Detection of Toxic Content in Russian Texts

An aggressive behavior of users towards each other is a serious problem of many websites. It is impossible to manually track every message to prevent users from online harassment. Therefore, the automation of this process became the main goal of this work. The main steps of this paper were to build a Russian text classifier of toxic content and publish the solution to the open source. Deep learning methods were applied using Tensorflow library. FastText, multilingual Universal Sentence Encoder, Wiki40B and Navec models were used to get word embeddings. In order to classify the text, convolutional and recurrent neural network architectures were used. The best model achieved an accuracy of 91.43% on the test sample. The model with the best quality-to-memory ratio was published as a python package. Everybody can install it using the following command «pip3 install toxicity». The main aim was to make the product as easy to use as possible. To start using the package, it is not necessary to have knowledge in field of deep learning. It is enough to import the class constructor, create the instance, and call its method “predict”.