2021 · Peer-reviewed journal article

BERT Modeli’nin Sınıflandırma Doğruluğunun Sıfır-Atış Öğrenmesi ile Artırılması

We examined how mismatches between an advertisement and its industry label affect a BERT text classifier.

Research question

What if the label on an ad is the mistake?

An online ad’s sector label does not always match what the ad actually says. The researchers trained BERT to identify industries from roughly 101,000 ads labelled across 44 sectors.

When some sectors proved hard to tell apart, they used a zero-shot model as a second opinion on whether each text fit its label. Filtering out mismatches and retraining BERT improved classification accuracy in their experiment.

Source attribution
Şükrü Ozan, Umut Özdil, Davut Emre Taşar, Büşra Arslan, Gökçe Polat
Publication status
Peer-reviewed journal article
Publication or platform
Türkiye Bilişim Vakfı Bilgisayar Bilimleri ve Mühendisliği Dergisi
Identifier
10.54525/tbbmd.1004781

Sources

Research areas · Language

Related work