Contents: Acquaintance with vector abstraction of words. Analysis of mathematical properties of vectors that can be used when working with embeddings. Output of CosineDistanceLoss.
Exploring word2vec, fasttest, glove. Comparison of language models and pre-trained embeddings in learning transfer. Mini-project: extract hashtags from comments.

Contents: Continued work with language models. Gather your own theme model with Gensim. Comparison of neural network prediction accuracy using embeddings pre-trained on different topics. Comparison of methods to work with texts written in languages other than English. Morphological text analysis using pymorphy2 for text pre-processing.

Contents: Consideration of the structure and meaning of popular recurrent layers: LSTM (and modifications), GRU. Comparison of operating modes of recurrent neural networks. Study of model operation at tensor level. The purpose of the lesson is to learn to understand how the data on a recursive neural network moves in keras, the necessary dimensions at input and output in different modes of neural network operation, for which each axis in input and output tensors is responsible. Theoretical part of the lesson in the future will speed up the creation of the project with its dataset and its architecture many times. Consideration of the Named Entity Recognition task. Mini-project: NER application on world news.

Contents: Exploring the second approach to text analysis - convolutional netowkrs: structure, why this approach works. Comparison of CNN and RNN based models.
Examination of NLP convolutional neural networks originally designed for computer vision. Object detection in texts. Mini-project: object detection in car reviews.

Contents: Exploring sequence-to-sequence transformation. Consideration of components of seq2seq model: encoder and decoder. Analysis of seq2seq tasks and ready implementations. Mini-project: creating your own chat-bot that answers questions.

Contents: The boundary between images and texts is finally erased. Extracting the message from the image and text with the subsequent transformation: search for the image by text, creation of the image by means of a generative network with the message of the text as a condition, creation of the text description by the image.

Contents: Studying the concept of "attention" in neural networks. Analysis of the principles of attention in seq2seq tasks. Comparison of the quality of neural networks with attention and without attention. Mini-project: creating your own translator.

Contents: Exploring an advanced attention mechanism called transformer. Building the basic version of the transformer. Examination of structure and possibilities of state-of-art implementations of transformer: BERT, GPT-2, T5. Mini-project: extraction of the gist: writing an abstract for a scientific article.

Contents: Analysis of the search for sentences with the necessary meaning in the text using recurrent neural networks. Mini-project: highlighting sentences in the news, which refer to a predetermined person.

Contents: Studying a scientific article and writing a neural network for it from scratch with a deep analysis of the theoretical part. Solving the problem of text generation using GAN. Text generation, successfully passing Turing test.
Parse the CNN+RNN ensemble of neural networks to solve the complex problems of обработки\генерации text. Customization of Keras' work with the help of low-level code on TensorFlow 2.0.