TODO

THERE'S A PROBLEM WITH THE EM OPTIMIZATION. SOMETIKMES THE LIKELIHOOD DEGRADES. THERE MIGHT BE  A SYNCHRONIZATION PROBLEM. THE TINY DATASET WOKRSK FINE. THE 10 SENT FR DATASET BREAKS. ALLOW_NULLL COULD BE A TRIGGER. NUMBER OF EPROCESSES COULD BE A TRIGGER.

EVEN THE TINY DATASET DOES NOT STEADILY IMPROVE WITH EM, WHEN LAMBDAS != 0



Paper writing:

phase 1: convince yourself that the model performs well with at least two tasks
DUE 6/17

- cache the features and rescore when the weights change
- write a report on where we are so that i can continue after going back to school
+ make sure the script wa-aren-latentCrfAligner.sh works well (with small training data).
+ use netbeans remote dev features https://netbeans.org/kb/docs/cnd/remotedev-tutorial.html
- use ducttape
- currently, it takes about 7 iterations over the dataset in order to make one step in coordinate descent
- currently, when you initialize with features and weights from a different data set (even for the same language), the program breaks. we need to be able to do that to jumpstart large-scale experiments.
+ "labeling the first 13263 in the corpus" takes for ever. it seems to be using a single process.
- use a command line library for the options
- use a 100K training set for aren, and evaluate WER of fastalign.
- regenerate word-pair features for the largest dataset
- use a 100K training set for aren, and evaluate WER of latentCRF.
- run controlled word alignment experiments with a medium dataset. compare to hmm aligner and ibm model 4 (initialized with model 1), and fastalign.
- use the learnt word alignments to train translation models and see how they compare to models trained with giza++ or fastalign
- run controlled POS tagging experiments without a tagging dictionary. compare to berg-kirkpatrick and my hmm implementation.
- find two state-of-the-art POS tagging results and replicate the experimental setup and try to beat them.

phase 2: run large scale experiments to convince the reviewers, in coordination with Chris

phase 3: in parallel with phase 2, write a draft