Hierarchical Topical Embeddings in Text Collections

This work discusses the method of creating hierarchical topical word embeddings of large collections of texts, that is, vector representations of words built on the basis of the topic model. In this work experiments have been conducted on the selection of hyperparameters of the algorithm, the application of the obtained vector representations to various problems of natural language processing. The method of constructing hierarchical word embeddings, which have not been considered by researchers, is described. As a result, obtained word embeddings showed high results on various applied tasks of natural language processing: text classification and word-similarity tasks. The pretrained word embeddings obtained in the course of this work, can be used by other researchers and commercial organizations to solve the problems of natural language processing. Keywords: topic modelling, NLP, word embeddings, ARTM.