Automation of Process of Symbols and Abbreviations Replacement by Words

Speech synthesisers have requirements for the input text in the Russian language. Numerals, abbreviations, and symbols in text must be transformed into word form in order to ensure the high quality audio output. In this research, the authors, to solve the symbol replacement task, have split it into two natural language processing tasks and have used a different approach for each one. The first task requires the use of regular expressions and the development of algorithms in Python to replace symbols that have a single-word form. The second task requires a more sophisticated approach, as the symbols now have different syntactic forms, which vary by case, gender, and number; therefore, machine learning methods were used, where neural networks were trained using PyTorch, transformers, and public T5-based models. The authors collected sentences with symbols from Wikipedia articles and developed datasets to test the models. The paper shows accuracy, training, and inference time for ruT5-small, ruT5-base-multitask, ruT5-large, fredT5-large models. Also, the authors tested the performance of an ensemble of models to replace symbols and abbreviations.