--- license: mit task_categories: - text-classification language: - en pretty_name: VOICED size_categories: - 1K Offensive speech detection is a key component of content moderation. However, what is offensive can be highly subjective. This paper investigates how machine and human moderators disagree on what is offensive when it comes to real-world social web political discourse. We show that (1) there is extensive disagreement among the moderators (humans and machines); and (2) human and large-language-model classifiers are unable to predict how other human raters will respond, based on their political leanings. For (1), we conduct a ***noise audit*** at an unprecedented scale that combines both machine and human responses. For (2), we introduce a first-of-its-kind dataset of ***vicarious offense***. Our noise audit reveals that moderation outcomes vary wildly across different machine moderators. Our experiments with human moderators suggest that political leanings combined with sensitive issues affect both first-person and vicarious offense. ## Dataset Details The train/dev/test splits are individual parquet files and there is a combined file as well with `voiced_complete.parquet`. ### Dataset Description This is an overview of the columns on this dataset. 1. `annotator_id` = The unique ID of the annotator of annotated this item, annotators come from Amazon Mechanical Turk. This is a hashed ID to maintain annotator information. 2. `comment_id` = Unique comment ID from YouTube. This is hashed in accordance to YouTube data sharing policies. 3. `batch_id` = This includes the batch the data item was part of during the human experiments. 4. `dataset` = This the YouTube channel source of the dataset, it is a pick between CNN, MSNBC, and FOXNEWS. 5. `duration` = The duration annotator took to annotate the entire batch (30 items) 6. `dataset_bin` and `dataset_kind` = Both are co-dependent. Since we did sampling of `general` randomly picked YouTube comments, `gun` is data items that are related to gun laws, and `abortion` is comments that are related to abortion laws. 1/2 of the entire dataset is general, and 1/4 gun and abortion related each. 7. `comment_text` = Text of the YouTube comment 8. `PERSON_TOXIC_raw` = This is the raw label from the annotator, if it is personally offensive to them. 9. `PERSON_TOXIC` = The aggregated version, here if the label choice is `not at all offensive`, then the value is 0. Everything else is 1. 10. `[political party]_TOXIC`,`[political party]_TOXIC_raw`, and `[political party]` refers to vicarious labeling and is either DEM, REP, or IND. This should be inferred with `annotator_political` which is the political leaning the annotator provided. For example, if the annotator `annotator_political` is Republican, then `REP_TOXIC` will be 1000 (empty value) and `REP_TOXIC_raw` will be empty. Since vicarious labeling asks for a REP annotator to annotate for DEM and IND. So `DEM_TOXIC` and `IND_TOXIC` will have values (including their `raw` columns). 11. `online` and `social` are questions where we asked if this content is OK to be shown online and on social networks respectively. 12. `2016_election` and `2020_election` are questions where we asked if the those presidential elections were conducted in a fair and democratic manner. 13. `published_at` is the timestamp the YouTube comment was published. 14. The rest of the columns are demographic information of the annotators. ### Dataset Paper - **Paper:** [EMNLP Paper](https://aclanthology.org/2023.emnlp-main.713/) ## Citation **BibTeX:** ```` @inproceedings{weerasooriya-etal-2023-vicarious, title = "Vicarious Offense and Noise Audit of Offensive Speech Classifiers: Unifying Human and Machine Disagreement on What is Offensive", author = "Weerasooriya, Tharindu and Dutta, Sujan and Ranasinghe, Tharindu and Zampieri, Marcos and Homan, Christopher and KhudaBukhsh, Ashiqur", editor = "Bouamor, Houda and Pino, Juan and Bali, Kalika", booktitle = "Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing", month = dec, year = "2023", address = "Singapore", publisher = "Association for Computational Linguistics", url = "https://aclanthology.org/2023.emnlp-main.713", doi = "10.18653/v1/2023.emnlp-main.713", pages = "11648--11668", abstract = "Offensive speech detection is a key component of content moderation. However, what is offensive can be highly subjective. This paper investigates how machine and human moderators disagree on what is offensive when it comes to real-world social web political discourse. We show that (1) there is extensive disagreement among the moderators (humans and machines); and (2) human and large-language-model classifiers are unable to predict how other human raters will respond, based on their political leanings. For (1), we conduct a ***noise audit*** at an unprecedented scale that combines both machine and human responses. For (2), we introduce a first-of-its-kind dataset of ***vicarious offense***. Our noise audit reveals that moderation outcomes vary wildly across different machine moderators. Our experiments with human moderators suggest that political leanings combined with sensitive issues affect both first-person and vicarious offense. The dataset is available through https://github.com/Homan-Lab/voiced.", } ```` ## Dataset Card Contact Tharindu Cyril Weerasooriya [cyril@mail.rit.edu]