Code for the paper:
Compound Probabilistic Context-Free Grammars for Grammar Induction
Yoon Kim, Chris Dyer, Alexander Rush
ACL 2019

The preprocessed datasets, trained models, and the datasets parsed with the trained models can be found here.

The code was tested in python 3.6 and pytorch 1.0. We also require the nltk package if creating the processed data from the raw PTB dataset.

The processed version of PTB can be downloaded at the above link. This contains the train/validation/test sets, as well as the vocabulary used (ptb.dict). If you want to create this from scratch, you can run

python process_ptb.py --ptb_path PATH-TO-PTB/parsed/mrg/wsj --output_path DATA-FOLDER

where PATH-TO-PTB is the location of your PTB corpus and OUTPUT-FOLDER is where the processed trees are saved. This will create ptb-train.txt, ptb-valid.txt, ptb-test.txt in DATA-FOLDER.

Now run the preprocessing script

python preprocess.py --trainfile data/ptb-train.txt --valfile data/ptb-valid.txt 
--testfile data/ptb-test.txt --outputfile data/ptb --vocabsize 10000 --lowercase 1 --replace_num 1

See preprocess.py for more options (e.g. batch size). 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.

To train the compound PCFG, run

python train.py --train_file data/ptb-train.pkl --val_file data/ptb-val.pkl 
--save_path compound-pcfg.pt --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).

To train the neural PCFG that does not use continuous latent variables, run

python train.py --z_dim 0 --train_file data/ptb-train.pkl --val_file data/ptb-val.pkl 
--save_path neural-pcfg.pt --gpu 0

Training will take 2-4 days depending on your setup.

To evaluate the trained model on the test set, run

python eval.py --model_file compound-pcfg.pt --data_file data/ptb-test.txt 
--out_file pred-parse.txt --gold_out_file gold-parse.txt --gpu 0

where --out_file is where you want to output the predicted parses. This will calculate the F1 scores between the predicted trees and the gold trees in ptb-test.txt.

To parse a new set of sentences, run

python parse.py --model_file compound-pcfg.pt --data_file sents-to-be-parsed.txt 
--out_file pred-parse.txt --gpu 0

Note that sents-to-be-parsed.txt should have one sentence per line, and be preprocessed in a way that roughly matches the processing in process_ptb.py (e.g. no punctuation).

To just evaluate F1 given the trees, run (for example)

python compare_trees.py --tree1 data/parsed-data/ptb-test-gold-filtered.txt
--tree2 data/parsed-data/ptb-test-compound-pcfg.txt

To be comparable to the numbers reported in PRPN/Ordered Neurons papers, we use the original sentence F1 evaluation code based on L83-89 of the PRPN repo. This has quirky behavior in corner cases where the gold tree is over a sentence of length > 2 but only has the sentence-level trivial span. In this case the sentence F1 for that example could be potentially nonzero according to the code. Corpus F1 does not have this issue.

We provide the best neural/compound PCFG models, under the data/trained-model folder. These can be used for eval.py or parse.py.

We also provide parsed train/val/test sets for the best run of each model for further analysis and RNNG training. These can be found under the data/parsed-data folder when you download the processed datasets from above:

• ptb-{train/valid/test}-gold-filtered.txt: Gold trees where length 1 sentences have been filtered out.
• ptb-{train/valid/test}-{neural-pcfg/compound-pcfg/prpn/on}.txt: Predicted trees from the best run of different models. For example ptb-test-compound-pcfg.txt is the test set parsed with the compound PCFG.
• ptb-test-{neural-pcfg/compound-pcfg/prpn/on}-{rnng/urnng}.txt: Predicted trees for the test only for RNNG and URNNG (i.e. train on induced trees with RNNG, then fine-tune with URNNG objective). For example ptb-test-compound-pcfg-urnng.txt contains the predicted trees from an an RNNG trained on compound PCFG trees then fine-tuned with the URNNG objective.

Here are the sentence-level F1 numbers on the PTB test set for the models that performed best on the validation set . F1 with Induced URNNG indicates training an RNNG on the induced trees and then fine-tuning with the URNNG objective (see below). This gave improvements across the board.

Training the RNNG on induced trees and fine-tuning with the Unsupervised RNNG uses code from Unsupervised Recurrent Neural Network Grammars. The below commands should be run from the urnng folder.

First preprocess the training set with induced trees, for example with the compound PCFG:

python preprocess.py --batchsize 16 --vocabfile data/ptb.dict --lowercase 1 --replace_num 1
--trainfile data/parsed-data/ptb-train-compound-pcfg.txt 
--valfile data/parsed-data/ptb-valid-compound-pcfg.txt 
--testfile data/parsed-data/ptb-test-compound-pcfg.txt 
--outputfile data/ptb-comp-pcfg

Note the use of --vocabfile to use the same vocabulary as the one used in the above experiments.

Then use the above files to train an RNNG (and fine-tune with URNNG) using instructions from the URNNG folder, e.g.

python train.py --train_file /compound-pcfg/data/ptb-comp-pcfg-train.pkl 
--val_file /compound-pcfg/data/ptb-comp-pcfg-val.pkl --save_path compound-pcfg-rnng.pt 
--mode supervised --train_q_epochs 18 --count_eos_ppl 1 --gpu 0

For this version of PTB we count the </s> token in PPL calculations, hence the use of --count_eos_ppl 1. Note that this only affects evaluation and not training.

For fine-tuning:

python train.py --train_from compound-pcfg-rnng.pt --save_path compound-pcfg-urnng.pt
--train_file /compound-pcfg/data/ptb-comp-pcfg-train.pkl 
--val_file /compound-pcfg/data/ptb-comp-pcfg-val.pkl
--mode unsupervised --train_q_epochs 10 --epochs 10 --count_eos_ppl 1 --lr 0.1 --gpu 0 --kl_warmup 0

For evaluation:

python eval_ppl.py --model_file compound-pcfg-urnng.pt --samples 1000 --is_temp 2 --gpu 0
--test_file /compound-pcfg/data/ptb-test.pkl --count_eos_ppl 1

For parsing F1:

python parse.py --model_file compound-pcfg-urnng.pt --data_file /compound-pcfg/data/ptb-test.txt 
--out_file pred-parse.txt --gold_out_file gold-parse.txt --gpu 0 --lowercase 1 --replace_num 1

• LSTM LM training in Table 3 uses code/hyperparameters from here.
• For Table 3, LSTM/RNNG is trained with SGD, while PRPN/ON is trained with Adam. However, it seems like models trained with ASGD, as in the AWD-LSTM, do quite a bit better in terms of perplexity. Thanks to Freda for pointing this out!
• Curriculum learning based on length does not always seem to help, and in subsequent experiments I am seeing similar results without it. For example, one could use --max_length 45 --final_max_length 45 to train on all sentences of length up to 45 from the first epoch (limit on 45 is due to memory reasons).
• Yanpeng has a much faster implementation of compound PCFGs utilizing the pytorch-struct library. Check it out here!

Much of our preprocessing and evaluation code is based on the following repositories:

• Recurrent Neural Network Grammars
• Parsing Reading Predict Network
• Ordered Neurons

MIT