Analysis and synthesis of machine learning models for solving the problem of searching for named entities

Problem setting. Building a dialog system is a voluminous and complex task that includes many classical tasks of natural language processing. One of the most important tasks is the Named Entity Recognition (NER), since entities are the hard ontologies for determining the meaning of the text. All entities can be divided into 3 groups: "flat" or regular entities, nested entities, and text-separated entities. When solving the NER problem, the difficulty arises with the construction of a universal model for the qualitative processing of all types of entities. In this paper different approaches for the recognition of each type of entities are studied and analyzed, benchmark solutions are identified and the structure of a hybrid model to handle all types of entities is proposed. Target. Analyzing and comparing the reference solutions for each type of entities. Synthesize different solutions based on the components of the reference models to solve the NER problem with a single model, compare and identify the best models in terms of speed and accuracy. Results. It was shown that each of the reference models for a particular type of entities does not give high quality in the processing of all types. Based on this, different architectures based on the components of the reference solutions were developed to find a universal model architecture for processing all types of entities. Through comparison and analysis, optimal architectures were identified, both in terms of speed and accuracy, capable of showing high quality performance for all types of entities. Practical significance. The developed model architectures allow to process all types of entities qualitatively with one universal model. The resulting models cover all types of entities and surpass the basic solutions in terms of average speed and accuracy.