BHDAI: same-event matcher for Turkish news

BHDAI decides whether two stories report the same event. BültenHub uses it to group stories from different sources.

Model

Method: logistic regression on 10 features: kos_baslik, kos_baslik_tum, kos_tum, ozel_oran, ozel_az, kaynak_baslik_kapsam, metin_baslik_dayanak, sayi_oran, sayi_yok, uzunluk_orani.

Version: bhdai-20261030, retrained nightly and promoted automatically after its exam.

Training data: 90,000 stories, 225,279 story pairs.

Exam

The exam uses the last 2 days: 5,229 stories.

Threshold	Positive (same event)	Hard pair (wrong match)	Random pair
0.05 (recommended)	0.454	0.014	0.000
0.1	0.666	0.031	0.000

Files

• agirlik.npz: weights
• sozluk.json: vocabulary
• bilgi.json: features, coefficients and exam details

License

CC BY-NC 4.0 (non-commercial). For commercial use, contact https://bultenhub.com.