This repository contains code and data for the paper entitled "A Group-Specific Approach to NLP for Hate Speech Detection." The starter code and data (and experimental approach) were copied from https://github.com/NasLabBgu/hate_speech_detection.
The main contributions of this paper on the data front are EchoKG and KnowledJe, both of which are further described in the paper.
The pertinent EchoKG files are this one, generated with no edits to the phrases found in the posts, and this one, generated with edits of distance <= 2 of the phrases in the posts.
The full KnowledJe knowledge graph is here in JSON form, and here in TXT form. An abridged version that yielded slightly better results in our experiments is here in JSON form and here in TXT form.
To generate your own versions of KG-augmented data, add the JSON-based KGs to the data folder, and then run detection/detection_utils/kg_to_data.py after changing the file paths on the last two lines, and then run detection/detection_utils/add_kg.py after changing the file paths appropriately on the last line. Both files can be run via python detection/detection_utils/file_name.py.
To avoid environment issues you can create a new conda env using the following line, replacing <env> with your desired environment name:
conda create --name <env> --file requirements.txt
Once you've activated the environment via conda activate <env>, you can run jupyter notebook to open up a Jupyter Lab session on your browser. From there, navigate to detection/experiments/echo_bert.ipynb, change the kg_data_conf and base_data_conf values as appropriate, and run each cell in that notebook in sequence to produce experimental results.