MTS AI, together with Skoltech, presented an original AI solution called “language detoxifier”. A special algorithm recognizes toxic words and expressions, and then either removes them and replaces them with words that are close in meaning, but not offensive. It can be used in social networks, as well as when communicating with voice assistants and chat bots.
The developers have created two types of AI models for creating bots and applications that clean up negative language. The first uses the BERT language model based on the Transformer neural network architecture. It edits the text pointwise, replacing negative words and expressions with neutral synonyms or deleting them. The second one performs conditional text generation on the input request, creating a neutral version of the toxic phrase.
“The Internet is a unique source of information, not only the largest and most accessible, but also spontaneously developing. Content is created by users themselves, so the problem of safe and non-toxic communication is always relevant. Disputes in the comments often develop into insults, users “on emotions” express their opinion in a rude form, because writing and sending a text is a matter of a few seconds.
The speed of content generation is so high that it is impossible to filter it manually. Often social networks simply block offensive language. Our solution allows not just deleting messages or banning users, but offering to replace the text with a more neutral one, without losing the essence of the message,” notes Irina Krotova, Senior Developer in the NLP MTS AI group.
According to the developers, the new algorithm can also be used to train chatbots so that they do not “pick up” unwanted vocabulary. Anyone can already test the capabilities of the language detoxifier with the help of a special bot in Telegram. You can learn more about the methods and models that were used to create the algorithm in an article prepared by MTS AI and Skoltech specialists, as well as at this link .
Alexander Panchenko, senior lecturer at Skoltech and head of the MTS-Skoltech joint laboratory, said that the proposed models can be used not only for text modification. For example, by recognizing obscene language, the bot can delay sending text to the chat and offer the user a less toxic wording of his message.