A word sketch is a summary of collocates and other lexical items commonly found in the surroundings of a given word. The Sketch Engine uses it to display a one-page view of the word’s grammatical and collocational behavior. Sketch grammar is a set of rules written in CQL (Corpus Query Language) which defines the grammatical relations required to capture the associated collocate words in a text corpus.

The main goal of this thesis is to identify the issues of the current Sketch Grammar of Japanese in the Sketch Engine and implement an improved Sketch Grammar with solutions to these problems. The secondary goal is to prepare a training corpus from the Aozora public library and use it together with the new Sketch Grammar to train the mT5 language model at the task of extracting word collocates. Finally, there will be an evaluation and comparison of both the rule-based and artificial intelligence approaches.