Usage of Large Language Models for Rules Induction based on the Examples from Text

The presented thesis addresses the possibility of using large language models for frequent pattern searches. Following the terms of reference of the thesis, which was to experimentally verify the ability of the LLM to induce general rules describing common characteristics of a given small number of examples, the thesis consists of three objectives. The main objective is to experiment with freely available language models (LLMs). The other objectives consist of a search of the procedures for the LLMs and an evaluation of the experiments, including a comparison with an existing approach for rule finding compared to Apriori and FP-Growth algorithms. In this thesis, experiments were conducted with three LLMs (Copilot, Gemini, and ChatGPT-3.5) in combination with eight prompt tasks focused on writing the frequent groups of items that occur in some versions of the small input dataset. In the original version, the dataset consisted of familiar column names and their values (names of the animals and their characteristics). In other versions, the actual information was masked and replaced with meaningless identifiers. The best results were obtained from the Copilot model, which scored up to 100% accuracy in the datasets with meaningless identifiers, but significantly lower than in the original data. A similar trend, where meaningless identifiers lead to better results could be spotted in the ChatGPT-3.5 model (if the result counts as the arithmetical mean of all eight experiments). It is questionable if this procedure would work on larger datasets and other domains.