search

hate-alert / Tutorial-Resources

Resources and tools for the Tutorial - "Hate speech detection, mitigation and beyond" presented at ICWSM 2021
MIT License
35 stars 6 forks source link

Performance for the Rationale predictor demo #1

Open aifangchai opened 6 months ago

aifangchai commented 6 months ago

Great work on the hate speech detection. However, the performance of the model in the Rationale Predictor demo seems less accurate than desired. I tried a few sentences, and the results were labeled as the normal class instead of the abusive class. I suspect that the model is biased toward the normal class. Could you please suggest ways to improve the model's performance?

Here are some examples of sentences and their results: I will kill you. {'Normal': 0.51631415, 'Abusive': 0.48368585} I hate the rich people. {'Normal': 0.8278808, 'Abusive': 0.17211922}

Hope to hear from you soon. Thank you.

punyajoy commented 6 months ago

For both these cases the ideal way would be train the rationale predictor model on such datapoints.

Although the first statement is very ambiguous and the target is not specified. It might be said as a friendly banter. In the second one the target does not represent any vulnerable groups hence it might misclassify it.

aifangchai commented 6 months ago

Dear Hate-Alert/Tutorial-Resources,

Thanks for your reply. May I know the data format for the training process? Or any other considerations for the training process?

Thanks for your response.

Best regards, Chai

On Fri, 16 Feb 2024 at 14:52, Punyajoy Saha @.***> wrote:

For both these cases the ideal way would be train the rationale predictor model on such datapoints.

Although the first statement is very ambiguous and the target is not specified. It might be said as a friendly banter. In the second one the target does not represent any vulnerable groups hence it might misclassify it.

— Reply to this email directly, view it on GitHub https://github.com/hate-alert/Tutorial-Resources/issues/1#issuecomment-1947847608, or unsubscribe https://github.com/notifications/unsubscribe-auth/A6ZF4EBM3SADQCUPVHGTAGTYT3625AVCNFSM6AAAAABDLQHMXGVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTSNBXHA2DONRQHA . You are receiving this because you authored the thread.Message ID: @.***>