DONE: line 142 (ml_dmv_parser.py ) batch_predict_lan_v should be manually given a value. I will revise it later.

DONE: line 308 (def evaluate_batch_score of ml_dmv_model.py) no use neural:

if self.initial_flag: ---> if True:

if set non_neural_iter>epochs, then it is DMV model.

python src/ml_dmv_parser.py --train en-fr --dev en --child_neural --em_type em --cvalency 2 --do_eval --ml_comb_type 1 --bidirectional --child_only --em_iter 1 --function_mask

---10.25 for ML-CLS2-1--- python src/ml_dmv_parser.py --train en --child_neural --em_type em --cvalency 2 --do_eval --ml_comb_type 2 --stc_model_type 1 --child_only --em_iter 1 --function_mask --non_dscrm_iter 20 --epochs 70

lv_dmv_parser

Dependency parser using DMV model with latent variables Parameters with best performance using vanilla DMV: --train data/wsj10_tr --dev data/wsj10_d --epoch 25 --split_epoch 4 --param_smoothing 0.1 --split_factor 2 --cvalency 1 --em_type viterbi --batch 1000 --sub_batch 1000 --do_eval

Parameters with best performance using split-DMV(viterbi): --train data/wsj10_tr --dev data/wsj10_d --epoch 25 --do_eval --use_lex --split_epoch 4 --param_smoothing 0.1 --split_factor 2 --cvalency 1 --do_split --em_type viterbi

Parameters with best performance using split-DMV(lateen): --train data/wsj10_tr --dev data/wsj10_d --epoch 25 --do_eval --use_lex --split_epoch 4 --param_smoothing 0.1 --split_factor 2 --cvalency 2 --do_split --em_type viterbi --em_after_split