From 6912768e34f12b615355d32b6976a56dca58a398 Mon Sep 17 00:00:00 2001 From: graehl Date: Wed, 28 Jul 2010 06:58:43 +0000 Subject: (probably not for talks) git-svn-id: https://ws10smt.googlecode.com/svn/trunk@445 ec762483-ff6d-05da-a07a-a48fb63a330f --- report/promise_power.txt | 25 +++++++++++++++++++++++++ 1 file changed, 25 insertions(+) create mode 100644 report/promise_power.txt (limited to 'report') diff --git a/report/promise_power.txt b/report/promise_power.txt new file mode 100644 index 00000000..8e243039 --- /dev/null +++ b/report/promise_power.txt @@ -0,0 +1,25 @@ +cdec option "promise_power" + +po::value()->default_value(0) + +Give more beam budget to more promising previous-pass nodes when pruning - but +allocate the same average beams. 0 means off, 1 means beam proportional to +inside*outside prob, n means nth power (affects just --cubepruning_pop_limit). +note: for the same poplimit, this gives more search error unless very close to 0 +(recommend disabled; even 0.01 is slightly worse than 0) which is a bad sign and +suggests this isn't doing a good job; further it's slightly slower to LM cube +rescore with 0.01 compared to 0, as well as giving (very insignificantly) lower +BLEU. TODO: test under more conditions, or try idea with different formula, or +prob. cube beams. + + +urdu mt09 2pass 1st pass tuned w/ unigram, prelm_beam_prune=1 scale_prune_srclen + +promise_power: BLEU rescoring time +0: 20.21 1.8s +0.01: 20.16 1.9s +0.1: 20.11 2.0s +(insignificant differences) + +10: 19.7 1.1s +(faster but much worse bleu) -- cgit v1.2.3