USAGE OF A GRAPHICS PROCESSOR TO ACCELERATE COREFERENCE RESOLUTION WHILE USING THE RoBERTa MODEL

Abstract

"The problem of finding coreferential objects (coreference resolution) in Ukrainian-language texts is considered, examples of co-referentiality in the Ukrainian language are given. Coreference in texts means the relationship between syntactic units that point to the same object (referent) in a given context. The differences between English-language and Ukrainian-language texts in solving the problem of searching for coreferenced objects and the difficulties that arise in finding coreferenced objects are noted. In particular, there are differences in the order of words in sentences: strict word order in English and arbitrary word order in Ukrainian. It is also pointed out the need to take into account the relationship between objects that are not part of the co-reference groups (predicates and secondary parts of sentence). Created an application that searches for reference objects in Ukrainian-language texts using the Transformers library and the RoBERTa language model. The use of a graphics processor to speed up the search for referential objects by the algorithm is considered. The work of the algorithm was tested on a set containing 2500 texts. 21.3 times acceleration reached due to the use of a graphics processor (GPU) in comparison to the performance of CPU. The resulting acceleration has greatly facilitated the development and testing of the application."