2025 , Volume 30, № 4, p.145-158

Purpose. This study addresses developing and evaluating neural network-based classification models for identifying sentences that are difficult to translate using machine translation systems. Methodology. We assemble and preprocess dataset sentences labeled by translation difficulty, apply tokenization and implement multiple neural network architectures for their classification. Three models are built: a simple recurrent network (A1) using SimpleRNN layers, a long short term memory network (A2), and a convolutional neural network (A3) with Conv1D layers. The models are trained and tested on the dataset using standard machine learning procedures, and their classification performance is evaluated using metrics such as accuracy and F1-score. Findings. The experimental results demonstrate that the LSTM-based architecture (A2) achieves the highest classification accuracy and F1-score among the proposed models, indicating its superior ability to capture complex features related to translation difficulty. All models yield satisfactory results, however clear differences in training dynamics and final performance metrics do occur. Detailed metric values for each architecture are reported, confirming the feasibility of using neural networks for this binary classification problem. Originality/value. A novel application of neural network classifiers to the problem of detecting translation-difficult sentences is presented. The developed dataset and models can improve pre-translation analysis and help optimize machine translation pipelines by flagging challenging inputs. The approach contributes to computational linguistics by exploring different neural architectures and offering a valuable resource for further study