README b/report/prune_results/README new file mode 100644 index 00000000..d448ce10 --- /dev/null +++ b/report/prune_results/README @@ -0,0 +1,21 @@ +tuned on urdu baselines (hiero) devtest, test on mt09 + +hdt: single pass (no pruning of -LM forest). these weights were used for the +final +LM rescoring in all cases. (tune BLEU 22.2) + +2pass-hdt: 1st pass weights same as final weights (bad) (tune BLEU ???) + +2pass-hdt0: 1st pass weights tuned without any lm (same final weights as hdt). (tune BLEU 14.7) + +2pass-hdt1: 1st pass weights tuned with unigram lm (same final). (tune BLEU 16.2) + +data files / graphs: 2nd column is BLEU + +2pass-hdt - first column is beam alpha scaled per source word. + +times.2pass-hdt - first column is +LM rescoring time + +space.2pass-hdt - first column is portion of -LM forest edges kept + +It would be nice to graph these with the same scale, or preferably as 3 lines in +the same graph. -- cgit v1.2.3