Detection of Low-toxic Texts in Similar Sets Using a Modified XLM-RoBERTa Neural Network and Toxicity Confidence Parameters
Yaroslav Aleksandrovich Seliverstov, Andrew A. Komissarov, Eleonora D. Poslovskaia, Alina A. Lesovodskaya, Artur Podtikhov · 2021
The article considers the problem of classifying low-toxic texts using a modified neural network of the XLM-RoBERTa transformer architecture, trained on highly toxic texts. Comments from the School of Pedagogical Design at the University of 20.35 were used as kits for identifying low-toxic texts. The network was not retrained on low-toxic texts. Instead, the classification of low-toxic texts was carried out by varying the toxicity confidence parameter. An approximation dependence of the number of low-toxic texts on the parameter of toxicity reliability was constructed and a threshold value of the toxicity reliability parameter was obtained, at which the quality of the classification of low-toxic texts is maximal. The hypothesis of the similarity of the toxicity of homogeneous information resources was also formulated and confirmed.