MISO-BR
An open corpus for misogyny detection in Brazilian Portuguese. It contains 896 annotated social media texts, construction protocols, annotation guidelines, models, and experiment scripts, with emphasis on auditability and reproducibility.
Timeline (last 90 days)
MISO-BR contains 896 balanced social media texts, all annotated with binary misogyny labels.
Three annotators, following a written guideline, reached a Fleiss' kappa of 0.788 agreement on 816 data items.
MISO-BR was constructed by combining lexicon-guided and target-guided crawling for the X platform, and incorporated entries from OLID-BR.
After collection, MISO-BR went through a large language model triage stage, and the prompts and decoding parameters of this stage are likewise publicly available.
Under a unified fixed protocol, the macro-F1 of classical classifiers (multinomial and Bernoulli naive Bayes, linear SVM, random forest) clustered between 0.651 and 0.676.
The fine-tuned BERTimbau model reached a macro-F1 and accuracy of 0.833, with a Wilson 95% confidence interval for accuracy of [0.77, 0.88].
The MISO-BR authors explicitly list three limitations: manual class balancing, the association between sources and labels, and the selection bias introduced by lexicon-guided sampling and automatic triage.
MISO-BR's corpus, annotation guidelines, models and experiment scripts are all publicly available.
All Facts (8)
MISO-BR contains 896 balanced social media texts, all annotated with binary misogyny labels.
50%UnverifiedThree annotators, following a written guideline, reached a Fleiss' kappa of 0.788 agreement on 816 data items.
50%UnverifiedMISO-BR was constructed by combining lexicon-guided and target-guided crawling for the X platform, and incorporated entries from OLID-BR.
50%UnverifiedAfter collection, MISO-BR went through a large language model triage stage, and the prompts and decoding parameters of this stage are likewise publicly available.
50%UnverifiedUnder a unified fixed protocol, the macro-F1 of classical classifiers (multinomial and Bernoulli naive Bayes, linear SVM, random forest) clustered between 0.651 and 0.676.
50%UnverifiedThe fine-tuned BERTimbau model reached a macro-F1 and accuracy of 0.833, with a Wilson 95% confidence interval for accuracy of [0.77, 0.88].
50%UnverifiedThe MISO-BR authors explicitly list three limitations: manual class balancing, the association between sources and labels, and the selection bias introduced by lexicon-guided sampling and automatic triage.
50%UnverifiedMISO-BR's corpus, annotation guidelines, models and experiment scripts are all publicly available.
50%