Skip to content

Latest commit

 

History

History
36 lines (26 loc) · 2.88 KB

README.md

File metadata and controls

36 lines (26 loc) · 2.88 KB
Raw

Detection of Normal, Hate, and Offensive Speeches

Dataset Source: https://www.kaggle.com/datasets/subhajournal/normal-hate-and-offensive-speeches

Normal, Hate, and Offensive Speeches The data has been scraped from Twitter through Apify regarding the tweets of normal, hate and offensive speeches. The complete database contains 12 datasets where the data are available for normal, hate and offensive speeches with four CSV files each. To create the complete standalone data, all 12 datasets need to be merged by relevant file labels.

Highlights:

  1. handled mild class imbalance
  2. data visulaizations
  3. weight normalization
  4. compared performance of various transformers
Model Accuracy Weighted F1 Weighted Recall Weighted Precision Micro F1 Micro Recall Micro Precision Macro F1 Macro Recall Macro Precision Training Time Evaluation Time
bert-base-uncased 0.9935 0.9935 0.9935 0.9935 0.9935 0.9935 0.9935 0.9933 0.9931 0.9936 83.87s 650ms
bert-large-uncased 0.9935 0.9935 0.9935 0.9936 0.9935 0.9935 0.9935 0.9932 0.9938 0.9927 275.92s 1.74s
distilbert-base-uncased 0.9935 0.9935 0.9935 0.9936 0.9935 0.9935 0.9935 0.9932 0.9938 0.9927 46.27s 365ms
diptanu/fBERT 0.9935 0.9935 0.9935 0.9936 0.9935 0.9935 0.9935 0.9932 0.9938 0.9927 81.92s 634ms
GroNLP/hateBERT 0.9902 0.9902 0.9902 0.9904 0.9902 0.9902 0.9902 0.9901 0.9903 0.9899 81.93s 637ms
roberta-large 0.9837 0.9837 0.9837 0.9839 0.9837 0.9837 0.9837 0.9832 0.9821 0.9845 281.49s 1.74s
Model Macro F1 Score Eval Loss Eval Runtime
bert-base-uncased 0.9933 0.0164 650ms
bert-large-uncased 0.9932 0.0097 1.74s
distilbert-base 0.9932 0.0178 365ms
fbert/hate 0.9932 0.0152 634ms
hateBERT 0.9901 0.0207 637ms
roberta-large 0.9832 0.0293 1.74s