AI Was Taught to Swear for a Good Cause—and Became More Concise
Russian researchers found a link between profanity and concise expression and applied it to training neural networks.
Researchers from the AIRI Institute, the Higher School of Economics, and Innopolis University have found that artificial intelligence models can use features of obscene language to represent meaning more compactly.
The study showed that in natural speech, a single coarse word can often replace an entire sentence, conveying its meaning briefly and with high intensity. This kind of “semantic density,” the researchers argue, can be leveraged to optimize how large language models process text.
Short and to the Point
According to the authors, the goal of the project was not to teach AI to swear, but to give neural networks a tool for saving computational resources by reducing text volume without a noticeable loss of information.
During the experiments, the researchers tested how models handle Russian-language texts of varying length, from short phrases to full-scale news articles. They found that using the semantic function of obscene vocabulary helps compress content efficiently.
Semantic Density
They emphasize that in practical scenarios, the working version of the model does not directly use coarse expressions when processing text. Instead, it benefits from analyzing the contrast between neutral language and more “dense” forms of expression.
This approach is becoming increasingly relevant as data volumes continue to grow faster than available computing power. Using more concise linguistic constructions can reduce the load on information-processing systems and speed up algorithms without sacrificing depth of understanding.
