Computational diachronic analysis of Turkish synonym words

Turkish has undergone significant structural changes from the 1920s to the present day, influenced by the Turkish Language Reform. In line with the aims of the Language Reform, steps were taken to reduce the influence of foreign languages on Turkish. As a result of these interventions, many new words were introduced into the language, while many words fell out of use over time. Beyond changes in vocabulary, modifications were also observed in the spelling and grammar rules of the language. This study takes the rapid evolution of Turkish as its motivation and conducts a diachronic computational analysis of synonym words. First, we created Turkronicles, the largest diachronic corpus of modern Turkish, covering the period from the 1920s to the 2020s, and performed frequency analyses on this corpus. Additionally, we made various resources available for diachronic analyses, such as n- grams, diachronic word vectors, a software library, and a digitized dictionary of synonym words. Our analyses revealed that Turkish vocabulary has significantly changed over the years. To observe changes in spelling rules, we specifically examined words ending in -b/-p and -d/-t and encountered similar results. Overall, our analyses demonstrated the suitability of the resources we created for diachronic analysis and highlighted the extent of evolution of Turkish. Using these resources,x we proposed two different methods to identify synonym words used in different time periods. These methods are essentially based on finding an orthogonal transformation matrix. We then compared the proposed methods to a baseline method, which relies on a linear transformation matrix. The effectiveness of our methods was tested using two different vector algorithms. Furthermore, we observed and evaluated how the performance of the methods varied with increasing temporal differences. Our experimental results showed that the proposed methods outperformed the baseline method and were more robust to temporal differences. Finally, we conducted various experiments to explore how the meanings of synonymous words change over time. Specifically, we examined the dynamics of semantic specialization and generalization of synonymous words. Initially, we trained an encoder-decoder model to correct misspelled words in the Turkronicles corpus and created a test set. After correcting the misspelled words, we modeled semantic specialization and generalization using a word association graph and a centrality measure. To observe how this measure changes over time for a pair of synonyms, we calculated the Spearman correlation coefficient. Lastly, we trained a Linear Mixed Model to measure the extent to which the changes in one synonym affect the other. Based on the experimental results, we found that when the meaning of one word generalizes, the meaning of its synonym tends to specializes.