Generation of Original Text with Text Mining and Deep Learning Methods for Turkish and Other Languages

Abstract

The amount of content on the web has increased dramatically since the Internet began providing users with the ability to produce content. Initial work on original text production has aimed at publishing the given data by putting in a certain mold. The most obvious example of this is the analysis reports on sporting events. However, preparing an original text compiled with general information about a subject has become a subject of interest to scientists as well. Although Neural Networks and Markov models were used previously for original text production, the original text generation process and comparison of the success rates weren't done using the Turkish language and the academic publication data repository dataset. In this study, it was tried to create summary information / original content about a specific subject by using Wikipedia TR for the Turkish language and the data pool created with hundreds of thousands of academic publications. In the study, texts were produced with Markov Model and LSTM, which were previously proposed, and the results are comparatively shared in detail. In the evaluation study, the performance of the proposed method was examined, and the correctness of the techniques was evaluated concerning syntactic accuracy and semantic preservation. The results are evaluated by presenting a mixture of original and machine-generated texts to the actual user for the success test of the proposed method. The success rate of the results is calculated with accuracy, recall, and f-measure. The results are very promising because it has been observed that the method can produce accurate and quality representations.