Author-book recognition with natural language processing techniques

Since the discovery of the manuscript, humanity has produced many written document in various ways. Every text carries traces of its author. There are author's thesaurus, thinking and logic execution, wrong or incomplete informations, spelling habits on the text. From this point of view, we can say that a written document is the textual fingerprints of the people who write it. However, just like fingerprints, these features are difficult to detect from the text with human abilities. It is difficult to determine the personal characteristics of texts before the computer revolution, but these processes can be achived with computers today. For the author recognition process, various author features are determined by training text and then compared with the features of other texts to look for similarities. Some of the authors features on the text: thesaurus, typographical errors, character n-gram traces, word n-gram traces. Thanks to the computers, these types of properties are extracted from the text and analysis of the status of the writer's ownership of a document. In this thesis, author detection process was studied. 120 different Turkish books written by 20 Turkish authors in different distributions were studied. Character n-gram for stylometry and Naïve Bayes classifier is used to recognize the author of the text. First of all, we gather 120 Turkish novels and convert them to txt format. Then, all book were pre-processed and gaps deleted, erroneous characters corrected, alphanumeric characters, punctuation, and non-Turkish characters removed. After preliminary operations, the books were divided into two group as 20 training and 100 test book. The training books have the author name label and author stylometric features extracted from them. These stylometric features are separately bi-gram, tri-gram, quadri-gram. We calculate the frequency of these features and get 200 for author's stylometry vector space. After all, our system is ready for the author recognition process. For the text, We take the test books one by one and extract bi-gram, tri-gram, quadri-gram for authorship properties as we do for educational books. Following, Extracted most frequent 200 n-gram pass to naive bayes classifier. Naïve Bayes classifier decides who write the book among the authors previously introduced to the system. The system's estimation label and the real author of the book is compared and the results are noted at the end of work. The comparison of the n-grams performances attained by Naïve Bayesian method is examined through this thesis. The achievements of the n-gram vector spaces about author detection were observed in this study. As a result of our observations, bigram vector space is failed. Besides, trigram and quadri-gram gave good accuracy result. It should be noted the best performance with 82% accuracy belongs to the quadri-gram. The other results and confusion matrix is located in the thesis. With the recent developments of computer architectures, the amount of proceedings and articles about the understanding of human languages by the computers has enormously increased. Unfortunately, a large amount of these papers are related with English language. From this perspective, the thesis contributes the Natural Language Processing in Turkish language and provides motivation for the further studies on the field. Keywords : Text classification, Author detection, Naïve Bayesian approach, N-gram.