This is an implementation of the paper:
Unsupervised Recurrent Neural Network Grammars
Yoon Kim, Alexander Rush, Lei Yu, Adhiguna Kuncoro, Chris Dyer, Gabor Melis
NAACL 2019
The code was tested in python 3.6 and pytorch 1.0.
Sample train/val/test data is in the data/ folder. These are the standard datasets from PTB.
First preprocess the data:
python preprocess.py --trainfile data/train.txt --valfile data/valid.txt --testfile data/test.txt
--outputfile data/ptb --vocabminfreq 1 --lowercase 0 --replace_num 0 --batchsize 16
Running this will save the following files in the data/ folder: ptb-train.pkl, ptb-val.pkl,
ptb-test.pkl, ptb.dict. Here ptb.dict is the word-idx mapping, and you can change the
output folder/name by changing the argument to outputfile. Also, the preprocessing here
will replace singletons with a single <unk> rather than with Berkeley parser's mapping rules
(see below for results using this setup).
To train the URNNG:
python train.py --train_file data/ptb-train.pkl --val_file data/ptb-val.pkl --save_path urnng.pt
--mode unsupervised --gpu 0
where save_path is where you want to save the model, and gpu 0 is for using the first GPU
in the cluster (the mapping from PyTorch GPU index to your cluster's GPU index may vary).
Training should take 2 to 3 days depending on your setup.
To train the RNNG:
python train.py --train_file data/ptb-train.pkl --val_file data/ptb-val.pkl --save_path rnng.pt
--mode supervised --train_q_epochs 18 --gpu 0
For fine-tuning:
python train.py --train_from rnng.pt --train_file data/ptb-train.pkl --val_file data/ptb-val.pkl
--save_path rnng-urnng.pt --mode unsupervised --lr 0.1 --train_q_epochs 10 --epochs 10
--min_epochs 6 --gpu 0 --kl_warmup 0
To train the LM:
python train_lm.py --train_file data/ptb-train.pkl --val_file data/ptb-val.pkl
--test_file data/ptb-test.pkl --save_path lm.pt
To evaluate perplexity with importance sampling on the test set:
python eval_ppl.py --model_file urnng.pt --test_file data/ptb-test.pkl --samples 1000
--is_temp 2 --gpu 0
The argument samples is for the number of importance weighted samples, and is_temp is for
flattening the inference network's distribution (footnote 14 in the paper).
The same evaluation code will work for RNNG.
For LM evaluation:
python train_lm.py --train_from lm.pt --test_file data/ptb-test.pkl --test 1
To evaluate F1, first we need to parse the test set:
python parse.py --model_file urnng.pt --data_file data/ptb-test.txt --out_file pred-parse.txt
--gold_out_file gold-parse.txt --gpu 0
This will output the predicted parse trees into pred-parse.txt. We also output a version
of the gold parse gold-parse.txt to be used as input for evalb, since sentences with only trivial spans are ignored by parse.py. Note that corpus/sentence F1 results printed here do not correspond to the results reported in the paper, since it does not ignore punctuation.
Finally, download/install evalb, available here.
Then run:
evalb -p COLLINS.prm gold-parse.txt test-parse.txt
where COLLINS.prm is the parameter file (provided in this repo) that tells evalb to ignore
punctuation and evaluate on unlabeled F1.
Note that some of the details regarding the preprocessing is slightly different from the original
paper. In particular, in this implementation we replace singleton words a single <unk> token
instead of using Berkeley parser's mapping rules. This results in slight lower perplexity
for all models, since the vocabulary size is smaller. Here are the perplexty numbers I get
in this setting:
- RNNLM: 89.2
- RNNG: 83.7
- URNNG: 85.1 (F1: 38.4)
- RNNG --> URNNG: 82.5
Some of our preprocessing and evaluation code is based on the following repositories:
MIT