code
code
DPTSG
You can download the code used in our 2009 ACL short paper, "Bayesian learning of a tree substitution grammar".
maxent
I modified the code to the excellent Edinburgh maxent toolkit (which does L-BFGS) to do the following:
1.(testing) Output the probability of the requested prediction (-l). Existing options only allowed outputting the most probable prediction or the entire distribution over predictions.
2.(training) Quit when perplexity on held-out test data starts to rise (--prob). More accurately, it quits when the perplexity is greater than that of two iterations ago, in order to allow for some noise.
You can download the patch against the original source. To apply the patch, download the source and type
$ cd maxent-20061005
$ patch -p1 -i maxent-20061005-patch
Usage:
Train until likelihood on the heldout test data starts to rise:
$ maxent -bm <model file> -l --heldout <heldout file> <training file>
Output probabilities instead of predictions or distributions:
$ maxent -pm <model file> --prob <feature file> -o <output file>
ITG decoder
This is the code described in our 2008 AMTA paper. It implements an inversion transduction parser that can use ngrams or parse trees as the target side language model. Available for download sometime in the future; in the meantime feel free to email me and I will send it to you.