Determining Of Semantically Close Texts Of Stock Market News Using Natural Language Processing

The paper considers the problem of finding seman-tically close texts in the domain of short stock market news. Two approaches to perform semantic search are proposed: with and without context awareness. The contextualized approach uses whole sentences in feature space mapping based on universal sentence encoder. The methodology of cluster analysis with determining optimal number of clusters is presented as well as construction of the classifier. The semantic search is illustrated with two experiments: for the complete dataset and within specific cluster. Another approach deals with isolated words of the sentence ignoring context and computes the vector representation of the whole sentence using classes of its words. The semantic search reduces to finding nearest neighbors in feature space using Euclidean distance as a measure of similarity. The experiments are performed with search keys of different length. The exper-iments illustrate that results obtained with contextualized and context-less approaches are of different thematic coverage.