diff options
| author | Patrick Simianer <p@simianer.de> | 2013-11-13 18:28:42 +0100 |
|---|---|---|
| committer | Patrick Simianer <p@simianer.de> | 2013-11-13 18:28:42 +0100 |
| commit | 4c7d24c9357f500839f04c7c8a8cfa0472801e18 (patch) | |
| tree | 19fceda02d09c9d4990aba692ed97193020d1e86 /training | |
| parent | a5909686af9fef41ca05abde519a06c5b5382225 (diff) | |
README
Diffstat (limited to 'training')
| -rw-r--r-- | training/dtrain/README.md | 30 |
1 files changed, 12 insertions, 18 deletions
diff --git a/training/dtrain/README.md b/training/dtrain/README.md index 2bae6b48..aa1ab3e7 100644 --- a/training/dtrain/README.md +++ b/training/dtrain/README.md @@ -1,10 +1,15 @@ This is a simple (and parallelizable) tuning method for cdec -which is able to train the weights of very many (sparse) features. -It was used here: - "Joint Feature Selection in Distributed Stochastic - Learning for Large-Scale Discriminative Training in - SMT" -(Simianer, Riezler, Dyer; ACL 2012) +which is able to train the weights of very many (sparse) features +on the training set. + +It was used in these papers: +> "Joint Feature Selection in Distributed Stochastic +> Learning for Large-Scale Discriminative Training in +> SMT" (Simianer, Riezler, Dyer; ACL 2012) +> +> "Multi-Task Learning for Improved Discriminative +> Training in SMT" (Simianer, Riezler; WMT 2013) +> Building @@ -17,20 +22,9 @@ To build only parts needed for dtrain do cd training/dtrain/; make ``` -Ideas ------ - * get approx_bleu to work? - * implement minibatches (Minibatch and Parallelization for Online Large Margin Structured Learning) - * learning rate 1/T? - * use an oracle? mira-like (model vs. BLEU), feature repr. of reference!? - * implement lc_bleu properly - * merge kbest lists of previous epochs (as MERT does) - * ``walk entire regularization path'' - * rerank after each update? - Running ------- -See directories under test/ . +See directories under examples/ . Legal ----- |