Layer Augmentation Save

Implementation of the NLI model in our ACL 2019 paper: Augmenting Neural Networks with First-order Logic

@inproceedings{li2019augmenting,
      author    = {Li, Tao and Srikumar, Vivek},
      title     = {Augmenting Neural Networks with First-order Logic},
      booktitle = {Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics},
      year      = {2019}
  }

For the QA model, check out here.

Prerequisites

pytorch 0.4.1
numpy
h5py
spacy 2.0.11 (with en model)
glove.840B.300d.txt (under ./data/)

Besides above, make sure snli 1.0 data is unpacked to ./data/snli_1.0/, e.g. ./data/snli_1.0/snli_1.0_dev.txt.

Also unzip the file ./data/snli_1.0/conceptnet_rel.zip and put all files directly under path ./data/snli_1.0/.

0. Preprocessing

preprocess with cached tokenizations (recommended)

For reproduction, it is recommended that you can unzip the snli_extracted.zip file into ./data/snli_1.0/ directory.

Then run the batching script:

python3 preprocess.py --glove ./data/glove.840B.300d.txt --dir ./data/snli_1.0/ --batch_size 48
python3 get_pretrain_vecs.py --glove ./data/glove.840B.300d.txt --dict ./data/snli_1.0/snli.word.dict --output ./data/snli_1.0/glove
python3 get_char_idx.py --dict ./data/snli_1.0/snli.allword.dict --token_l 16 --freq 5 --output ./data/snli_1.0/char

preprocess on your own (requires ConceptNet installation)

Alternatively you can run the tokenization:

python3 snli_extract.py --data ./data/snli_1.0/snli_1.0_dev.txt --output ./data/snli_1.0/dev
python3 snli_extract.py --data ./data/snli_1.0/snli_1.0_train.txt --output ./data/snli_1.0/train
python3 snli_extract.py --data ./data/snli_1.0/snli_1.0_test.txt --output ./data/snli_1.0/test

which will produce files with the same name as those in the snli_extracted.zip. But due to the evolved SpaCy tokenizer, the tokens no longer align with the cached conceptnet_rel.zip files, so you will have to extract ConceptNet relations on your own (see the below ConceptNet section).

1. Training

mkdir ./models

python3 -u train.py --gpuid [GPUID] --dir ./data/snli_1.0/ --train_data snli-train.hdf5 --val_data snli-val.hdf5 --word_vecs glove.hdf5 \
--encoder rnn --rnn_type lstm  --attention local --classifier local --dropout 0.2 --epochs 100 --learning_rate 0.0001 --clip 5 \
--save_file models/lstm_clip5_adam_lr00001 | tee models/lstm_clip5_adam_lr00001.txt

Expect to see dev accuracy around 87.

2. Evaluation

First redo evaluation on the dev set to make sure we can get exactly the same F1 as reported during training:

python3 -u eval.py --gpuid [GPUID] --dir ./data/snli_1.0/ --data snli-test.hdf5 --word_vecs glove.hdf5 \
--encoder rnn --rnn_type lstm --attention local --classifier local --dropout 0.0 \
--load_file ./models/lstm_clip5_adam_lr00001 | tee models/lstm_clip5_adam_lr00001.evallog.txt

Expect to see test accuracy to be around 87.

3. Augmented Models

To train augmented models using the constraints N1, N2, and N3 in our paper, simply run:

GPUID=[GPUID]
CONSTR_W=n2
RHO_W=2
CONSTR_C=n3
RHO_C=1
RATIO=1
PERC=$(python -c "print(int($RATIO*100))")
SEED=1
python3 -u train.py --gpuid $GPUID --dir ./data/snli_1.0/ --train_res train.content_word.json,train.all_rel.json \
--val_res dev.content_word.json,dev.all_rel.json \
--within_constr ${CONSTR_W} --rho_w ${RHO_W} --cross_constr ${CONSTR_C} --rho_c ${RHO_C} --constr_on 1,2,3 \
--encoder rnn --rnn_type lstm --dropout 0.2 --epochs 100 --learning_rate 0.0001 --clip 5 \
--percent ${RATIO} --seed ${SEED} \
--save_file models/${CONSTR_W//,}_rho${RHO_W}_${CONSTR_C//,}_rho${RHO_C//.}_bilstm_lr00001_perc${PERC}_seed${SEED} | tee models/${CONSTR_W//,}_rho${RHO_W}_${CONSTR_C//,}_rho${RHO_C//.}_bilstm_lr00001_perc${PERC}_seed${SEED}.txt

For evaluation, remeber to change corresponding parameters in the eval.py. Expect to see accuracies as reported in our paper.

ConceptNet

Before proceeding, please make sure you have a local instance of ConceptNet running. An example of setup can be found here

For extracting edges from ConceptNet, you can refer to the following code:

DATASET=train
python3 -u conceptnet.py --sent1_lemma ./data/snli_1.0/${DATASET}.sent1_lemma.txt --sent2_lemma ./data/snli_1.0/${DATASET}.sent2_lemma.txt --worker 4 --rel syn --output ./data/snli_1.0/conceptnet.syn.txt --continu 1
python3 -u conceptnet.py --sent1_lemma ./data/snli_1.0/${DATASET}.sent1_lemma.txt --sent2_lemma ./data/snli_1.0/${DATASET}.sent2_lemma.txt --worker 4 --rel distinct --output ./data/snli_1.0/conceptnet.distinct.txt --continu 1
python3 -u conceptnet.py --sent1_lemma ./data/snli_1.0/${DATASET}.sent1_lemma.txt --sent2_lemma ./data/snli_1.0/${DATASET}.sent2_lemma.txt --worker 4 --rel related --output ./data/snli_1.0/conceptnet.related.txt --continu 1
python3 -u conceptnet.py --sent1_lemma ./data/snli_1.0/${DATASET}.sent1_lemma.txt --sent2_lemma ./data/snli_1.0/${DATASET}.sent2_lemma.txt --worker 4 --rel isa --output ./data/snli_1.0/conceptnet.isa.txt --continu 1

python3 constraint_preprocess.py --dir ./data/snli_1.0/ --src dev.sent1_lemma.txt --targ dev.sent2_lemma.txt --output_rel all_rel --output dev
python3 constraint_preprocess.py --dir ./data/snli_1.0/ --src train.sent1_lemma.txt --targ train.sent2_lemma.txt --output_rel all_rel --output train
python3 constraint_preprocess.py --dir ./data/snli_1.0/ --src test.sent1_lemma.txt --targ test.sent2_lemma.txt --output_rel all_rel --output test

python3 constraint_preprocess.py --dir ./data/snli_1.0/ --src dev.sent1_lemma.txt --targ dev.sent2_lemma.txt --src_pos dev.sent1_pos.txt --targ_pos dev.sent2_pos.txt --output_rel content_word --output dev
python3 constraint_preprocess.py --dir ./data/snli_1.0/ --src train.sent1_lemma.txt --targ train.sent2_lemma.txt --src_pos train.sent1_pos.txt --targ_pos train.sent2_pos.txt --output_rel content_word --output train
python3 constraint_preprocess.py --dir ./data/snli_1.0/ --src test.sent1_lemma.txt --targ test.sent2_lemma.txt --src_pos test.sent1_pos.txt --targ_pos test.sent2_pos.txt --output_rel content_word --output test

python3 constraint_preprocess.py --dir ./data/snli_1.0/ --src dev.sent1_lemma.txt --targ dev.sent2_lemma.txt --output_rel excl_ant --output dev
python3 constraint_preprocess.py --dir ./data/snli_1.0/ --src train.sent1_lemma.txt --targ train.sent2_lemma.txt --output_rel excl_ant --output train
python3 constraint_preprocess.py --dir ./data/snli_1.0/ --src test.sent1_lemma.txt --targ test.sent2_lemma.txt --output_rel excl_ant --output test

Issues & To-dos

• Add the machine comprehension model here.
• Add the text chunking model.