GroC — a Pytorch implementation of GroC, a grounded compositional output model for adaptive language modeling presented at EMNLP 2020 [1]. The model has a fully compositional output embedding layer that is optionally further grounded in information from a structured lexicon (WordNet), namely semantically related words and free-text definitions. It can be applied to both conventional language modeling as well as challenging cross-domain settings with an open vocabulary.

@inproceedings{pappas-etal-2020-grounded,
    title = "Grounded Compositional Outputs for Adaptive Language Modeling",
    author = "Pappas, Nikolaos and Mulcaire, Phoebe and Smith, Noah A.",
    booktitle = "Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)",
    publisher = "Association for Computational Linguistics",
    url = "https://www.aclweb.org/anthology/2020.emnlp-main.96",
    pages = "1252--1267",
    year = "2020"
}

Note that this repository is based on the codebase of awd-lstm, for a general purpose language model consider using that library directly [2]. Before running the code make sure you have installed the conda environment first:

conda env create -f environment.yml

To obtain the datasets for conventional language modeling please follow the instructions from [2], e.g. the datasets can be easily obtained by running this script.

Example:

echo "- Downloading Penn Treebank (PTB)"
wget --quiet --continue http://www.fit.vutbr.cz/~imikolov/rnnlm/simple-examples.tgz
tar -xzf simple-examples.tgz``
mkdir -p penn
cd penn
mv ../simple-examples/data/ptb.train.txt train.txt
mv ../simple-examples/data/ptb.test.txt test.txt
mv ../simple-examples/data/ptb.valid.txt valid.txt
cd ..
rm -rf simple-examples/

Below you can find a few example training commands for training the baseline language model or GroC with different options.

Baseline (tied)

python -W ignore  main.py --data penn --dropouti 0.4 --dropouth 0.25 --seed 28 --batch_size 20 --epoch 1000\
 --save tied --cuda --cuda_device 0

GroC (char)

python -W ignore  main.py --data penn --dropouti 0.4 --dropouth 0.25 --seed 28 --batch_size 20 --epoch 1000\
--save groc_char --char_emb --char_update_ratio 0.3 --cuda --cuda_device 0

GroC (char, rel, def)

python -W ignore  main.py --data penn --dropouti 0.4 --dropouth 0.25 --seed 28 --batch_size 20 \
--epoch 1000 --save groc_full --char_emb --rel_emb --def_emb --char_update_ratio 0.3 --cuda --cuda_device 0

GroC for adaptation (char, rel, def, deep residual net, bias estimator)

python -W ignore  main.py --data penn --dropouti 0.4 --dropouth 0.25 --seed 28 --batch_size 20 \
--epoch 1000 --save groc_full --char_emb --predict_bias --joint_emb 400 --joint_emb_depth 4 --joint_dropout 0.6\ 
--joint_locked_dropout --joint_emb_activation Sigmoid --char_update_ratio 0.3 --cuda --cuda_device 0

For those who are interested, we also make our pretrained models and configurations publicly available for our experiments on conventional language modeling and cross-domain language modeling: Pretrained GroC models (Google Drive).

Download the News Crawl data here (http://statmt.org/wmt14/translation-task.html, under “Monolingual language model training data”). Place the downloaded files in a directory called “raw” and run scripts/create-data.sh to recreate our splits.

python evaluate.py --test_data data/news2007_train.news2008_test/ --save saved-models/[our model] --cuda  
--cuda_device 0 --seed 1234 --adapt_method [change_vocab|interpolate_neural|interpolate_unigram]

• [1] Nikolaos Pappas, Phoebe Mulcaire, Noah A. Smith Grounded Compositional Outputs for Adaptive Language Modeling, Conference on Empirical Methods in Natural Language Processing, 2020
• [2] Stephen Merity, Nitish Shirish Keskar, Richard Socher, Regularizing and Optimizing LSTM Language Models, Sixth International Conference on Learning Representations, Vancouver, Canada, 2018

For questions and requests please email: npappas@cs.washington.edu, pmulcaire@cs.washington.edu.