MTS AI and Skoltech: Language detoxifier

Main articles:

• Speech Recognition (Technology, Market)
• Speech technology: from recognition to understanding

2022: Detoxifying Agent Presentation

On March 22, 2022, MTS AI announced that, together with Skoltech, it had created a language detoxifier - an AI solution that recognizes toxic words and expressions, and then removes them or replaces them with near-meaning, but not offensive vocabulary. This solution allows you to maintain the meaning of the phrase and reduce the degree of aggression in communication. Detoxification can be used not only on social networks, but also in communication with voice assistants, chat and troop bots.

Examples of substitution of vocabulary with detoxifying agent

"The Internet is a source of information, not only quite large and accessible, but also spontaneously developing. Content is created by users themselves, so the problem of safe and non-toxic communication is always relevant. Controversy in the comments often escalates into insults, users "on emotions" express their opinion in a rude form, because writing and sending a text is a matter of seconds. The speed of content generation is so high that it is impossible to qualitatively filter it manually. Often, social networks simply block offensive statements. Our solution makes it possible not just to delete messages or banter users, but to propose replacing the text with a more neutral one, while not losing the essence of the message, "- said Irina Krotova, senior developer in the NLP MTS AI group.

The development of MTS AI and Skoltech is relevant for the Russian market, since most of these solutions are intended for the English language. There are practically no such services for Russian-speaking users, and the solutions developed earlier, according to the company, were ineffective.

MTS AI and Skoltech proposed two types of models for creating bots and applications that remove negative vocabulary from speech. The first approach uses the BERT language model, based on the Transformer neural network architecture. The model makes point editing of the text - finds negative words and expressions in it, replaces them with neutral synonyms or deletes them altogether.

The second approach is also based on the Transformer architecture, but it solves another problem - conditional generation of text on an input request. That is, the language model creates a neutral version of the toxic phrase. For the academic competition, a detoxification model based on the language model ruT5 was prepared.

"As part of this project, together with colleagues from MTS AI, it was possible to create a number of technologies for detecting and rewriting toxic content. The proposed methods and models can be used to prevent reputational risks of the company (a chatbot trained on texts from the network can respond toxic). Other applications are possible. For example, before a comment is sent, the user may be offered a less toxic wording of his message. In a similar use case, freedom of expression does not suffer, but the number of emotionally written negative comments can be significantly reduced. As a result of this collaboration, in addition to methods, models and data sets, many scientific articles appeared published by a joint team of researchers, engineers and students from Skoltech and MTS, "- said Alexander Panchenko, PhD, senior lecturer at Skoltech and head of the MTS-Skoltech joint laboratory.

You can test the capabilities of the language detoxifier yourself. For example, you can simply scold the bot in. messenger Telegram More information about the methods and models that were used in this approach can be found in the article "Methods of detoxification of texts for Russian the language," prepared by MTS AI and Skoltech specialists, as well as on the page of the joint laboratory "MTS-Skoltech" in the field of AI.