The efficiency of multitask deep learning methods for combined neural network models is comprehensively studied in application to a selected set of tasks: generating headers, defining lemmas, and keywords. The multitask model is built using Multi-head Attention layers and is used to develop models for generating headers and a model based on LSTM layers for lemmatization. Open corpuses RIA Novosti, containing news texts and headings for them, and a corpus with morphological, syntactic markup and lemmas for word forms SynTagRus from the universal dependencies project are used. For the task of highlighting keywords, we have assembled a new corpus consisting of news texts using the crowdsourcing platform. The results of the work show an increase in the accuracy by 1 % in the F1 score for the lemmatization problem when using features from the multitask model compared to using only morphological features, state-of- art accuracy (0.42 ROUGE F1 score) is achieved for the title generation task. An algorithm for highlighting keywords without additional network training is proposed on the basis of the model for generating headings obtained in this work.