For many educational applications, such as learning resource recommendation systems, the textual difficulty of a text is a key information. Today, the best results for this task are obtained by using NLP and deep learning techniques. However, the use of these methods can result in the loss of statistical linguistic information that is important for determining text readability more accurately. In our work, we propose an approach for assessing text readability by combining neural network models with linguistic features extracted from the text and integrated into the model to improve the quality of the neural network models. Experimental results show that this combination improves system performance.