The thesis will study how the transformer architecture will be able to learn hyphenation. The student will concentrate on the languages, where ground truth data (hyphenated Czechoslovak and German word lists) are available, together with existing effective near-to-optimal solutions by competing patterns generated by the program Patgen.

The student will find the minimal architecture of the transformer comparable to the current solution (minimal patterns generated by Patgen with optimizations suggested in Ondřej Sojka's thesis). He will design a series of architectural experiments and compare their precision, recall, generalization, effectiveness, and efficiency. Then, in the minimal architecture, he will describe how the cascade of context-dependent exceptions is handled in the transformer without memorization to achieve both high accuracy and generalization (measured by cross-validation).