kg-hateguard-eval

Try out our hybrid models in this Demo! 👉

This is the data and code in the paper Knowledge-Grounded Target Group Language Recognition in Hate Speech, together with conference slides and a poster!

We use a Knowledge Graph (KG), and adapt it, to provide interpretability to deep learning predictions as to why hate speech texts may refer to particular target group identities (Figure 1). Specifically, by enriching probability scores with the entities at input that contribute to a prediction.

Grounding the classification task in semantic knowledge gives context about the identity groups most impacted by these technologies. Our experiments show how knowledge-grounded interpretations help better understand model outcomes, the training data, and the ambiguious cases in human annotations.

Project description

The project is organised in three folders. data contains the hate speech datasets, baselines the lexicon-based (System B) and transformer-based (System A) models used in the paper, and models the files from model training.

There are two main code files:

• kg_adaptation.py: script to learn weights for the KG entities based on a classification task (./models/adaptation).
• identity_group_identification.py: script to train text classifiers based on adapted KG features or Huggingface transformer embeddings [1].

If you are planning to use a KG as feature extractor in a text classification task, get in touch!

Requirements

Conda Environment

Python 3.8.2 and install requirements.txt (generated using pipreqs).

    $ conda create --name <env_name> python=3.8.2
    $ conda activate <env_name>
    (<env_name>) $ python -m pip install -r requirements.txt
    # Inside virtual environment to run notebooks with this conda environment
    (<env_name>) $ python -m ipykernel install --user --name=<env_name>
    (<env_name>) $ jupyter notebook

Docker

    # Build docker image
    $ docker build . -t <docker-image>:tag

    # Run interactive docker container
    $ docker run -it --name <docker-container>:tag -v `pwd`:/app <docker-image>:tag
    # ... with port to open jupyter notebooks
    $ docker run -it -p 8888:8888 --name <docker-container>:tag -v `pwd`:/app <docker-image>:tag

    # Inside the container to run notebooks
    $ jupyter notebook --ip 0.0.0.0 --port 8888 --no-browser --allow-root
    # ... or from command line with nbconvert
    $ jupyter nbconvert --to notebook --execute my_notebook.ipynb --output=result.ipynb

Optional (To train Transformer models on GPU)

Follow instructions to install TF with GPU-enabled using a Virtual Environment if applicable. Ensure to install requirements after setting up the virtual environment as indicated above (Conda Environmnt). To use a specific GPU, adjust CUDA_VISIBLE_DEVICES in identity_group_identification script.

Paper experiments

All results from the paper are in notebooks folder.

We provide two bash scripts to reproduce all models (hate-speech-identities <username>$ bash notebooks/<script-name>.sh &> notebooks/<script-name>.log), including the export of Jigsaw Sample (jigsaw_0.5_gendersexualorientation.csv). Please contact us if you would prefer us to share it directly.

All paper results are shown in the following Jupyter notebooks:

• 1_data_statistics.ipynb: statistics for the datasets used in the paper (Table 2).
• 2_KG_adaptation.ipynb: evaluation for different weighting schemes in the KG adaptation phase.
• 3_identity_group_identification.ipynb: evaluation of hybrid, transformers, and lexicon-based models (Figure 3, Table 3, Table 4, Table 5).

The excel files in models/interpretations contain the qualitative analyses of model errors and interpretability.

Resources

• Jigsaw Toxicity Corpus (all_data.csv): download directly via Kaggle
• Gab Hate Corpus (GabHateCorpus_annotations.tsv): download through this link
• HateXplain (dataset.json): obtained from this repo
• XtremeSpeech (kenya_re.csv): keep an eye on its repo, and contact the authors to request for access.
• Gender, Sex, and Sexual Orientation Ontology (gsso.owl): v2.0.10, latest release at repo

Meauring Hate Speech (measuring-hate-speech.csv) downloads directly from Huggingface when using identity_group_identification.py for model training.

Baseline repositories

[1] Targeted Identity Group Prediction in Hate Speech Corpora by Sachdeva et al (2022). In Proceedings of the Sixth Workshop on Online Abuse and Harms (WOAH). Association for Computational Linguistics, Seattle, Washington (Hybrid), 231–244.

[2] Challenges in Automated Debiasing for Toxic Language Detection by Zhou et al (2021). Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume.

Citation

@incollection{reyero2023knowledge,
    title       = {Knowledge-Grounded Target Group Language Recognition in Hate Speech},
    author      = {Reyero Lobo, Paula and 
                   Daga, Enrico and 
                   Alani, Harith and 
                   Fernandez, Miriam},
    booktitle   = {Knowledge Graphs: Semantics, Machine Learning, and Languages},
    pages       = {1--18},
    year        = {2023},
    publisher   = {IOS Press}
}