数据集:
mediabiasgroup/mbib-base
Task | Model | Micro F1 | Macro F1 |
cognitive-bias | ConvBERT/ConvBERT | 0.7126 | 0.7664 |
fake-news | Bart/RoBERTa-T | 0.6811 | 0.7533 |
gender-bias | RoBERTa-T/ELECTRA | 0.8334 | 0.8211 |
hate-speech | RoBERTA-T/Bart | 0.8897 | 0.7310 |
linguistic-bias | ConvBERT/Bart | 0.7044 | 0.4995 |
political-bias | ConvBERT/ConvBERT | 0.7041 | 0.7110 |
racial-bias | ConvBERT/ELECTRA | 0.8772 | 0.6170 |
text-leve-bias | ConvBERT/ConvBERT | 0.7697 | 0.7532 |
All datasets are in English
An example of one training instance looks as follows.
{ "text": "A defense bill includes language that would require military hospitals to provide abortions on demand", "label": 1 }
We believe that MBIB offers a new common ground for research in the domain, especially given the rising amount of (research) attention directed toward media bias
@inproceedings{ title = {Introducing MBIB - the first Media Bias Identification Benchmark Task and Dataset Collection}, author = {Wessel, Martin and Spinde, Timo and Horych, Tomáš and Ruas, Terry and Aizawa, Akiko and Gipp, Bela}, year = {2023}, note = {[in review]} }