RoBERTa — BABE — HA-FT
This repository provides a RoBERTa-base model fine-tuned on the BABE (Bias Annotations By Experts) dataset for sentence-level lexical/loaded-language bias detection in English news text. BABE was introduced in the paper Neural Media Bias Detection Using Distant Supervision With BABE – Bias Annotations By Experts.
Labels
0→ neutral / non-lexical-bias1→ lexical-bias
Intended use & limitations
- Intended use: research and benchmarking of lexical bias at the sentence level on news-like English text.
- Out-of-scope: detection of informational/selection bias, stance, political leaning, or factuality; production deployments without human oversight.
How to use
from transformers import AutoTokenizer, AutoModelForSequenceClassification
m = "mediabiasgroup/roberta-babe-ft"
tok = AutoTokenizer.from_pretrained(m)
model = AutoModelForSequenceClassification.from_pretrained(m)
text = "Democrats shamelessly rammed the bill through Congress."
probs = model(**tok(text, return_tensors="pt")).logits.softmax(-1).tolist()[0]
print({"neutral": probs[0], "lexical_bias": probs[1]})
Training data & setup
- Data: BABE (expert-annotated, sentence-level lexical bias).
- Backbone:
roberta-basewith a standard sequence-classification head. - Training: single-run fine-tuning; standard hyperparameters (update with your exact config if desired).
Safety, bias & ethics
Media-bias perception is subjective and context-dependent. This model may over-flag emotionally charged wording. Keep a human in the loop and avoid punitive or outlet-level decisions without careful validation.
Citation
If you use this model or the dataset, please cite:
@article{spinde2022neural,
title = {Neural Media Bias Detection Using Distant Supervision With BABE -- Bias Annotations By Experts},
author = {Spinde, Timo and Plank, Manuel and Krieger, Jan-David and Ruas, Terry and Gipp, Bela and Aizawa, Akiko},
journal = {arXiv preprint arXiv:2209.14557},
year = {2022},
url = {https://arxiv.org/abs/2209.14557}
}