mediabiasgroup/roberta-anno-lexical-ft

RoBERTa-SA-FT (Anno-lexical)

This model is a sentence-level media (lexical) bias classifier. It is the Synthetic-Annotations Fine-Tuned (SA-FT) classifier from the paper “The Promises and Pitfalls of LLM Annotations in Dataset Labeling: a Case Study on Media Bias Detection” (NAACL Findings 2025; arXiv:2411.11081). It’s a roberta-base encoder with a 2-layer classification head trained on the Anno-lexical dataset (48,330 sentences; 70/15/15 splits). Labels are binary: 0 = neutral/non-lexical-bias, 1 = lexical-bias.

Paper: The Promises and Pitfalls of LLM Annotations in Dataset Labeling: a Case Study on Media Bias Detection
Dataset: mediabiasgroup/anno-lexical

Intended use & limitations

• Intended use: research on lexical/loaded-language bias in news-like English sentences; benchmarking SA-FT vs human-annotation fine-tuning.
• Out-of-scope: detecting non-lexical forms of media bias (e.g., informational/selection bias), political leaning, stance, or factuality; production deployments without human oversight.
• Known caveats: Compared with a model fine-tuned on human labels (HA-FT), SA-FT emphasizes recall over precision; robustness to perturbations is weaker than HA-FT.

How to use

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

m = "mediabiasgroup/roberta-anno-lexical-ft"
tok = AutoTokenizer.from_pretrained(m)
model = AutoModelForSequenceClassification.from_pretrained(m)

text = "Democrats shamelessly rammed the bill through Congress."
inputs = tok(text, return_tensors="pt")
probs = model(**inputs).logits.softmax(-1).tolist()[0]
print({"neutral": probs[0], "lexical_bias": probs[1]})

Training data & setup

Training data: Anno-lexical (48,330 sentences; binary lexical-bias labels aggregated via majority vote from LLM annotators).

Base encoder: roberta-base; head: 2-layer classifier.

Hardware: single A100; single-run training.

Evaluation

BABE (test): P 0.875 / R 0.814 / F1 0.843 / MCC 0.662

BASIL (all): P 0.171 / R 0.502 / F1 0.254 / MCC 0.205 (Positive class = lexical bias; BASIL informational bias treated as neutral.)

Safety, bias & ethics

Media bias perception is subjective and culturally dependent. This model may over-flag biased wording and should not be used to penalize individuals or outlets. Use with human-in-the-loop review and domain-specific calibration.

Citation

If you use this model, please cite:

@inproceedings{horych-etal-2025-promises,
  title = "The Promises and Pitfalls of {LLM} Annotations in Dataset Labeling: a Case Study on Media Bias Detection",
  author = "Horych, Tom{\'a}{\v{s}}  and
    Mandl, Christoph  and
    Ruas, Terry  and
    Greiner-Petter, Andre  and
    Gipp, Bela  and
    Aizawa, Akiko  and
    Spinde, Timo",
  editor = "Chiruzzo, Luis  and
    Ritter, Alan  and
    Wang, Lu",
  booktitle = "Findings of the Association for Computational Linguistics: NAACL 2025",
  month = apr,
  year = "2025",
  address = "Albuquerque, New Mexico",
  publisher = "Association for Computational Linguistics",
  url = "https://aclanthology.org/2025.findings-naacl.75/",
  doi = "10.18653/v1/2025.findings-naacl.75",
  pages = "1370--1386",
  ISBN = "979-8-89176-195-7"
}

---