GPT3-HateEval

Code implementation for the paper "Evaluating GPT-3 Generated Explanations for Hateful Content Moderation" (IJCAI'23).

Getting Started

• The dataset used in this project is HateXplain. We performed clustering on the tweets from the HateXplain dataset based on their annotation labels and stored the clustered data in the data folder.

• The project's code can be found in the scripts folder. The ExplanationGeneration.ipynb file is responsible for generating explanations, which are then stored in the explanations folder. The explanations folder contains two subfolders:

  • The demonstration_examples subfolder stores the candidate pool for demonstration examples as well as the selected demonstration examples. These examples are used in few-shots prompting to generate explanations.

  • The evaluation_explanations subfolder stores the explanations generated for all evaluation tweets. The tweet and explanation pairs in this folder will be evaluated by human annotators based on the quality of the explanations and the hatefulness of the tweets.

• The annotation folder contains all the annotation results collected from the crowd sourcing platform. To preprocess and analyze the human annotations, refer to the ResultsAnalysis.ipynb file. The folder evaluation_table comprises the table presenting the results of the analysis. A visually enhanced representation of the data is presented in the subsequent section.

Findings derived from human evaluation.

The assessment of the quality of explanations.

The assessment of the persuasive impact of explanations on the perception of hatefulness in the same set of tweets.

The manifestation of misclassification of tweets influenced by misleading explanations.

The labeling process for each tweet entails initially classifying each annotation as either 'non-hateful' or 'hateful' based on their respective hatefulness scores of 1-2 and 4-5. Each tweet undergoes evaluation by three distinct human evaluators, and the majority vote among them determines the annotated label assigned to the tweet.

Acknowledgements

To cite:

@inproceedings{gpt3hateeval2023,
    title={Evaluating GPT-3 Generated Explanations for Hateful Content Moderation},
    author={Wang, Han and Hee, Ming Shan and Awal, Md Rabiul and Choo, Kenny Tsu Wei and Lee, Roy Ka-Wei},
    year={2023},
    booktitle = {In Proceedings of the International Joint Conference on Artificial Intelligence (IJCAI)},
}

Correspondence

Wang Han ([email protected])