Abstract: Early-season crop yield prediction is essential for agricultural planning but remains uncertain due to incomplete observations and evolving crop conditions. This thesis examines the trade-off between forecast timing and yield prediction accuracy within a satellite-based yield forecasting framework, focusing on how uncertainty changes when predictions are issued earlier in the growing season. Multi-year Sentinel-2 imagery was integrated with meteorological data from the Copernicus Climate Data Store to construct weekly time series of vegetation indices and environmental variables. Cloud contamination was addressed through a cloud masking and temporal reconstruction procedure, and the proportion of reconstructed pixels was quantified to assess data reliability. Yield prediction models were developed for forecast months ranging from March to October. Multiple modeling approaches, including Random Forests, Gradient Boosting, and Neural Networks, were evaluated, with Linear Regression selected as the final model due to its robustness, interpretability, and stable performance across prediction horizons. Model performance was assessed using R², RMSE, MAE, and relative error metrics. Results indicate a strong dependence of prediction accuracy on forecast timing. Early-season predictions exhibit high uncertainty, while performance improves substantially later in the season. July and August provide a practical balance between timeliness and reliability, whereas September and October yield high-confidence estimates. All forecasts target the final end-of-season (December) yield of the corresponding production year.
Abstract: Knowledge Graph Question Answering (KGQA) sits at the intersection of natural language processing and structured knowledge representation, but existing approaches trade off flexibility for faithfulness: pure Large Language Model (LLM) methods hallucinate and are expensive, while symbolic methods lack semantic flexibility. This thesis presents HyGRA (Hybrid Graph Reasoning Agent), a dual-layer cognitive architecture that combines a Graph Neural Network (GNN)-enhanced hybrid retrieval layer with an autonomous ReAct agent to deliver graph-grounded, hallucination-resistant question answering over industrial knowledge graphs. Inspired by Kahneman's dual-process theory, HyGRA routes queries through a fast Reflex Layer (FAISS over GraphSAGE embeddings plus Cypher lexical matching) and falls back to a deliberate Reasoning Layer in which a ReAct agent navigates the graph through SEARCH, EXPAND, and READ operations. On a domain-specific industrial KG (516 questions, 0–2-hop complexity), HyGRA reaches an answer rate of 88.2% with a local Llama-3-8B backbone and 87.0% with GPT-5.1, outperforming Think-on-Graph (28.7%) and Paths-over-Graph (32.6%). A stricter source-based evaluation over 342 questions exposes confident-but-wrong answers and reduces HyGRA's margin over a Text-to-Cypher baseline to 0.7pp on F1. A second evaluation on Freebase WebQSP reverses the ranking: ToG and PoG reach 50.0%, HyGRA drops to 33.0%. The two boundaries point the same direction: paradigm choice is not universal, it follows the graph.
Abstract: History-taking is a foundational clinical skill in medical education, yet opportunities for safe, repeatable practice remain limited due to the resource-intensive nature of traditional standardized patient methods. Recent advances in large language models have enabled the development of scalable AI-driven alternatives that can simulate realistic patient encounters on demand. This study presents ECOSIM, a web-based virtual patient system designed to support history-taking skills training through fine-tuned language models and automated rubric-based assessment. OpenAI GPT-4o-mini and Google Gemini 1.5 Flash were fine-tuned on 183 patient-doctor question-and-answer pairs collected through ECOSIM interactions and four model configurations were comparatively evaluated by domain experts. A rubric adapted from the Calgary-Cambridge Observation Guide was integrated into an automated scoring engine that provides students with immediate, personalised feedback at session end. Results show that fine-tuning substantially improves response naturalness and contextual consistency, while the automated assessment engine achieves strong agreement with expert evaluations.
Abstract: Authorship attribution, genre classification, and recommendation systems based on literary content have become increasingly important in natural language processing (NLP). This thesis presents a system capable of both classification and content-based recommendation by representing book texts in a vector space. Comprehensive datasets of Turkish and English novels were compiled. All texts were lowercased and preprocessed to remove punctuation and special characters, then transformed into fixed-length vectors using the Doc2Vec algorithm. To ensure generalization, a custom train/test split was applied, guaranteeing each author was represented in the test set. These document vectors were used for authorship and genre classification with classifiers such as Logistic Regression, LinearSVC, Random Forest, Gaussian Naive Bayes, and K-Nearest Neighbors. This study also adopts a centroid-based similarity approach that extends traditional cosine similarity principles. For each author and genre, a representative vector was computed from the training data, and test books were classified based on their angular proximity to these centroids. This approach improved conceptual sensitivity and yielded high accuracy. In the second phase, a content-based recommendation system was developed that, unlike collaborative filtering, relies solely on textual similarity. It recommends books with similar narrative styles to those already enjoyed by the reader, helping raise the visibility of lesser-known authors. This research shows that document embedding-based methods can be effectively applied in digital humanities, authorship attribution, and recommendation systems.
Abstract: In this thesis, a comprehensive performance comparison is presented on which large language model, vector database, embedding model and similarity method in a retrieval augmented generation system, in the setting of a Turkish supported conversation bot for the students in higher education. Although there are studies in this field for many languages, the number of studies conducted in the Turkish language and especially in the field of education is low. Therefore, there is a need to identify the best alternatives for the components of such systems when they are used under Turkish language rules. This study gives answers to this need which is obtained through scientifical experimentation and is practically realizable. The most commonly used, well known and reliable large language models, embedding algorithms associated with them, vector databases and similarity measures were selected and several question sets based on university regulations were tested on different combinations of them. Based on the achieved results and analysis made, the best combination was obtained. In this way, the optimal Turkish supported retrieval augmented generation based virtual academic advisor bot was created so that students can easily access what they are inquiring about academic rules. Furthermore, the workload of academics and administrative staff was lightened with this automation.
Abstract: The frequent occurrence of natural disasters, particularly seismic events, leads to significant destruction and brings major catastrophes. Addressing the critical need for fast information during such crises is of paramount importance. Therefore, this study focuses on mitigating disaster-related losses through an innovative system that processes information available on social platforms. Development and comparative evaluation of the components of a system designed to rapidly identify and prioritize vital information from high amounts tweet data are detailed. A specific dataset was compiled, consisting of tweets posted in the aftermath of the major earthquake in Türkiye in February 2023. Using this data, messages were classified into three primary categories: disaster-related and urgent, disaster-related but not urgent, and non-disaster-related, utilizing both prompt engineering and fine-tuning methodologies with large language models. Performance of Gemini-1.0-pro and GPT-4o-mini models, as well as fine-tuned versions of these models were compared. The results indicated that the fine-tuned GPT-4o-mini model, when used with advanced prompting strategies, achieved a classification success of 93.64%. Following this classification, an impact score algorithm was implemented. This algorithm utilizes various tweet engagement metrics, such as reposts, likes, and views, to further rank and highlight more significant messages. This research is anticipated to offer substantial support to rescue and aid organizations by enhancing their situational awareness and response capabilities. Furthermore, it aims to make a significant contribution to the advancement of Turkish language processing in the field of crisis informatics.