This algorithm is derived from the BiRNN-HateXplain algorithm, and this project is based on the HateXplain project https://github.com/hate-alert/HateXplain/tree/master (Associated article : https://arxiv.org/pdf/2012.10289)
The results of our studies can be found in the models_and_results/BiAtt_BiRNN_max_2 folder.
File example_hatexplain_with_BiRNN-HateXplain_vs_BiAtt_BiRNN-HateXplain.ipynb contains examples of comparison between ground truth attention, attention predicted by BiRNN-HateXplain and that predicted by BiAtt-BiRNN-HateXplain.
To train a proposed model use the file Example_HateExplain.ipynb
The objective of this project is to improve the results of the BiRNN-HateXplain and BERT-HateXplain algorithms in terms of detection performance, unintentional bias, and explainability.
In current algorithms such as BiRNN-HateXplain, we observe a large variation in the estimated attention when it should be constant.
We observe for example that in the interval [8, 20] of the plot above, the attention estimated by the BiRNN-HateXplain model varies a lot when it should be constant.
Our hypothesis is that considering the sequential aspect of input data in HateXplain models could resolve the variability of attention and improve explainability. And then, it can also improve classification performance and unintentional biases related to communities indexed in hate speech because it uses multi-task learning(classification and explainability tasks).
The results show that the proposed approach improves explainability, prediction performance, and metrics that measure unintentional biases of the model. We also observed that the attention estimated by the proposed approach estimates constant attention when it should be.
In the above figure in the interval [8, 20] compared to BiRNN-HateXplain, the attention of the proposed model is constant when it should be in reality.
It is recommended to use a tool like conda to create a virtual environment and facilitate conflict management. Install the appropriate packages contained in the requirements.txt file