Merikatori/hate-speech-knn

Hate Speech Detector — KNN Pipeline

KNN classifier cho bài toán phân loại hate speech trên Twitter.

Labels

• 0 — Hate Speech: ngôn ngữ thù ghét
• 1 — Offensive: xúc phạm nhưng không phải hate speech
• 2 — Neither: bình thường

Pipeline

• TF-IDF (15k features) + Chi2 selection (top 5000)
• Sentence Embeddings: all-MiniLM-L6-v2 (384 chiều)
• Meta features: word count, uppercase ratio, mention count, v.v.
• KNN (k=3, euclidean, distance-weighted, BallTree)
• Imbalance: sample_weight='balanced' (không ADASYN — tránh overfit)

Kết quả

Metric	Score
Accuracy	0.8574
Macro F1	0.6396
Weighted F1	0.8437

Load pipeline

import joblib
from huggingface_hub import hf_hub_download

path = hf_hub_download(repo_id="Merikatori/hate-speech-knn", filename="knn_pipeline.pkl")
pipeline = joblib.load(path)

# Predict
knn   = pipeline['knn']
# (cần chạy feature extraction trước — xem gradio_demo.py)