PAN 2021: Profiling Hate Speech Spreaders on Twitter

This repository summarizes the work presented to the shared task "Profiling Hate Speech Spreaders on Twitter". Pull requests or issues are encouraged.

What to find in this repository

Implementations to tackle this problem by means of:

• n-grams: src/ngrams
• fasttext: src/fasttext-supervised
• BERT: src/simpletransformers-bert

Please see the next section to check which dependencies you should install before executing the programs.

Dependencies

The whole project has several dependencies, most of which are covered by executing pip install -r requirements.txt:

• nltk
• fasttext
• simpletransformers
• pandas
• KenLM (not covered by requirements.txt).
  • Install KenLM through your package manager of choice, or build it from source.
  • Tested: Manjaro Linux (installed through pamac-manager, AUR package here).

data/ structure

data/plain_text/

• Raw data (only sentences)
• {hater}_{lang}.txt: one sentence for each line

data/tok/

• Tokenized data by NLTK's TweetTokenizer
• {hater}_{lang}.tok.txt: all tokenized sentences
• {hater}_{lang}.tok_grouped.txt: each line has all tokenized sentences from a writer (i.e. 100 tweets, or lines, become a single line)

data/tok/partitioned_data/

• {hater}_{lang}.tok.{part}.txt: sentences divided in partitions train/dev/eval
• {hater}_{lang}.{part}.tok_grouped.txt: same but grouped as in data/tok/

Extract and partition the data

# Extract the data (assuming you have the source files)
tar xzf data.zip

# Extract the sentences from the data
python src/01-extract_text.py --extract_to data/plain_text

# Clean the text
python src/02-clean.py --dataset data/plain_text/nonhaters_es.txt
python src/02-clean.py --dataset data/plain_text/nonhaters_en.txt
python src/02-clean.py --dataset data/plain_text/haters_es.txt
python src/02-clean.py --dataset data/plain_text/haters_en.txt

mkdir data/tok
mv data/plain_text/*tok* data/tok

# Split the dataset in a train/dev/eval partition
python src/03-split_dataset.py --dataset data/plain_text/nonhaters_es.tok.txt
python src/03-split_dataset.py --dataset data/plain_text/nonhaters_en.tok.txt
python src/03-split_dataset.py --dataset data/plain_text/haters_es.tok.txt
python src/03-split_dataset.py --dataset data/plain_text/haters_en.tok.txt

Example: n-grams

cd src/ngrams

# Extract n-grams
./04-extract_ngrams.sh

# Obtain the accuracy of each developed n-gram with respect to the text
./05-reco.sh

# Get the accuracy
./06-get_accuracy.sh

DISCLAIMER

The scripts provided in the src/ directory are meant to be used as a guide. If you want to directly use them, you may want to adjust some things e.g. the directory where to launch them.