OBJECT-ORIENTED SYSTEM FOR NEURAL NETWORK DETECTION OF HATE SPEECH USING CLOUD TECHNOLOGIES

Abstract

The article presents the results of the development and experimental study of an object-oriented neural network system for hate speech detection using cloud technologies. A method for neural network detection of hate speech is proposed, which involves two-stage processing: training a stable neural network model by modularly introducing noise into the training data and further using this model for inference in a cloud environment. Introducing noise allows you to simulate typical distortions characteristic of social platforms (spelling variations, symbolic substitutions, partial masking), which increases the stability of the classifier to real text conditions. The system architecture is implemented on the basis of the TextIndexDataset, BatchNoisyCollator and TemperatureScaler modules, which are responsible for data encapsulation, the formation of batches with distortions and the calibration of probabilistic forecasts, respectively. Cloud deployment ensures scalability of calculations, centralized storage of models and parameters, as well as repeatability of experiments.

Experimental studies were conducted on the datasets “Hate Speech Detection curated Dataset” (for training) and “Hate Speech and Offensive Language Detection” (for external validation). The obtained results prove that training models in mixed mode (clean and noisy examples) provides better generalization: on the internal test, models without noise show a higher F1-measure, however, on the external dataset, the advantage of models trained with distortions is 1.5–1.7%. This confirms the effectiveness of modular noise injection to increase the robustness of models and reduce the effect of overfitting to the training corpus.

The proposed approach combines the principles of object-oriented design, cloud computing and deep learning, which makes it suitable for scalable content monitoring and moderation systems. Prospects for further research are to expand the set of noise reduction strategies, improve the calibration of predictions and verify the proposed solution on multilingual corpora and real message streams.