--- annotations_creators: - expert-generated language_creators: - found language: - th license: - cc-by-nc-3.0 multilinguality: - monolingual size_categories: - 1K, for anonymity to prevent individual bias. #### Who are the source language producers? Twitter users in Thailand ### Annotations #### Annotation process We manually annotated our dataset with two labels: Toxic and Non-Toxic. We define a message as toxic if it indicates any harmful, damage, or negative intent based on our definition of toxicity. Furthermore, all the tweets were annotated by three annotators to identify toxicity; the conditions used for this identification are presented in the following list. - A toxic message is a message that should be deleted or not be allowed in public. - A message’s target or consequence must exist. It can either be an individual or a generalized group based on a commonality such as religion or ethnicity, or an entire community. - Self-complain is not considered toxic, because it is not harmful to anyone. However, if self-complain is intended to indicate something bad, it will be considered as toxic. - Both direct and indirect messages including those with sarcasm are taken into consideration. We strictly instructed all the annotators about these concepts and asked them to perform a small test to ensure they understood these conditions. The annotation process was divided into two rounds. We asked the candidates to annotate their answers in the first round to learn our annotation standard. Then, we asked them to annotate a different dataset and selected the ones who obtained a full-score for the second round as an annotator. From among these annotators, 20% of the annotators failed the first round and were not involved in the final annotation. #### Who are the annotators? Three annotators hired by [Sirihattasak et al (2019)](https://www.ta-cos.org/sites/ta-cos.org/files/1_W32.pdf) ### Personal and Sensitive Information Despite all tweets being public, due to the nature of toxic tweets, there might be personal attacks and toxic language used. ## Considerations for Using the Data ### Social Impact of Dataset - toxic social media message classification dataset ### Discussion of Biases - Users are masked before annotation by the annotators to prevent biases based on tweet authors ### Other Known Limitations - The data is included into [huggingface/datasets](https://www.github.com/huggingface/datasets) in Dec 2020. By this time, 506 of the tweets are not available publicly anymore. We denote these by `TWEET_NOT_FOUND` in `tweet_text`. ## Additional Information ### Dataset Curators [Sirihattasak et al (2019)](https://www.ta-cos.org/sites/ta-cos.org/files/1_W32.pdf) ### Licensing Information CC-BY-NC 3.0 ### Citation Information Please cite the following if you make use of the dataset: ``` @article{sirihattasak2019annotation, title={Annotation and Classification of Toxicity for Thai Twitter}, author={Sirihattasak, Sugan and Komachi, Mamoru and Ishikawa, Hiroshi}, year={2019} } ``` ### Contributions Thanks to [@cstorm125](https://github.com/cstorm125) for adding this dataset.