This model was released with the following paper:
@proceedings{feedbackloop,
title = "Feedback Loops and Complex Dynamics of Harmful Speech in Online Discussions",
author = {Rong-Ching Chang, Jonathan May, and Kristina Lerman},
publisher = {Proceedings of the 16th International Conference on Social Computing, Behavioral-Cultural Modeling & Prediction and Behavior Representation in Modeling and Simulation.}
venue = {Pittsburgh, PA},
month = sep,
year = {2023}
}
We combined several multilingual ground truth datasets for misogyny and sexism (M/S) versus non-misogyny and non-sexism (non-M/S) [3,5,8,9,11,13, 20]. Specifically, the dataset expressing misogynistic or sexist speech (M/S) and the same number of texts expressing non-M/S speech in each language included 8, 582 English-language texts, 872 in French, 561 in Hindi, 2, 190 in Italian, and 612 in Bengali. The test data was a balanced set of 100 texts sampled randomly from both M/S and non-M/S groups in each language, for a total of 500 examples of M/S speech and 500 examples of non-M/S speech.
References of the datasets are:
Bhattacharya, S., et al.: Developing a multilingual annotated corpus of misog- yny and aggression, pp. 158–168. ELRA, Marseille, France, May 2020. https:// aclanthology.org/2020.trac- 1.25
Chiril, P., Moriceau, V., Benamara, F., Mari, A., Origgi, G., Coulomb-Gully, M.: An annotated corpus for sexism detection in French tweets. In: Proceedings of LREC, pp. 1397–1403 (2020)
Fersini, E., et al.: SemEval-2022 task 5: multimedia automatic misogyny identification. In: Proceedings of SemEval, pp. 533–549 (2022)
Fersini, E., Nozza, D., Rosso, P.: Overview of the Evalita 2018 task on automatic misogyny identification (AMI). EVALITA Eval. NLP Speech Tools Italian 12, 59 (2018)
Guest, E., Vidgen, B., Mittos, A., Sastry, N., Tyson, G., Margetts, H.: An expert annotated dataset for the detection of online misogyny. In: Proceedings of EACL, pp. 1336–1350 (2021)
Jha, A., Mamidi, R.: When does a compliment become sexist? Analysis and classification of ambivalent sexism using Twitter data. In: Proceedings of NLP+CSS, pp. 7–16 (2017)
Waseem, Z., Hovy, D.: Hateful symbols or hateful people? Predictive features for hate speech detection on Twitter. In: Proceedings of NAACL SRW, pp. 88–93 (2016)
Please see the paper for more detail.
license: mit tags: - generated_from_trainer metrics: - accuracy - f1 - precision - recall model-index: - name: xlm-roberta-base-misogyny-sexism-indomain-mix-bal results: []
xlm-roberta-base-misogyny-sexism-indomain-mix-bal
This model is a fine-tuned version of xlm-roberta-base on the None dataset. It achieves the following results on the evaluation set:
- Loss: 0.8259
- Accuracy: 0.826
- F1: 0.8333
- Precision: 0.7996
- Recall: 0.87
- Mae: 0.174
- Tn: 391
- Fp: 109
- Fn: 65
- Tp: 435
Model description
More information needed
Intended uses & limitations
More information needed
Training and evaluation data
More information needed
Training procedure
Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 2e-05
- train_batch_size: 16
- eval_batch_size: 16
- seed: 42
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- num_epochs: 2
Training results
| Training Loss | Epoch | Step | Validation Loss | Accuracy | F1 | Precision | Recall | Mae | Tn | Fp | Fn | Tp |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 0.2643 | 1.0 | 1603 | 0.6511 | 0.82 | 0.8269 | 0.7963 | 0.86 | 0.18 | 390 | 110 | 70 | 430 |
| 0.2004 | 2.0 | 3206 | 0.8259 | 0.826 | 0.8333 | 0.7996 | 0.87 | 0.174 | 391 | 109 | 65 | 435 |
Framework versions
- Transformers 4.20.1
- Pytorch 1.12.0+cu102
- Datasets 2.3.2
- Tokenizers 0.12.1