BLM-emotions

Official code and data of the paper "An Analysis of Emotions and the Prominence of Positivity in #BlackLivesMatter Tweets" (PNAS 2022)

We release 1) the weights of the BERT model pre-trained with our BLM tweets with a masked language modeling objective and 2) emotion recognition models fine-tuned with our human-annotated data. You can find them from this gdrive folder

However, we do not make the raw data freely available to preserve anonymity and privacy as much as possible. Still, we will make tweet ids and annotated data available only for academic research upon request. Please fill out this form to request the data.

Prerequisite

• change path_root in src_model/configs.py to your project directory path
• install the following python packages: torch, tensorboard, pandas, transformers, emoji, wordsegment, scikit-learn

Training new emotion recognition models

1. Prepare GoEmotions and HurricaneEmo dataset

bash scripts/download_emotion_datasets.sh
python src_model/preprocess_emotion_data.py

Executing the commands above will create a processed HurricaneEmo and GoEmotions dataset under data/processed-emotions, which can be used to train emotion recognition models. The two data sets use different emotion categories, but here we map both data sets into Ekman basic emotions using the mappings we defined in data/emo-mapping.

2. Train emotion recognition models

bash scripts/train_binary_model.sh $target_emotion $bert_model

• $target_emotion: one of six Ekman emotions (disgust, fear, anger, sadness, surprise, joy)
• $bert_model: {blm, none}. blm indicates the bert-base-uncased model from HuggingFace transformers further pre-trained with our BLM tweets. If you do not provide anything to the bert_model argument (i.e., none) the code will use the base pre-trained bert model (bert-base-uncased).

Generating emotion labels for given input texts

bash scripts/generate_binary_emotion_labels.sh $target_emotion $trained_model_path $path_input_txt_file $path_output

• $target_emotion: {disgust, fear, anger, sadness, surprise, joy}
• $trained_model_path: path to a trained model you want to generate labels from, e.g., blm_joy.pt.
• $path_input_txt_file: .txt file separated by lines
• $path_output: path to save the output and their predictions