SimileGeneration-EMNLP2020

SET UP ENVIRONMENT

conda create --name simile python=3.7.3

conda activate simile

#point your LD_LIBRARY_PATH to your miniconda or anaconda library

export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/nas/home/miniconda3/lib/

install nltk

SET UP COMET

cd comet-commonsense

To run the setup scripts to acquire the pretrained model files from OpenAI, as well as the ATOMIC and ConceptNet datasets

  bash scripts/setup/get_atomic_data.sh
  bash scripts/setup/get_conceptnet_data.sh
  bash scripts/setup/get_model_files.sh
  Then install dependencies (assuming you already have Python 3.6 ):

  pip install torch==1.3.1
  pip install tensorflow
  pip install ftfy==5.1
  conda install -c conda-forge spacy
  python -m spacy download en
  pip install tensorboardX
  pip install tqdm
  pip install pandas
  pip install ipython
  pip install inflect
  pip install pattern
  pip install pyyaml==5.1
  pip install regex==2018.2.21
  pip install pytorch_pretrained_bert==0.6.0

Making the Data Loaders Run the following scripts to pre-initialize a data loader for ATOMIC or ConceptNet:

python scripts/data/make_atomic_data_loader.py
python scripts/data/make_conceptnet_data_loader.py

Download pretrained COMET First, download the pretrained models from the following link:

https://drive.google.com/open?id=1FccEsYPUHnjzmX-Y5vjCBeyRt1pLo8FB Then untar the file:

tar -xvzf pretrained_models.tar.gz

SET UP DATA PROCESSING FOR SIMILE

If you don't want to generate new data, you can download train and val files from https://drive.google.com/drive/folders/1KSANJ7XiPo0xqFCUG5WDhB3763EgEVnC?usp=sharing Put inside simile folder

To generate data on your own

            python scrape_reddit_for_similes.py #Scrapes self labeled similes
            python convert_to_literal.py #Converts simile to literal

FINETUNE BART

You can download our finetuned model directly from here and put checkpoint_best.pt in checkpoint-simile if you don't want to finetune. https://drive.google.com/drive/folders/1KSANJ7XiPo0xqFCUG5WDhB3763EgEVnC?usp=sharing

  Use data from previous step for fine-tuning BART
  Sample data in simile folder (train.source --> LITERAL , train.target---> SIMILE)

  Use the encoder.json and dict.txt already provided as a part of the repo, since it contains additional delimeter tokens relevant for simile generations


  Now for BPE preprocess:
  sh create_bpe.sh


  Binarize dataset:

  sh preprocess.sh


  Download Pretrained BART from here:

  https://github.com/pytorch/fairseq/tree/4b986c51ed89f9895df2f30e57909301b4a4f19b/examples/bart

  Train models:

  sh finetune.sh

Update the field BART_PATH to suit where your pretained model.pt file is You can customize MAX_TOKENS and UPDATE_FREQ based on gpu memory / no of gpus

For Inference:

Test Data There are two files gen1.source , gen2.source which contains 160 similes. We did 10 generations extra for backup , but 150 of them were used for evaluation. Outputs of all baselines along with human written simile for 150 utterances is available in SimileEMNLP.csv

The auto eval script is also available

Change the input file names n output file names based on what you are giving When you run generate.py give the two files (gen1.source , gen2.source) seperately in each run for exact reproducibility. If you want to test on your own literal sentence, simply give a sentence per line in any text file as input

python generate.py

For Citation I will update bibtex once ACL anthology is out . For now to cite arxiv paper

         @inproceedings{chakrabarty-etal-2020-generating,
            title = "Generating similes effortlessly like a Pro: A Style Transfer Approach for Simile Generation",
            author = "Chakrabarty, Tuhin  and
              Muresan, Smaranda  and
              Peng, Nanyun",
            booktitle = "Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)",
            month = nov,
            year = "2020",
            address = "Online",
            publisher = "Association for Computational Linguistics",
            url = "https://www.aclweb.org/anthology/2020.emnlp-main.524",
            pages = "6455--6469",
            abstract = "Literary tropes, from poetry to stories, are at the crux of human imagination and communication. Figurative language such as a simile                               go beyond plain expressions to give readers new insights and inspirations. In this paper, we tackle the problem of simile generation.                               Generating a simile requires proper understanding for effective mapping of properties between two concepts. To this end, we first                                   propose a method to automatically construct a parallel corpus by transforming a large number of similes collected from Reddit to their                             literal counterpart using structured common sense knowledge. We then propose to fine-tune a pre-trained sequence to sequence model,                                 BART (Lewis et al 2019), on the literal-simile pairs to gain generalizability, so that we can generate novel similes given a literal                               sentence. Experiments show that our approach generates 88{\%} novel similes that do not share properties with the training data. Human                             evaluation on an independent set of literal statements shows that our model generates similes better than two literary experts 37{\%}                               of the time when compared pairwise. We also show how replacing literal sentences with similes from our best model in machine-generated                             stories improves evocativeness and leads to better acceptance by human judges.",
                    }