Skip to content

Latest commit

 

History

History
99 lines (69 loc) · 3.91 KB

README.md

File metadata and controls

99 lines (69 loc) · 3.91 KB
Raw
MTNT

MTNT: A Testbed for Machine Translation of Noisy Text

Codacy Badge

This repo contains the code for the EMNLP 2018 paper MTNT: A Testbed for Machine Translation of Noisy Text. It will allow you to reproduce the collection process as well as the MT experiments. You can access the data here.

Prerequisites

For preprocessing, you will need Moses (for tokenization, clean-up, etc...), sentencepiece (for subwords) and KenLM (for n-gram language modeling). If you want to work with japanese data you should also install Kytea (for word segmentation)

To run the collection code, you will need the following python modules:

kenlm
langid
numpy
pickle
praw
sentencepiece>=0.1.6
yaml

Finally, for the MT experiments, refer to the README in the recipes folder

Downloading and Preparing the Data

From this folder, run

# Monolingual en data from WMT17
bash scripts/download_en.sh config/data.en.config
bash scripts/prepare_model config/data.en.config

# Monolingual fr data from WMT15
bash scripts/download_fr.sh config/data.fr.config
bash scripts/prepare_model config/data.fr.config

# Prepare en<->fr parallel data
bash scripts/prepare-en-fr.sh config/data.fr.config path/to/moses/scripts

# Download and prepare the en<->ja monolingual and parallel data
bash scripts/download_ja.sh config/data.ja.config path/to/moses/scripts

# Download and extract MTNT
wget http://www.cs.cmu.edu/~pmichel1/hosting/MTNT.1.0.tar.gz && tar xvzf MTNT.1.0.tar.gz && rm MTNT.1.0.tar.gz
# Split the tsv files
bash MTNT/split_tsv.sh

You can edit the config/data.{en,fr,ja}.config files to change filenames, subword parameters, etc...

Running the Scraper

Edit the config/{en,fr,ja}_reddit.yaml to include the appropriate credentials for your bot. You can also change some of the parameters (like subreddits, etc...).

Then run

bash scripts/start_scraper.sh [config_file]`

When running the scraper, be mindful of the Reddit API terms.

Analysing the Data

You can analyse the collected data using the various scripts in the analysis folder, for example:

# Count the number of profanities (should return 38)
cat MTNT/test/test.en-fr.en | python3 analysis/count_keywords.py resources/profanities.en
# Count the number of emojis (should return 46)
cat MTNT/test/test.en-fr.en | python3 analysis/count_emojis.py
# Check the ration US/UK spelling (for ise/ize which is a good indicator) (should return 35.7% 64.3%)
cat MTNT/test/test.en-fr.en | python3 analysis/uk_us_ratio.py
# Count the number of informal pronouns (in japanese) (should return 35)
kytea -model /path/to/kytea/data/model.bin -out tok MTNT/test/test.ja-en.ja | python3 analysis/count_keywords.py resources/informal_pronouns.ja

Citing

If you use this code or the MTNT dataset, please cite the following publication:

@InProceedings{michel2018mtnt,
  author    = {Michel, Paul  and  Neubig, Graham},
  title     = {{MTNT}: A Testbed for {M}achine {T}ranslation of {N}oisy {T}ext},
  booktitle = {Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing (EMNLP)},
  year      = {2018}
}

License

The code is released under the MIT License. The data is released under the terms of the Reddit API