Comparative Analysis of Turkish Proverbs and Idioms Using Natural Language Processing-Based Direct and Semantic Matching Methods

This study examined the use of Turkish idioms and proverbs using datasets obtained from various digital environments. The study was conducted using three different datasets, with the idiom and proverb dictionary prepared by the Turkish Language Association serving as a reference. The dataset was created using Turkish news articles, Twitter data, and data from the Ekşi Sözlük website. Two different methods were used in the study: direct matching and semantic-based matching. Five different language models based on semantic similarity were used. The matching performance of proverbs and idioms was evaluated using the SBERT, LaBSE, USE, E5, and DistilBERT models. The results showed that idioms are more widely used in language than proverbs. Models with higher coverage yielded higher matching values, while precision values decreased; however, selective matching models achieved higher precision values. When the model performances were evaluated using the F1-score, DistilBERT demonstrated the most balanced, while the SBERT and E5 models stood out with their high coverage, and the LaBSE and USE models achieved higher precision values despite their lower recall values. The results provide an assessment of the way proverbs and idioms are delivered in different environments and the language models' perception of these elements.