TRF-semi.md

TRF semi-supervised learning on PTB

This is semi-supervised learning on PTB dataset. TRFs are used to model the joint probabilities of the word sequences and the POS sequences. TRFs are compared with CRFs in POS tagging task, and compared with ngram/LSTM in language modeling task.

folders

• full: supervised training on all the training data
• semi: semi-supervised training on part of training data

Scripts

full/

scrptes	description
data_process.py	(run first) prepare the training data
eval.py	including the code to compute the error rate
plot.py	plot the results
plot_thesis.py	plot the results used in thesis
run_baseline_ngrams.py	run ngram LMs
run_lsmtlm.py	run lstm LMs
run_crf_discrete.py	run CRF with discrete features
run_crf_neural.py	run neural CRF
run_crf_neural2.py	run neural CRF, a better realization than run_crf_neural.py
run_hrf_*.py	the old version of training TRFs. The new scripts are in semi/

semi/

scrptes	description
task.py	(run first) prepare the training data, i.e. selecting part of the training data including tags
eval.py	including the code to compute the error rate
plot.py	plot the results
run_lsmtlm.py	run lstm LMs
run_crf_neural.py	train neural CRFs on part data set
run_hrf_neural.py	train neural TRFs by AugSA method (very slow)
run_hrf_neural_nce.py	train neural TRFs by NCE/DNCE (fast)