Theses supervised by Dr. Öğr. Üyesi Serdar Arslan
11 theses · Alanya Alaaddin Keykubat University, Çankaya University
Kaderciliğin örgütsel sessizlik ve işten ayrılma niyeti üzerine etkisi: Konaklama işletmeleri üzerine bir araştırma
Bu araştırma, kadercilik inancının örgütsel sessizlik ve işten ayrılma niyeti üzerindeki etkisini incelemeyi amaçlamaktadır. Özellikle hizmet sektörünün önemli bir alanını oluşturan otel işletmelerinde, çalışanların kaderci eğilimlerinin işyeri içindeki tutum ve davranışlarına yansımaları ele alınmıştır. Kadercilik, bireylerin yaşadıkları olayları dışsal güçlere atfetmeleri ve bu durumları değiştirme konusunda sınırlı bir inanca sahip olmaları şeklinde tanımlanır. Literatürde bu tür inançların, çalışanların örgütsel ortamdaki tutumlarını etkileyebileceği öne sürülmektedir. Çalışmanın örneklemini Alanya'daki otel işletmelerinde görev yapan 409 çalışan oluşturmaktadır. Veriler, yapılandırılmış anket formu aracılığıyla nicel yöntemle toplanmış; analizlerde açımlayıcı faktör analizi, korelasyon ve regresyon teknikleri kullanılmıştır. Bulgular, kaderciliğin örgütsel sessizliğin alt boyutlarından biri olan kabullenici sessizlik ile istatistiksel olarak anlamlı ve orta düzeyde güçlü bir ilişki içinde olduğunu ortaya koymuştur. Ayrıca korunma amaçlı sessizlik üzerinde de anlamlı etkiler saptanmıştır. İşten ayrılma niyeti ile kadercilik arasında ise pozitif yönlü, ancak düşük düzeyde bir ilişki belirlenmiştir. Sonuç olarak, kaderciliğin hem örgütsel sessizlik hem de işten ayrılma niyeti ile anlamlı ilişkiler kurduğu görülmüştür. Elde edilen bulguların, insan kaynakları yönetimi ve çalışan bağlılığı konularında politika geliştiricilere yol gösterici olabileceği düşünülmektedir.
Çatışma çözümü, müzakere süreci ve üçüncü taraflar: Uzlaştırmacı ve arabulucular üzerine bir inceleme
Bu çalışma, üçüncü tarafların çatışma çözüm ve müzakere süreçlerindeki deneyimlerini, üstlendikleri rolleri, geliştirdikleri çözüm stratejilerini ve süreci şekillendiren bağlamsal faktörleri anlamayı amaçlamaktadır. Nitel araştırma yöntemlerinden fenomenolojik desenin kullanıldığı bu çalışma, Alanya Adliyesi'ne bağlı uzlaştırmacı ve arabuluculardan oluşan 22 katılımcı ile gerçekleştirilen derinlemesine görüşmelere dayanmaktadır. Katılımcılar amaçlı örnekleme yöntemlerinden kartopu örnekleme yöntemi ile belirlenmiş, elde edilen veriler içerik analiz yöntemi ile incelenmiştir. Çalışma kapsamında, çatışma çözüm ve müzakere sürecine etki eden psikolojik, sosyal ve kültürel boyutlar ile üçüncü tarafların sahip olması gereken temel yetkinlikler detaylı bir biçimde ele alınmıştır. Kollektivist kültürel yapıya sahip toplumlarda çatışma çözüm ve müzakere sürecine etki eden sosyal eşitsizlik, uzlaşı kültürü, üçüncü taraf geleneği ve güven sorunları gibi dinamikler analiz edilerek, literatürde bu alandaki boşluğun giderilmesine katkı sunulmuştur. Bulgular; çatışma çözüm ve müzakere süreçlerinde üçüncü tarafların iletişimi kolaylaştırıcı, adaleti sağlayıcı, ilişkiyi onarıcı ve sürece yön verici roller üstlendiğini göstermekte, süreci etkileyen unsurlara dikkat çekmektedir. Ayrıca çalışma, çatışma ortamında liderlik pozisyonunda bulunan bireyler için duyguların yönetimi, güven inşası, ikna becerileri, empatik yaklaşım, iletişim, güç dengesi ve kültürel duyarlılık konularında yol gösterici veriler sunmaktadır.
A new fusion reranking pipeline for Turkish datasets using fine-tuned RAG components
This study addresses the gap in the multilingual capabilities of Retrieval Augmented Generation (RAG) systems for the Turkish language, particularly in the medical domain. With the rise of Large Language Models (LLMs) and their widespread applications, the reliance on external knowledge through retrieval components has become crucial to mitigate hallucinations and improve response accuracy. However, most existing retrieval components, including embeddings and rerankers, are predominantly trained on English datasets, highlighting a significant limitation in multilingual and domain-specific capabilities. To address this, the study introduced Pubmed-RAG-TR, a Turkish-language medical dataset, and fine-tuned retrieval components on both Pubmed-RAG-TR and WikiRAG-TR, a Turkish RAG dataset. A novel RRF-based reranker pipeline was also developed to improve the context construction for LLMs. Experimental results demonstrated that fine-tuning retrieval components on domain-specific datasets significantly enhanced the retrieval and post-retrieval quality, improving the accuracy of LLM responses. The study concludes that incorporating domain-specific semantics into retrieval and reranking models can substantially boost the performance of RAG systems in multilingual contexts.
Multi-class ship image classification using a hybrid vit-resnet architecture with attention mechanisms and semantic gated fusion
In this thesis, a hybrid model based on Vision Transformer (ViT) and ResNetRS50 is developed for multi-class classification of ship images. While ViT extracts high-level semantic information, ResNetRS50 captures low- and mid-level spatial features; these two structures are integrated through attention mechanisms and a Gated Fusion layer. During training, advanced techniques such as MixUp and CutMix data augmentation, Focal Loss combined with knowledge distillation loss, the OneCycleLR scheduler, automatic mixed precision (AMP), and exponential moving average (EMA) of model weights are employed. Experiments conducted on a dataset consisting of eight ship classes demonstrate that the proposed architecture outperforms single-stream CNN and ViT models in terms of both accuracy and F1-score. The results indicate that hybrid architectures and attention-based fusion strategies provide an effective solution to the ship classification problem.
Designing, analysis, and evaluating OSINT software architectures through scenario-based methods: A QAW, ADD and ATAM approach
This paper proposes a novel software architecture designed specifically for Open-Source Intelligence (OSINT) systems, addressing critical challenges in performance, scalability, and security that arise from handling large volumes of diverse and dynamic open-source data. Leveraging established system-centric methodologies including the Quality Attribute Workshop (QAW), Attribute-Driven Design (ADD), and Architecture Trade-off Analysis Method (ATAM), this study systematically identifies and addresses architectural drivers, integrating them into an optimized OSINT architecture. A significant innovation of this architecture is the integration of an on-premise Large Language Model (LLM) within the data analysis layer, employing the Retrieval-Augmented Generation (RAG) architectural pattern to effectively process and analyse unstructured textual data. The architecture emphasizes scalability through distributed multi-process data acquisition systems, robust security protocols minimizing digital footprints, and advanced methods for semantic information retrieval and analysis. Empirical evaluations conducted through rigorous experiments involving multi-threaded and multi-process environments demonstrate the superior scalability, high availability, and fault tolerance of the proposed system. These results validate the hypothesis that distributed processing, combined with stringent access controls and on-premise data analysis solutions, substantially enhances operational performance and security. This research fills significant gaps in the existing literature by offering detailed conceptual and practical insights into the design and implementation of system-centric OSINT architectures, presenting a comprehensive framework applicable to cybersecurity, intelligence operations, and business intelligence sectors.
Zero-shot and few-shot named entity recognition in environmental sciences domain
Novel architectures in natural language processing enable to transfer knowledge of the model for specific tasks. For many downstream tasks, training the model from scratch has become unnecessary since transfer learning can be leveraged for such cases. This can be achieved by finetuning a pretrained Large Language Models (LLM). In this study, a lightweight version of BERT, DistilBERT which is pretrained to predict next sentence was fine-tuned to handle Named Entity Recognition, as one of the most important information extraction task in context of textual data. Transfer learning also enable to transfer knowledge of the model to unseen domains. In this context, we created a domain-specific dataset in the environmental sciences domain. Also, to recognize specific entities, custom NER labels for entities in environmental sciences domain have been defined. To evaluate transfer learning ability of the model, zero-shot, one-shot and ten-shots learning procedures have been conducted on created dataset. To improve transfer learning, we have pre-trained the model a generic Turkish dataset. Finally, artificially generated data that specific to environmental sciences domain have been combined with our created dataset to improve the prediction performance of the model in zero-shot and few-shot setups. In the study, pretraining the model with generic dataset and introducing artificially generated dataset evaluated individually and together. In addition, presence of semantically related entities in the dataset have been investigated and improvements in prediction performance regardless of shot number are seen. The evaluation of tests demonstrates promising results and enlightens improvements in terms of transfer learning.
Integrating hierarchical gated attention network and recurrent neural network for improved sentiment classification of diverse Turkish texts
Sentiment classification is a significant process of determining contexts and trends in texts. Due to recent advancements in natural language processing and deep learning, methods such as neural networks have gained much more importance in sentiment classification. This study specifically highlights embedding methods, attention networks and mechanisms, and Recurrent Neural Networks (RNNs). The study investigates how embeddings of character and word improve models and the analysis of text, how attention mechanisms lay stress on and give more attention to certain parts of information, and how Long Short-Term Memory (LSTM) networks capture long-term dependencies and temporal changes in the text rather effectively. In order to analyze the sentiments of both short, unstructured and long, structured Turkish texts in a single model, a hybrid model is developed by integrating LSTM and Hierarchical Gated Attention Network. Furthermore, to evaluate the performance of the proposed model a new Turkish dataset has been created and labeled using Twitter data. Hierarchical Gated Attention Network plays a dual role by prioritizing essential information at both the word and sentence levels, thereby capturing the hierarchical structure of text more effectively. This approach, combined with LSTM's strong capabilities, creates a robust framework for sentiment classification. The experiment results show that the integration of these methods increases the performance of sentiment classification and provides a more comprehensive understanding of contexts in Turkish texts. The study examines in detail how these techniques are applied and the effects of these applications on the sentiment classification of the model and compares them to other configurations, methods and models.
Content-based image retrieval using deep learning and multidimensional indexing
Recent technological advancements and reductions in hardware and software costs have propelled visual search applications into the spotlight, making them both popular and indispensable. Consequently, the rapid and precise retrieval of images from vast databases through image queries has become a critical task. We introduce a novel end-to-end retrieval architecture that significantly enhances retrieval performance compared to a baseline system conducting database searches at the video frame level. Leveraging a pre-trained Convolutional Neural Network (CNN) model, we employ unsupervised image retrieval processes to extract and store low-level features for efficient indexing. To facilitate fast and efficient retrieval, we implement a tree-based indexing structure that leverages low-level features known as the Vantage Point Tree (VP Tree). To make these features compatible with our system, we employ dimension-reduction techniques to represent them in a lower-dimensional space. Our experiments, conducted on a benchmark image dataset, demonstrate that this approach leads to faster and more accurate retrieval when compared to a state-of-the-art search method known as K-Nearest Neighbor (KNN) search. Furthermore, we assess the proposed technique against KNN using two real-world video datasets, and it consistently outperforms KNN by a significant margin.
Named entity recognition with neural networks and pretrained word embeddings
NER (Named Entity Recognition) is a critical component of natural language processing that identifies and categorizes essential information elements inside text such as names, places, and organizations. Utilizing these embeddings, neural networks obtain a greater knowledge of the complex relationships between words within the vast range of unstructured text, through meaningful semantic representations that include rich contextual variations. The main goal of this research is to increase the precision and robustness of NER systems by utilizing the capabilities of cutting-edge word embeddings, such as Word2Vec, GloVe, FastText, and BERT. Word embeddings play an essential role in this process given that they turn words into a high-dimensional space where semantic and syntactic commonalities are captured, allowing machine learning models to successfully handle and analyze unstructured textual input. This research suggests a popular design in sequence labeling tasks—the Bi-directional Long Short-Term Memory (BiLSTM) network—added with a Conditional Random Field (CRF) layer to accomplish this goal. Combining these two elements enables the model to accurately represent the context necessary for precise entity recognition by capturing local and global dependencies in unstructured textual data, thereby assessing and improving the identification and classification of entities within the input text.
Stance detection in Turkish dataset on Russia-Ukraine war
Social media has evolved into a crucial informational resource to understand public opinion on various issues in recent years. Therefore, the importance of automatic information extraction from these data has increased. Stance detection, one of the subtasks of natural language processing, is also a crucial issue for automatic information extraction. Stance detection automatically determines the user's side regarding a particular subject, event, or person. In this study, a Turkish-labelled data set focusing on the stance determination task to determine social media users' attitudes towards the Russia-Ukraine War was created, and various machine learning methods were evaluated on this data set. For this study, 8215 tweets were collected on Twitter and cleaned. The dataset then was tagged with two targets Russia, and Ukraine. Support Vector Machines, Random Forest, k-Nearest Neighbour, XGBoost, Long-Short Term Memory (LSTM), and Gated Recurrent Unit (GRU) models are employed with GloVe and Fastext word embedding. Since the dataset is unbalanced between the targets, undersampling and oversampling methods were also used with these algorithms. With an F1 score of 0.73 for Russia and 0.81 for Ukraine, the results showed the Support Vector Machines algorithm to produce the best outcomes. In addition to these results, LSTM and GRU also produced outcomes that were highly comparable to those of the Support Vector Machines algorithm. The newly created Turkish corpus can be regarded as a valuable resource for this research area and in the future, transformer-based approach can be used with this corpus. Therefore, this study advances the field of stance detection research using Turkish text.
Predicting house prices in Ankara using machine learning
The focus of this thesis is to investigate whether machine learning predictions are accurate and viable enough to replace traditional real estate appraisal reports. To do this, we compare two datasets, one scraped from a real estate website and the other created from appraisal reports, and use various machine learning and neural network methods to find the best performing one and to determine the practicality of the study. Bagging and boosting ensemble methods are compared with the implementation of Extreme Gradient Boosting and Random Forest Models. Also, an Artificial Neural Network with five layers and Relu activation function is built as well as ensemble learning models. Hyperparameters of all models built throughout the study are chosen diligently for a comprehensive comparison. We evaluate the success of the models using root mean square error and accuracy score. Findings suggest that this approach has potential for improving the real estate valuation process, but further research is needed to determine its viability in the real world.