Unnatural Language Understanding (UNLU)

This repository contains the code and data used in the paper "UnNatural Language Inference", to appear in ACL 2021 (Long paper).

Demo Colab Notebook

Getting Started

For reproducing our experiments, follow the outlined steps. We use Pytorch Model Hub checkpoints (trained using FairSeq) to infer RoBERTa and BART models. For DistilBERT, we train the model using code derived from ANLI repository. For pre-Transformer models, we train the models using InferSent repository.

Install

1. Install Pytorch 1.7.0

conda install pytorch==1.7.0 torchvision==0.8.0 -c pytorch

2. Install dependencies

pip install -r requirements.txt
pip install ray[tune]

and download the Spacy English and Chinese Tokenizer:

python -m spacy download en_core_web_md
python -m spacy download zh_core_web_lg

NB: For training, please configure Apex on your system to enable fp16 training.

Set Output Path

In config.yaml, first set the proper outp_path where you will download our files.

Download the models

We provide a list of models used in the paper here. There are three model types: Hub for Pytorch Hub models, HF for trained HuggingFace models, and RNN for InferSent models. For Pytorch Hub models, you do not need to manually download the models as it is taken care by fairseq.

English language models trained on MNLI

Chinese language models trained on OCNLI corpus

InferSent based models, trained on MNLI and OCNLI corpus

(Same location containing all folders and files)

Maximum Entropy Trained Models

Pre-requisites

Download and build the original MNLI, SNLI, ANLI and OCNLI data following the code in ANLI, which places the data in anli/build/{data name}. You can use the following script to prepare the datasets in one go:

cd anli
source setup.sh
bash ./anli/script/download_data.sh
python src/dataset_tools/build_data.py

Run Evaluation

We make use of Pytorch FairSeq, ANLI scripts and InferSent repository script for running evaluation on the word shuffled data and computing our metrics. Below we provide the steps to evaluate from each of these settings.

• In each of the datasets, the script first generates the word shuffled corpus and creates it at {outp_path}/{eval_data}/rand_100_p_1.0_k_0.0_stop_False_punct_True/rand.jsonl
• For each of the below cases, after the run evaluation is complete, you can view the outputs in {outp_path}/{eval_data}/rand_100_p_1.0_k_0.0_stop_False_punct_True/outputs

Hub Evaluation

• Models: roberta.large.mnli, bart.large.mnli

• Run evaluation, where eval_data can be any one of Data names listed above. For example, to run evaluation on MNLI Matched development set, run:

  python main.py --eval_data mnli_m_dev --model_type hub --model_name roberta.large.mnli

  Note : You can modify the batch_size parameter in config.yaml to your liking for faster inference.

• You can use any FairSeq and Pytorch Hub model trained on MNLI to explore similar setups.

ANLI Evaluation

• Models: distilbert.mnli, roberta.large.ocnli

• Download the appropriate model from the above table and place it in anli/models folder

• For DistilBERT, run the following

  ./scripts/distilbert_eval.sh

• For OCNLI, run the following

  ./scripts/ocnli_eval.sh

• You can use any model trained using HuggingFace Transformers as done in ANLI repository in this section following the above scripts. To add your own model, add the path to your trained model in model_store.

InferSent Evaluation

• Models: infersent.mnli, convnet.mnli, bilstmproj.mnli

• Download the appropriate model from the table above and extract in the root folder (you should have rnn_models) in the main folder

• For InferSent evaluation, we need a slightly different format of the data. For MNLI and OCNLI, we provide run.sh in dataset_utils/mnli/ and dataset_utils/ocnli to convert the files appropriately. These scripts require rnn_models to exist in the main folder.

• Run evaluation, where eval_data can be any one of Data names listed above. For example, to run evaluation on MNLI Matched development set using InferSent model, run:

  python main.py --eval_data mnli_m_dev --model_type rnn_infersent --model_name infersent.mnli

• Model types are defined in config.yaml

• Example script to run other models is available in scripts/transformer_eval.sh

Compute Table

You can use the table script to generate the aggregate table as in the paper:

python codes/tables.py --data_loc <outp_path> python --datas mnli_m_dev --models roberta.large.mnli

• For more rows in the table, provide the eval_data in comma separated way in --datas and provide the models in --models in a comma separated way. Check the defaults in codes/tables.py for more information.

Compute Plots

You can use the plotting script to generate the plots used in the paper.

python codes/plot.py --plot entropy

POS Tree Analysis

To run the POS Tree Analysis experiments, run the following script, which first computes the train statistic on MNLI train data and then computes the POS Tree analysis on model outputs.

python codes/min_tree.py -k 4 --loc data/ --compute_train_stats --mp

If you have already computed the train statistics (output file in data/mnli_m_dev/min_tree_train.pkl), omit the --compute_train_stats flag.

Results

For reproducibility, you can check your results with our published results in the paper.

Citation

If you find our work useful in your research, please cite us using the following bibtex.

@article{sinha2021unnat,
  title           = {{UnNatural Language Inference}},
  author          = {Koustuv Sinha and Prasanna Parthasarathi and Joelle Pineau
                  and Adina Williams},
  journal         = {ACL},
  year            = 2021,
  url             = {https://arxiv.org/abs/2101.00010},
  arxiv           = {2101.00010},
}

Acknowledgement

• We thank the amazing ANLI repository, which we use for various prototyping and experiments.
• We also use the InferSent repository for non-Transformer experiments

Terms of Use

https://opensource.facebook.com/legal/terms

Privacy Policy

https://opensource.facebook.com/legal/privacy

License

UNLU is licensed under Creative Commons-Non Commercial 4.0. See the LICENSE file for more details.