Search

The approach presented in this study is about the calculation of the similarities among languages by using the new feature template to be obtained from the syntactic analysis phase. Studies were conducted on 6 different language sets from two different language families in order to show the calculability of similarity of languages with the help of the recommended new feature template. In the first study, triplet-pattern template which is obtained from the syntactic analysis of Turkey, Kazakh, and Uyghur Turkish languages from Turkic languages families belonging to the Ural-Altaic linguistic family, could be formed automatically through developed software, and also similarity analysis of the desired languages could be made thanks to a different module developed within the same software. Consequently, not only similar structural relations of the languages from the same language family but also structural differences among the languages can also be revealed. Even if the first aim is to determine the similarities among languages when developing an approach, the real aim of the desired structure is to form a system independent from the language. In order to show that the formed system has a structure independent from the language, another study was carried out. In the second study, the similarities among the languages were determined by using treebanks of English, Swedish and Norwegian from the Germen language family. When the language family is Turkic and the metrics are Jaccard, Dice, and Similarity Matching, the highest similarity is Turkish-Uyghur, and the values of the metrics are 25.21%, 40.27%, and 50.42%, respectively. When the language family is Germen, the highest similarity is Norwegian-Swedish, and the values of the metrics are 37.15%, 54.17%, and 74.3, respectively.

Mobile data collectors that can be used without a central control mechanism currently have common use in many fields. Because they do not need a central unit, each node in a network can move independently. The field literature offers various group- or entity-based models to define the functioning of mobile data collectors. In this study, a random entity mobility model (REMM) research was performed. The study was based on the models random walk (RW) and random waypoint (RWP), used in several former studies mentioned in the literature. Furthermore, the models random point (RP) and random journey (RJ) proposed by Bilgin [1] for two-dimensional (2D) space were transferred to three-dimensional (3D–cubic) to be used in the study. Study findings obtained by defining a various number of fixed nodes in areas of various sizes were analyzed using 4 different metrics. It was observed that 4 different metric values decreased for 4 REMMs when the cubic area was enlarged by increasing the edge lengths (150–200–250 pixel) of the cubic. When the cubic’s edge length is 150-200-250 pixel, respectively, connected node ratio (CNR) metric value is 98.04%–95.8%–91.34% for RP and 96.83%-83.23%-70% for RJ. Provided that the cubic area remains constant, the increases in the number of nodes generally tend to increase, although there are slight fluctuations on the results. When the cubic edge is 200 and the node numbers are 4–6–8–10, the message delay is 13.345–16.566–27.386–40.050 seconds for RW and 6.579–9.124–11.431–13.456 seconds for RWP. In the comparisons made by taking the average of the values obtained according to the size of the cubic area and the number of nodes, the RP model reached the highest values for all metrics. For example, the visited node ratio (VNR) metric average for the cubic edge 200 pixels is 98.76% for RP and 94.68%–87.38%–94.78% for RW–RWP–RJ. The VNR metric for the cubic edge 250 is 96.55%–93.7%–87.45%–51.27% for the RP–RW–RWP–RJ. Similarly, the average values obtained for other metrics prove this situation. In addition, when the results of the study are examined, it has been measured that the RP model can deliver the message to the base with less delay than other models. The average delay for the cubic edge 150 is 2.933–27.667–23.236–5.698 second for the RP–RW–RWP–RJ and 2.846–24.337–10.148-4.293 second when the edge is 200. When the average results obtained were examined, the success ranking in the delay metric was RP–RJ–RWP and RW, while the other metrics were formed as RP–RJ–RW–RWP. Considering all the obtained results, it was seen that the proposed two models achieved better results than the existing models in 3D after 2D.

In recent years, deep neural networks have been successful in both industry and academia, especially for computer vision tasks. Humans and animals learn much better when gradually presented in a meaningful order showing more concepts and complex samples rather than randomly presenting the information. The use of such training strategies in the context of artificial neural networks is called curriculum learning. In this study, a strategy was developed for curriculum learning. Using the CIFAR-10 and CIFAR-100 training sets, the last few layers of the pre-trained on ImageNet Xception model were trained to keep the training set knowledge in the model’s weight. Finally, a much smaller model was trained with the sample sorting methods presented using these difficulty levels. The findings obtained in this study show that the accuracy value generated when trained by the method we provided with the accuracy value trained with randomly mixed data was more than 1% for each epoch. Keywords: Curriculum learning, model distillation, deep learning, academia, neural networks.

Object detection applications started to take their places in our lives with the improvements in technology and artificial intelligence. In daily life, some applications such as license plate detection, optical character recognition become indispensable. In parallel with ongoing technological developments today, the technologies which soon will be a part of our daily life such as detection of suspicious situations through security cameras and autonomous cars, are improving rapidly. The studies in the object detection area are have generally focused on opaque objects. The number of studies on transparent objects is very limited. In this study, a system which is trained by the dataset that contains transparent and non-transparent glasses, is used to detect transparency of these glasses. Recently introduced capsule networks are used to develop a proposed system. To compare the obtained results, LeNet, AlexNet, and ResNet are trained and tested with the same dataset. When the result of the study is evaluated, it was seen that CapsNet got better results on classification accuracy than other deep learning methods which are used in this study. The obtained test accuracy, precision, recall, and F-score values for CapsNet are 91.58%, 91.61%, 91.58%, and 91.56%, respectively. The best obtained test accuracies for LeNet, AlexNet, and ResNet are 84.58%, 88.58%, and 87.58%, respectively. Thus, AlexNet got the second-best result on classification accuracy. While the ResNet took third place, LeNet stayed behind all of them.

Named entity recognition is a subject, inside of information retrieval which is a subdomain of natural processing. It pertains to identifying and labeling of location, person, organization, etc., inside of text content. Named entity recognition provides identifying and classifying of person, area, etc. inside of formal and informal text content and it can be used for different purposes as question answering systems and removal of the relation between events. In this work, named entity recognition is performed and one method is suggested and results are discussed for assignment to unlabeled name entities by using OpenNLP library with the help of KNIME program in the data set.

İnternetin günlük hayatımızdaki artan kullanımı ile beraber sosyal medya organlarının gelişimi de paralellik göstermektedir. Mikroblog adı verilen facebook ve twitter benzeri uygulamaları ile anlık duyguları ve düşünceleri ifade etmek son derece yaygın bir hale gelmiştir. Mikroblog sitelerinin en yaygın kullanıma sahip olanlarından birisi de Twitter uygulamasıdır. Twitter üzerinden paylaşılan mesajlar bir ürün ya da hizmet hakkında olabileceği gibi bir kişiyle ilgili bir yorumda olabilmektedir. Yapılan yorumun belirtmek istediği anlamı ve duyguyu belirleyebilmek son dönemdeki gözde konulardan biridir. Bir ürün ya da hizmet hakkında yapılan binlerce yorumun tek tek okunup anlamlandırılması ve yorumlayanların fikirlerinin sınıflandırılması geleneksel yöntemlerde oldukça zaman ve emek alan bir alandır. Gerek makine öğrenmesi ve derin öğrenme algoritmalarındaki gelişmeler gerekse de bunları işleyip yorumlayacak bilgisayar sistemlerinin gelişimine parallel olarak milyonlarca veri üzerinde duygu sınıflandırılması mümkün hale gelmiştir. Gerçekleştirdiğimiz çalışmada Türkçe ve İngilizce tivitler üzerinde duygusal sınıflandırma çalışması gerçekleştirilmiştir. Döküman vektörleri (Doc2Vec) kullanılarak yapılan çalışmada hem DBoW ve DM gibi iki farklı döküman vektörü yönteminin çalışması hemde Yarı Danışmanlı ve Danışmanlı öğrenmenin etkileri araştırılmıştır. Çalışma sonuçları doğruluk, kesinlik, anma, özgünlük ve F-ölçütü metrikleri ile raporlanmıştır. Gerçekleştirilen çalışma sonucunda Yarı Danışmanlı öğrenme yöntemi hem Türkçe hemde İngilizce veri kümesinde Danışmanlı öğrenmeye göre daha başarılı sonuçlar elde etmiştir.

Bu çalışmada duygu analizi için oluşturulan iki farklı veri kümesi, kelime vektörü algoritması olan Word2Vec ile modellenmiştir. Model oluşturulurken Word2Vec’in iki farklı yöntemi olan CBoW (Continous Bag of Words) ve Skip-Gram kullanılmıştır. Word2Vec ile bir metnin modelini oluşturmak için genellikle Ortalama yöntemi kullanılmaktadır. Bu çalışmada hem CBoW hem de Skip-Gram yöntemleriyle bir metni modellemek için üç farklı yöntem önerilmiştir. Model oluşturma (eğitim zamanı) süreleri her ikisi içinde ölçülmüştür. Sonuç olarak modelleme süresi açısından CBoW’un Skip-Gram’dan daha başarılı olduğu deneysel olarak gösterilmiştir.

The development of information and communication technologies (ICT) has led to many innovative technologies. The integration of technologies such as the internet of things (IoT), cloud computing, and machine learning concepts have given rise to Industry 4.0. Fog and edge computing have stepped in to fill the areas where cloud computing is inadequate to ensure these systems work quickly and efficiently. The number of connected devices has brought about cybersecurity issues. This study reviewed the current literature regarding edge/fog-based cybersecurity in IoT to display the current state.

In this study, the problem of vehicle routing, which arises from supplying a heterogeneous fleet of vehicles with different capacities and features, is discussed. First, a two-level method has been developed to address the transport-carrying assignment and then the route approach. In the first stage of the proposed method, the transported persons are matched with the transported vehicles considering their suitability. In the second stage, the routing process is completed by using the simulated annealing algorithm, which is a metaheuristic method. The novelty aspect of this study is the creation of a solution model by simulating the scheduling model of parallel machines. The implementation of the study was carried out on the distribution plans obtained from a large-scale logistics company performing cost analysis over the total distance. The proposed solution model is compared with the solution method used by the firm and it is concluded that the total distance and the resulting costs are reduced.