This thesis explores the use of small to medium-sized language models (LLMs) for multi-hop question answering, focusing on overcoming their limitations in knowledge and reasoning compared to larger models like GPT-4. The research explores using knowledge graphs to enhance these models' performance, specifically by integrating entity-relation triplets extracted from text. Techniques like supervised fine-tuning and reinforcement learning are evaluated for improving entity-relation extraction. A new prompting technique, Connect-the-Entities (CTE), is introduced to improve question answering performance with lower computational costs.

Date: 29.08.2024 / 10:00 Place: A-212

There are two goals of the thesis: (i) to understand the descriptive properties of the so-called negative dependencies in Turkish, which are expressions that basically require the presence of a negative-like element or environment to be well-formed; and (ii) to understand the so-called intervention effects on negative dependencies, where the term refers to the anti-licensing of negative dependencies in an otherwise adequate structure due to the presence of some elements (e.g. conjunction, universal quantifier, etc.).

Date: 06.09.2024 / 14:00 Place: A-212

In the literature, it is widely accepted that there is a direct relationship between the aspectual interpretation of the predicate and the internal argument of the verb. For languages such as Turkish and Finnish, it is argued that the accusative case marker on the direct object functions as a measurer and elicits a delimited event interpretation. However, in this thesis, we will argue that the correlation between the accusative case and the delimited interpretation of events is not a result of the supposed direct relationship between case and aspect, and show that the said relationship is only indirect following MacDonald’s (2006) work. 

Date: 06.09.2024 / 12:00 Place: A-212

In software development, the metaphor of technical debt (TD) describes the cost of additional work caused by choosing a non-ideal solution. This study aims to identify how cultural dynamics affect TD accumulation and management. Using both quantitative and qualitative methods, responses from 30 software industry practitioners across six different domains revealed that organizations with clan and market cultures tend to accumulate higher levels of TD. These findings demonstrate that organizational culture has a significant impact on TD outcomes and emphasize the importance of management strategies tailored to specific cultural dynamics within organizations.

Date: 16.08.2024 / 13:30 Place: A-212

This thesis investigates the integration of Model-Based Systems Engineering (MBSE) and Product Line Engineering (PLE) into Model-Based Product Line Engineering (MBPLE) for systematic variability management. It develops a robust MBPLE methodology by adapting variability management techniques, enhancing visualization, and ensuring traceability across abstraction levels. The methodology is applied to a case study and validated against established requirements, aiming to optimize business processes, improve product quality, and reduce engineering efforts within the aerospace industry.

Date: 06.08.2024 /10:00 Place: A-212

Brain-computer interfaces (BCIs) offer solutions for motor impairments and enhance human-computer interaction in virtual reality and cognitive augmentation. Adoption is hindered by user fatigue and the unnatural feel of visual stimuli, necessitating comfortable, intuitive paradigms. This study develops an imperceptible steady-state motion visual evoked potential (SSMVEP) stimulus for multi-directional BCI control. Using sinusoidal gratings with high-frequency motion, the first experiment shows robust cortical responses with reduced discomfort. The second experiment combines eye-tracking, EEG, and advanced machine learning to decode attentional responses accurately. Findings support naturalistic, high-performance BCIs for assistive technologies and human-computer interaction.

Date: 05.07.2024 / 14:00 Place: A-212

We employ the XGBoost algorithm to predict driver non-coding mutations based on multiple engineered features, augmented with features from existing annotation and effect prediction tools. The resulting dataset is passed through a feature selection and engineering pipeline and then trained to predict driver versus passenger non-coding mutations. We also use this model within the architecture of a known driver discovery model from existing literature. We then use non-coding driver mutations found in previous studies and predict their driver-ness using our models. Furthermore, we use Explainable AI methodologies to perform an in-depth analysis of the generated predictions.

Date: 07.06.2024 / 10:00 Place: A-212

This thesis explores attribution mechanisms in Turkish discourse through the adaptation of the Penn Discourse TreeBank (PDTB) framework, resulting in the Turkish Discourse Bank (TDB 1.2). Utilizing insights from lexical control and eventuality specific to Turkish, a custom annotation scheme was developed, facilitating robust data annotation. Analysis shows the predominance of communicative verbs in attribution instances, highlighting novels and news as rich domains for study. Achieving high inter-annotator agreement, this work advances the field by enriching the TDB and laying groundwork for future automated text analysis in Turkish.

Date: 04.09.2024 / 10:00 Place: B-116

This study investigates the association between SNPs and COVID-19 susceptibility in the Turkish population, focusing on patients from Hacettepe University Hospital. Using NGS, we analyzed SNP data from various scientific publications, performing variant calling, linkage analysis, and statistical comparisons with non-Finnish European allele frequencies. Key findings indicate that certain variants have different frequencies compared to the European population, suggesting genetic predispositions affecting disease susceptibility in the Turkish population. Linkage disequilibrium analysis revealed strong correlations between specific genetic loci.

Date: 23.07.2024 / 15:00 Place: A-212

This study examines infrastructure issues and system malfunctions in Cloud Computing systems exacerbated by the COVID-19 pandemic, which acts as a stress test due to increased demand. It is argued that task scheduling algorithms are the main source of these problems. Post-pandemic Cloud Computing task scheduling algorithms were systematically reviewed and analyzed using the Meta-Synthesis method. A global categorization schema for these algorithms was presented, comparing their advantages, disadvantages, applications and vulnerabilities. Current task scheduling approaches and trends in Cloud Computing were analyzed comparatively.

Date: 27.05.2024 / 13:30 Place: B-116