NiGuLa/Russian_Inappropriate_Messages

--- language: - ru tags: - toxic comments classification license: cc task_categories: - text-classification size_categories: - 100K<n<1M --- ## General concept The **'inappropriateness'** substance we tried to collect in the dataset and detect with the model **is NOT a substitution of toxicity**, it is rather a derivative of toxicity. So the model based on our dataset could serve as **an additional layer of inappropriateness filtering after toxicity and obscenity filtration**. You can detect the exact sensitive topic by using [this model](https://huggingface.co/Skoltech/russian-sensitive-topics). Generally, an inappropriate utterance is an utterance that has not obscene words or any kind of toxic intent but can still harm the reputation of the speaker. Find some samples for more intuition in the table below. Learn more about the concept of inappropriateness [in this article ](https://www.aclweb.org/anthology/2021.bsnlp-1.4/) presented at the workshop for Balto-Slavic NLP at the EACL-2021 conference. Please note that this article describes the first version of the dataset, while the model is trained on the extended version of the dataset open-sourced on our [GitHub](https://github.com/skoltech-nlp/inappropriate-sensitive-topics/blob/main/Version2/appropriateness/Appropriateness.csv) or on [kaggle](https://www.kaggle.com/nigula/russianinappropriatemessages). The properties of the dataset are the same as the one described in the article, the only difference is the size. | Text | Inappropriate | Comment | Unsafe topic | |------------------------------------------------------------------------------------------------------|---------------|----------------------------------------------------------------|---------------| | Ладно бы видного деятеля завалили а тут какого то ноунейм нигру преступника | yes | Оправдание убийства | offline crime | | Это нарушение УКРФ! | no | Осуждение нарушения закона | offline crime | | Религия - это насилие высшего уровня. Верующего принуждают к вере еще в младенчестве. | yes | Оскорбление чувств верующих | religion | | Христиане, как вы относитесь к иноверцам, в частности к Гамалиилу и к уверовавшим в Иисуса фарисеям? | no | Обычные вопросы к верующим без намека на оскорбление их чувств | religion | | Хорошо что проституция запрещена | no | Осуждение простуитуции | prostitution | | Возьму бабки и сниму шлюх на выходных | yes | Популяризация проституции | prostitution | ## Licensing Information [Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License][cc-by-nc-sa]. [![CC BY-NC-SA 4.0][cc-by-nc-sa-image]][cc-by-nc-sa] [cc-by-nc-sa]: http://creativecommons.org/licenses/by-nc-sa/4.0/ [cc-by-nc-sa-image]: https://i.creativecommons.org/l/by-nc-sa/4.0/88x31.png ## Citation If you find this repository helpful, feel free to cite our publication: ``` @inproceedings{babakov-etal-2021-detecting, title = "Detecting Inappropriate Messages on Sensitive Topics that Could Harm a Company{'}s Reputation", author = "Babakov, Nikolay and Logacheva, Varvara and Kozlova, Olga and Semenov, Nikita and Panchenko, Alexander", booktitle = "Proceedings of the 8th Workshop on Balto-Slavic Natural Language Processing", month = apr, year = "2021", address = "Kiyv, Ukraine", publisher = "Association for Computational Linguistics", url = "https://www.aclweb.org/anthology/2021.bsnlp-1.4", pages = "26--36", abstract = "Not all topics are equally {``}flammable{''} in terms of toxicity: a calm discussion of turtles or fishing less often fuels inappropriate toxic dialogues than a discussion of politics or sexual minorities. We define a set of sensitive topics that can yield inappropriate and toxic messages and describe the methodology of collecting and labelling a dataset for appropriateness. While toxicity in user-generated data is well-studied, we aim at defining a more fine-grained notion of inappropriateness. The core of inappropriateness is that it can harm the reputation of a speaker. This is different from toxicity in two respects: (i) inappropriateness is topic-related, and (ii) inappropriate message is not toxic but still unacceptable. We collect and release two datasets for Russian: a topic-labelled dataset and an appropriateness-labelled dataset. We also release pre-trained classification models trained on this data.", } ``` ## Contacts If you have any questions please contact [Nikolay](mailto:[e-mail])--- 语言： - 俄语 标签： - 恶意评论分类（toxic comments classification） 许可协议：cc 任务类别： - 文本分类（text-classification） 数据规模： - 10万<n<100万 --- ## 核心概念 我们在本数据集中收集并拟通过模型检测的**‘不当性（inappropriateness）’**并非“恶意性（toxicity）”的替代概念，而是恶意性的衍生范畴。基于本数据集训练的模型，可作为恶意内容与低俗内容过滤后的**额外不当性过滤层级**。您可通过[该模型](https://huggingface.co/Skoltech/russian-sensitive-topics)识别具体的敏感话题。 一般而言，不当言论指不含低俗词汇或任何恶意意图，但仍可能损害发言者声誉的言论。您可通过下表的示例获得更直观的理解；如需深入了解不当性的概念，可参阅[EACL-2021会议波罗的海-斯拉夫语自然语言处理研讨会](https://www.aclweb.org/anthology/2021.bsnlp-1.4/)上发表的相关论文。 请注意，该论文描述的是本数据集的第一版，而模型训练所用的是扩展版数据集，该扩展版已在我们的[GitHub仓库](https://github.com/skoltech-nlp/inappropriate-sensitive-topics/blob/main/Version2/appropriateness/Appropriateness.csv)或[Kaggle平台](https://www.kaggle.com/nigula/russianinappropriatemessages)上开源。本数据集的属性与论文中描述的版本完全一致，唯一差异在于数据规模。 | 文本 | 是否不当 | 备注 | 敏感话题类别 | |------------------------------------------------------------------------------------------------------|---------------|----------------------------------------------------------------|---------------| | 就算把知名人士拉下马也就算了，可这次却是个无名小卒的黑人罪犯 | 是 | 为谋杀辩解 | 线下犯罪 | | 这是违反俄罗斯联邦刑法的行为！ | 否 | 对违法行为的合法谴责 | 线下犯罪 | | 宗教是最高级别的暴力。信徒在婴儿时期就被强迫信仰宗教。 | 是 | 伤害信徒的宗教情感 | 宗教 | | 基督徒们，你们如何看待其他信仰者，尤其是迦玛列和信奉耶稣的法利赛人？ | 否 | 向信徒提出的普通问题，未暗示伤害其宗教情感 | 宗教 | | 幸好卖淫是被禁止的 | 否 | 对卖淫行为的谴责 | 卖淫 | | 周末我要找些姑娘，找个妓女玩玩 | 是 | 宣扬卖淫行为 | 卖淫 | ## 许可信息 [知识共享署名-非商业性使用-相同方式共享4.0国际许可协议（Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License）][cc-by-nc-sa]. [![CC BY-NC-SA 4.0][cc-by-nc-sa-image]][cc-by-nc-sa] [cc-by-nc-sa]: http://creativecommons.org/licenses/by-nc-sa/4.0/ [cc-by-nc-sa-image]: https://i.creativecommons.org/l/by-nc-sa/4.0/88x31.png ## 引用信息 若您认为本仓库的内容对您的研究有所帮助，可引用我们的论文： @inproceedings{babakov-etal-2021-detecting, title = "Detecting Inappropriate Messages on Sensitive Topics that Could Harm a Company{'}s Reputation", author = "Babakov, Nikolay and Logacheva, Varvara and Kozlova, Olga and Semenov, Nikita and Panchenko, Alexander", booktitle = "Proceedings of the 8th Workshop on Balto-Slavic Natural Language Processing", month = apr, year = "2021", address = "Kiyv, Ukraine", publisher = "Association for Computational Linguistics", url = "https://www.aclweb.org/anthology/2021.bsnlp-1.4", pages = "26--36", abstract = "Not all topics are equally ``flammable`` in terms of toxicity: a calm discussion of turtles or fishing less often fuels inappropriate toxic dialogues than a discussion of politics or sexual minorities. We define a set of sensitive topics that can yield inappropriate and toxic messages and describe the methodology of collecting and labelling a dataset for appropriateness. While toxicity in user-generated data is well-studied, we aim at defining a more fine-grained notion of inappropriateness. The core of inappropriateness is that it can harm the reputation of a speaker. This is different from toxicity in two respects: (i) inappropriateness is topic-related, and (ii) inappropriate message is not toxic but still unacceptable. We collect and release two datasets for Russian: a topic-labelled dataset and an appropriateness-labelled dataset. We also release pre-trained classification models trained on this data.", } ## 联系方式 如有任何疑问，请联系[尼古拉（Nikolay）](mailto:[e-mail])

原始信息汇总

数据集概述

数据集名称

• 未明确提供数据集的具体名称。

数据集内容

• 该数据集专注于收集和检测“不适当”内容，这些内容并非直接的毒性或淫秽，而是毒性的衍生。
• 数据集用于训练模型，该模型作为毒性和淫秽过滤后的额外不适当内容过滤层。
• 数据集包含一系列敏感话题的不适当言论样本，如宗教、犯罪、性交易等。

数据集特点

• 数据集中的不适当言论不包含淫秽词汇或明显的毒性意图，但仍可能损害发言者的声誉。
• 数据集的扩展版本已开源，可在GitHub或Kaggle上获取。

数据集使用

• 用户可以通过使用特定的模型来检测具体的敏感话题。
• 数据集适用于文本分类任务。

数据集规模

• 数据集大小介于10万到100万之间。

语言

• 数据集主要使用俄语。

许可证

• 数据集遵循Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License。

引用信息

• 如需引用，请参考以下出版物：

  @inproceedings{babakov-etal-2021-detecting, title = "Detecting Inappropriate Messages on Sensitive Topics that Could Harm a Company{}s Reputation", author = "Babakov, Nikolay and Logacheva, Varvara and Kozlova, Olga and Semenov, Nikita and Panchenko, Alexander", booktitle = "Proceedings of the 8th Workshop on Balto-Slavic Natural Language Processing", month = apr, year = "2021", address = "Kiyv, Ukraine", publisher = "Association for Computational Linguistics", url = "https://www.aclweb.org/anthology/2021.bsnlp-1.4", pages = "26--36", abstract = "Not all topics are equally {``}flammable{} in terms of toxicity: a calm discussion of turtles or fishing less often fuels inappropriate toxic dialogues than a discussion of politics or sexual minorities. We define a set of sensitive topics that can yield inappropriate and toxic messages and describe the methodology of collecting and labelling a dataset for appropriateness. While toxicity in user-generated data is well-studied, we aim at defining a more fine-grained notion of inappropriateness. The core of inappropriateness is that it can harm the reputation of a speaker. This is different from toxicity in two respects: (i) inappropriateness is topic-related, and (ii) inappropriate message is not toxic but still unacceptable. We collect and release two datasets for Russian: a topic-labelled dataset and an appropriateness-labelled dataset. We also release pre-trained classification models trained on this data.", }

联系方式

• 如有疑问，请联系Nikolay。

搜集汇总

数据集介绍

构建方式

该数据集聚焦于俄语文本中的‘不当性’（inappropriateness）概念，其构建并非简单替代毒性检测，而是作为毒性及秽语过滤后的补充层。研究团队基于敏感话题（如离线犯罪、宗教、卖淫等）设计标注体系，通过人工标注区分文本是否具有不当性——即虽不含脏话或明显恶意，但可能损害发言者声誉的表述。数据集包含超过10万条样本，源自公开语料库及社交媒体，并经过多轮质量校验，确保标注一致性与话题覆盖的广度。

特点

数据集的核心特色在于对‘不当性’的精细化界定，区别于传统毒性分类。它强调话题相关性，例如同样提及犯罪，辩护杀人可能被视为不当，而谴责违法则否。样本附带敏感话题标签，支持多维度分析。此外，数据集规模适中（100K-1M），平衡了标注成本与模型训练需求，且与同名论文中描述的初版一致，仅扩展了数量，保证了研究可复现性。

使用方法

该数据集主要用于训练文本二分类模型，判断俄语消息是否具有不当性。用户可加载CSV格式数据，通过HuggingFace的transformers库微调预训练语言模型（如BERT）。推荐搭配敏感话题检测模型（如Skoltech/russian-sensitive-topics）使用，形成级联过滤管道：先识别敏感话题，再评估不当性。数据划分建议按80/10/10比例分割训练、验证与测试集，并采用宏平均F1分数评估性能。

背景与挑战

背景概述

在自然语言处理领域，毒性内容检测一直是研究热点，然而，许多对话中虽不包含显性脏话或恶意攻击，却仍可能对个人或机构声誉造成潜在损害。为填补这一研究空白，由Skoltech团队主导，Nikolay Babakov、Varvara Logacheva等研究人员于2021年在EACL-2021会议的波罗的-斯拉夫自然语言处理研讨会上提出了“不恰当性”概念，并发布了首个俄语不恰当消息数据集。该数据集的核心研究问题在于识别那些虽无毒性意图但能损害说话者声誉的言论，其影响力在于为内容安全过滤提供了更精细的层次，超越了传统毒性检测的边界。数据集创建后不断扩展，并开源在GitHub和Kaggle上，成为俄语敏感话题内容审核领域的重要基准资源。

当前挑战

该数据集面临的核心挑战源于其定义的“不恰当性”概念本身：不同于显性的毒性或冒犯性语言，不恰当性高度依赖话题上下文，例如对犯罪的轻描淡写或对宗教的隐晦贬低，均需模型理解微妙的社会规范与语义差异。在构建过程中，挑战尤为突出：首先，标注者需在缺乏明确脏话或攻击性词汇的情况下，依据社会常识判断言论是否“不恰当”，这导致标注一致性难以保证；其次，数据集需覆盖如犯罪、宗教、卖淫等多类敏感话题，每一话题下不恰当性的边界模糊且易受文化差异影响；最后，平衡数据集规模与标注质量也是一大难题，扩展版本虽增大了数据量，却可能引入更多噪声，增加了模型泛化的难度。

常用场景

经典使用场景

该数据集专为俄语文本中‘不当性’（inappropriateness）的分类任务而设计，其核心应用场景是在完成毒性（toxicity）与淫秽内容（obscenity）过滤之后，作为一道额外的检测层，用于识别那些虽不含粗俗词汇或明显恶意，却仍可能损害说话者或组织声誉的言论。研究者通常利用该数据集训练二分类模型，以区分对话中微妙的不当表达，尤其关注涉及敏感话题（如犯罪、宗教、卖淫等）的语境。

衍生相关工作

该数据集衍生了一系列经典工作，包括其发布团队基于扩展版数据集训练的俄语敏感话题检测模型（如Skoltech/russian-sensitive-topics），该模型可精准识别言论所涉的敏感领域。此外，该研究提出的‘不当性’定义框架被后续多语言毒性检测、对话安全分析等任务引用，例如在Balto-Slavic NLP研讨会上的相关论文中，该数据集常被用作对比实验的基准，推动了对非显性有害内容的建模研究。

数据集最近研究

最新研究方向

在当前自然语言处理领域，针对俄语社交媒体的不当言论检测正成为一项前沿课题，尤其是围绕敏感话题的细粒度识别。该数据集聚焦于“不当性”这一概念，它超越了传统毒性检测的范畴，强调那些虽不含脏话或明显恶意，却可能损害发言人声誉的言论。相关研究紧密关联企业声誉管理、内容审核自动化等热点事件，例如在政治、宗教、犯罪等敏感话题中，模型能精准区分“正常讨论”与“隐性不当”。该数据集的意义在于推动了从粗粒度毒性过滤向话题相关的不当性检测的范式转变，为构建更安全、更负责任的俄语在线交流环境提供了关键资源。