mohummedalee / twitteraae-sentiment-data Goto Github PK

This repo contains code to construct sentiment data for African American Vernacular English (AAE) vs. Standard American English (SAE) from tweets.

The original TwitterAAE dataset is from Blodgett et. al. (2016). While the original paper aimed to train a dialect-estimation model, once dialect has been estimated, the data can be used to study disparities in performance of AAVE vs. SAE for NLP tasks --- sentiment is one such task. Prior work that uses sentiment from TwitterAAE includes Bagdasaryan et. al. (2019) and Elazar et. al. (2018).

The heuristic to estimate sentiment is from Elazar et. al. and is fairly simple (even though the unicode wrestling to achieve it is not). If a tweet includes 🙂, 😈, 😆, 😅, 😂, 😀, 😺 etc., it's positive. If it includes 😡, 😤, 😰, 🤢, 👿 etc., it's negative. If both types of emojis occur, sentiment cannot be inferred. Appendix C in Elazar et. al. has the full emoji list.

Code in this repo is taken from Elazar et. al.'s Github. I did the labor of running an extinct version of Python to re-create this dataset and share it.

• data/raw/sentiment_race contains full tweet texts, files for AAE and SAE are separate, files for positive ("happy") and negative ("sad") sentiments are separate.
• data/processed/sentiment_race contains the tokenized dataset (using twokenize.py). Each row contains token IDs corresponding to line number invocab.

First, build the conda environment to run this code. You will need vintage Python 2.7

CONDA_SUBDIR=osx-64  # this is for backwards compatibility with Apple silicon
conda create -f env_py27.yml
conda activate py27

If conda create -f doesn't work and you'd like to instantiate a new Python 2 environment, you can do so with CONDA_SUBDIR=osx-64 conda create -n py27 python=2.7 and then conda install --file env_py27.yml.

• Download the original dataset (TwitterAAE-full-v1.zip) from http://slanglab.cs.umass.edu/TwitterAAE/ and unzip it
• Run python make_data.py <path_to_unzipped_folder>/twitteraae_all raw/sentiment_race processed/sentiment_race sentiment race. The parameters are: source data path, raw output path, tokenized output path, task type (sentiment), demographic type (race).

If anyone has the time to port this codebase and all its unicode regexes to Python 3, it would be a good thing to do --- happy to take PRs here.