ИСТИНА

Relatively small labelled text arrays can be used as training samples for contemporary neural network language models. A content analysis methodology thus can be scaled to assess large collections of documents. A multi-criteria approach is proposed to control the quality of an automatic markup. Whether the model reaches an expert level of quality is estimated for the case when each text is labelled by a few of experts. The applications of the method include "understanding the meaning" of school writings in Russian, literature, social studies, history and English studies (UpGreat PRO // READING contest (https://ai.upgreat.one /) as well as detecting human values and propaganda techniques in social media.