19
Archived Theses
0
DOIs Assigned
0%
DOI Rate
Discipline
Yüz yüze iletişime bakış merkezli çok modlu yaklaşım
Face-to-face conversation implies that interaction should be characterized as an inherently multimodal phenomenon involving both verbal and nonverbal signals. Gaze is a nonverbal cue that plays a key role in achieving social goals during the course of conversation. The purpose of this study is twofold: (i) to examine gaze behavior (i.e., aversion and gaze on face) and relations between gaze and speech in face to face interaction, (ii) to construct computational models to predict gaze behavior using high-level speech features. We employed a job interview setting, where pairs (a professional interviewer and an interviewee) conducted mock job interviews. Twenty-eight pairs of native speakers took part in the experiment. Two eye-tracking glasses recorded the scene video, the audio and the eye gaze position of the participants. To achieve the first purpose, we developed an open-source framework, named MAGiC (A Multimodal Framework for Analyzing Gaze in Communication), for the analyses of multimodal data including video recording data for face tracking, gaze data from the eye trackers, and the audio data for speech segmentation. We annotated speech with two methods: (i) ISO 24617-2 Standard for Dialogue Act Annotation and, (ii) using tags employed by the previous studies that examined gaze behavior in a social context. We then trained simplified versions of two CNN architectures (VGGNet and ResNet) by using both speech annotation methods.
Kelimeleri hecelerine bölmenin okuma güçlügü çeken çocukların okuma performansı üzerine etkileri: Okuma becerileri geri kalmış çocuklar için yardımcı bir teknik
There is a growing body of evidence supporting the idea that dyslexia exhibits itself differently in different languages. Based on how reading is taught in Turkish Language, the symptoms exhibited by students with reading difficulties may be different than the ones exhibited by English speaking students. Even though there is no support for the role of reading syllables as a reading unit in the literature, it is clear that the students in Turkey are taught how to read syllables and are actively using the strategy of reading a word syllable by syllable in their early reading development. Decoding words syllable by syllable, naturally, requires segmentation of the words into their subunits, namely syllables. During the slow and struggling serial decoding of syllables, which is also based on serial decoding of phonemes to corresponding graphemes, the problem is transformed into bringing the correct number of phonemes together since the syllable length is not constant. It entails an increase in the number of the mistakes while reading a word in relation to its complexity in terms of the number and the variability of the syllables it includes. Deciding how many phonemes are supposed to be brought together and isolating them from the other graphemes in the word until the decoding of the syllable finishes is of critical importance during this process. The current study tested and found significant effects of segmenting the words into its syllables on behalf of the learner. The results suggest that aiding the segmentation process significantly improved the pronunciation of the syllables and words and decreased the number of mistakes during reading. It is hoped the results will help a better understanding of reading difficulties in Turkish, which in turn might help the development of more effective intervention techniques to the problem at hand.
Oryantiring sporunda rota özelliklerinin karar vermeye etkisi üzerine bir durum çalışması
Orienteering is a sport where athletes need to find located targets at certain points on a predetermined terrain or in the city with the help of a map. In this sport where performance is measured with time, it is important to combine their physical endurance with mental processes and their ability to adapt to the environment and optimize them correctly. One of the outcomes that we can best observe these choices is the routes chosen and each route has its own environmental characteristics. Therefore, athletes need to analyze these characteristics well and as a result, they need to choose the most suitable route for themselves. In this thesis, the components affecting route selection are investigated. For this purpose, athletes' data was obtained through GPS containing watches from an orienteering race held by Turkey Orienteering Federation. The collected data were examined by quantitative and qualitative methods, and a general understanding of athletes' behavior was obtained and the distinction of modelable and subject-dependent factors in decision making of athletes was made. From the modelable components, a model that computes the shortest distance based on the distance and terrain surface has been created and then its compatibility with the behaviors of athletes was examined. Additionally, the results were supported by various statistical analyzes. According to the results of the study, environmental variables play a major role in the decision-making of athletes and model performance is more accurate in short-distance routes than long-distance routes in greater need of reducing the cognitive load.
Morfolojik analiz, sözcük türü işaretleme ve bağlılık ayrıştırmanın eş zamanlı öğrenilmesi
In agglutinating languages, there is a strong relationship between morphology and syntax. Inflectional and derivational suffixes have a significant role while determining the syntactic role of the word in the sentence. This connection enables the joint learning of morphology and syntax. Apart from that, the complex morphology poses a sparsity problem. In this respect, morphological analysis and segmentation are vital for various natural language processing applications. All of these have provided the primary motivation to develop a joint learning model for morphological segmentation, morphological tagging, part-of-speech (POS) tagging, and dependency parsing. The proposed model consists of a multi-layered neural network structure. In each level, there is a bidirectional long-short memory unit (BiLSTM) to encode sequential information. Additionally, attention networks are used to compute soft alignment between encoder-decoder states in the morphological tagging component. Finally, the obtained results from each layer of the network are compared with other works from the literature. The results are very competitive on Universal Dependencies (UD) dataset.
Etkileşimli erişilebilirlik için göz takibi: İletişim arayüzlerinin kullanılabilirlik analizi
At the late stages of ALS, i.e., Amyotrophic Lateral Sclerosis, individuals may lose their ability to speak, eat, move, and even breathe without ventilator support. Alternative modalities of communication, such as gaze, may have a potential to provide a communication interface that may increase the quality of life in cases as such. This study aims at developing interactive interfaces for gaze-based communication, and it reports a small-scale usability analysis for a comparison of the interfaces. In particular, we focus on two dimensions in the experimental investigation: The interface layout design and gaze recording interfaces. The interface layouts include a virtual QWERTY keyboard, an alphabetical one, as well as a cascades menu. The gaze recording interfaces include a wearable eye tracker and a webcam-based eye tracker. More generally, we focus on the role of the simplicity of the interface, the functionality of it, and the usability, which may be three contradicting aspects of interface design and development.
Toplama ve çıkarma problemleri için bir matematiksel kelime problemi çözücüsü
Mathematical Word Problems are mathematical problems written in natural languages. These problems are crucial for mathematics education. This thesis aims to develop a model to solve addition and subtraction math word problems automatically. At first, an intermediate representation is introduced to mimic the human representation computationally. Dependency structures and part-of-speech tags are used to provide the auxiliary linguistic information about the problem. Circumscription idea is used to fill the blanks in intermediate representation. Then, an automata-based system has been designed to generate the equations for given problems with the help of domain-specific verb categories. The model has been developed and evaluated with different datasets. In conclusion chapter, possible future studies that may use our system were discussed.
Yönetici kontrol alıştırmasının çatışma çözümleme sırasında emosyonel dikkat dağılımına etkisi: Göz bebeği çapı ölçüm çalışması
Task-irrelevant emotion distracts executive processing on the one hand and executive control suppresses emotional processing on the other. This reciprocal cognition-emotion link underlies emotion regulation, which is critical for mental health. The present dissertation investigated the intricate relationship between cognitive load and emotional distraction, using non-emotional executive control training and taking mood-related effects into account. Two groups of participants were either trained on a high-load conflict resolution task or completed a low-load simple identification task. Both groups then completed an intermediate-load conflict resolution task with (Experiment 1) and without (Experiment 2) distracters. Distracter pictures were manipulated considering both valence and arousal dimensions. During training session and subsequent experiments, participants' accuracy scores, reaction times, and pupil diameters were recorded. At the behavioral level, the results of Experiment 1 revealed impaired performance in the presence of all task-irrelevant emotional pictures irrespective of cognitive load and training paradigm. Furthermore, after adjusting for state anxiety levels, cognitive load and training interacted with emotional distraction. At the physiological level, the effects of executive training were observed exclusively for negative arousing distracters, for which executive training reduced pupillary responses. Importantly, sustained effects of training were observed in Experiment 2; training reversed the association between behavioral and physiological responses and depression and anxiety levels. Together, the findings of the present study have important implications for emotion regulation and treatment of mental disorders by demonstrating the modulation of emotional distraction and mood-related behavioral and physiological responses by single-session non-emotional executive control training.
Okumada fonolojik dolayım: Teorik bir çerçeve
A set of perceptual and cognitive processes at various levels, from low-level visual perception to high-level discourse comprehension, underlie reading. Accordingly, research on reading focuses on different aspects of reading ranging from prelexical processing and word recognition to syntactic parsing, sentence comprehension, and discourse comprehension with various technical and theoretical tools including behavioral experiments, neuroimaging techniques, and computational models of eye movement controlling, which reflect the variety of the levels involved during reading. The focus of the current study was the early prelexical and lexical processing and postlexical integration processes involved in the word recognition process during text reading from a perspective of eye-movement control modeling. A framework for a computational model of guidance by attentional gradient (GAG) eye-movement control model that includes the role of phonological processes during reading was presented. The assumptions of the framework were tested by two sets of linear mixed models (LMMs) with data from Turkish Reading Corpus: (1) an LMM of fixation speech interval (FSI), and three LMMs of eye movement measures among oral reading data, and (2) three LMMs of eye movement measures among silent reading data. The results of the LMMs were compatible with the canonical findings frequently reported in the literature. Influences of the neighboring words on eye movement measures in the current study were mixed. The results indicated an effect of prelexical phonological processing on eye movements and the involvement of phonological representations on postlexical processing.
Konuşma ihlalleri bulunduran etkileşimlerin Gricein niyet-merkezli semantik bakış açısı bağlamında görsel algı aracılığıyla incelenmesi
Linguistic principles at various levels are crucial in maintaining a reliable and transparent communication for dyadic interactions. However, violating these principles might result in unwieldy and problematic communications. Gaze can be a medium of reflecting the cognitive responses when conversational violations occur. An eye-tracking study was conducted to investigate visual patterns in communication in response to social communication errors, specifically Grice's Maxims violations. This study investigates how social-communicative errors affect task performance and gaze during social interaction. The results suggest participants' visual exploration patterns shift towards the violator speaker for the maxim of Relation violation, and the gaze stays mostly within the task area for the Quantity, Quality, and Manner violations. In addition, the response time increased for Quantity and Quality violations, followed by the Manner and Relation violations. Overall, this work ties visual cognition and linguistic principles for conversational Gricean Maxim violations in which the findings contribute to the design space in conversational agent interactions, particularly in human-computer interactions.
Türkçe okuma sırasında nefes alma örüntüleri
Breathing is a process that is actively involved in speech production and has an impact on pace of reading. The main purpose of this thesis is to study the relationship between text complexity and the breathing pattern of the readers. This study approaches this question from two distinct aspects. First, it asks whether the syntactic complexity of a text affects the breathing frequency during reading in Turkish. Secondly, it investigates whether text complexity influences the readers' decisions on where to re-breath within the sentences. In order to answer these questions, three experimental conditions were generated: reading two simple texts aloud, reading two middle-complex text aloud, reading two complex texts aloud. The results reveal significant effect of the text complexity level both on the breathing frequency and the locations chosen to re-breath by the readers. The outcomes may provide an input for designing speech generation patterns that may improve naturalness of synthetic speech.
Türkçede statik uzamsal dil semantiğinin kelime çeşitliliği açısından incelenmesi: Bir göz hareketleri takip çalışması
The semantics of spatial terms has been attracting the attention of researchers for the past several decades. As an understudied language, Turkish presents an appropriate test bed for studying the generalizability of semantic characterization of spatial terms across languages. Turkish also exhibits unique characteristics, such as the use of locative case markers and being an agglutinative language. The present study reports an eye-tracking investigation of comprehension of spatial terms in Turkish by employing Topological Relations in Picture Series (Bowerman & Pederson, 1992). The major research question of the study is the presence of a relationship between the variety of spatial expressions produced by native speakers and fixation patterns on the stimuli. The findings reveal that the richness of the spatial expression is related to longer and more frequent fixations on the stimuli. The findings also show that the presence of a locative case marker in the utterances plays a role in this relation. We also investigated whether Turkish native speakers exhibit any sensitivity on Core and Non-Core distinction in spatial term semantics (Landau, 2017). Our findings showed that some of the Non-Core term categories reveal more variety of spatial expressions and longer fixation durations on the Figure, supporting cross-linguistic Core categorization.
Dans sporu değerlendirmesinde uzmanlık: Sözel betimlemelere dayalı bir sınıflandırma yaklaşımı
Dance sport is a performance in which a male and a female aim at exhibiting certain body techniques in a limited time period. The judges, who report the performance evaluation, focus on the body lines, the bodily communication between the partners, and the sense of rhythm. The evaluation process of dance sport may vary across judges' level of experience. The goal of the thesis is to investigate the modeling capabilities of dance performance evaluation, which eventually may lead to a binary classification of expert and novice evaluators through their verbal descriptions of the dance activity. In particular, the models presented in this thesis aim to classify the evaluator as an expert or a novice through the analysis of speech data. For this, we trained two binary classification models, namely Multinomial Naïve Bayes and DistilBERT. The findings reveal that both models may return acceptable results for English and Turkish, though their different performance in accuracy.
Benzetim analiz laboratuvarında karşı kuvvetlerin bilişsel modellenmesi
In modern military systems, engagement simulations are used to assist in conceptual design and verification of weapon systems. The quality of these simulations is directly proportional to the level of authenticity of the behavior of the components that are modeled. For this reason, it is important to represent the behavior of human controlled components by accounting for the cognitive aspects including human limitations and flaws. In various studies, the behaviors of these components are modeled using cognitive architectures. These representations of behavior can contain several cognitive processes, but in many studies, a single cognitive process has been modeled and the interacting cognitive processes have not been adequately studied. In this thesis, computational cognitive model of situational awareness and surprisal related cognitive processes were implemented using SOAR cognitive architecture. The behaviors of agents guided by these cognitive models have been observed by performing various experiments in the STAGE simulation environment. Interaction of the two cognitive processes modelled has been discussed.
Türkçe cümlelerdeki kelimelerin LSTM tabanlı dil modellemesiyle tahmini
Language comprehension is affected by predictions because it is an incremental process. Predictability has been an important aspect of studying language processing and acquisition in cognitive science. In parallel, Natural Language Processing field takes advantage of advanced technology to teach computers how to understand natural language. Our study investigates if there is an alignment between human predictability and artificial language model predictability results. This thesis solely focuses on the Turkish language. Therefore, we have built a word-level Turkish language model. Our model is based on Long Short-Term Memory (LSTM), which is a recently trending method in NLP. Alternative models are trained and evaluated with their prediction accuracy on test data. Finally, the best performing model is compared to human predictability scores gathered from the cloze-test experiment. We have shown a promising correlation and analyze the cases where the correlation is high or low.
Türkçedeki yüksek frekanslı yapım eklerinin semantiğinin keşfi için dağılımsal incelenmesi
In agglutinating languages such as Turkish, the process of derivation is mostly performed by adding suffixes at the end of words. Most of the derivational suffixes carry a distinctive semantic content and representing them has an important role in computational tasks, such as question answering. In this thesis, we aim to explore the structure of some frequent Turkish derivational suffixes in distributional vector space by clustering word embedding vectors of them and analyzing their underlying semantic properties. Suffix vectors are obtained by subtracting the vector of the base form of the derived word from the derived word's word vector. We used a pre-trained word embedding model for obtaining word vectors and multiple unsupervised clustering algorithms with different parameters for clustering them. Our assumption is if a derivational suffix category manages to dominate one or more clusters, it is possible to obtain reliable representations of it in the distributional vector space. Our results show that many Turkish derivational suffix categories have this capability. We analyzed the underlying semantic structure of the generated clusters in terms of the thematic roles the suffixes are selecting, the UCCA labels and the UD relations the stem and the derived word can get.
Kolaboratif robotlarda güven özelliği: sanal insan robot etkileşim ortamında, sözsüz ipuçlarının deneysel araştırması
This thesis reports the development of non-verbal HRI (Human-Robot Interaction) behaviors on a robotic manipulator, evaluating the role of trust in collaborative assembly tasks. Towards this end, we developed four non-verbal HRI behaviors, namely gazing, head nodding, tilting, and shaking, on a UR5 robotic manipulator. We used them under different degrees of trust of the user to the robot actions. Specifically, we used a certain "head-on neck posture" for the cobot using the last three links along with the gripper. The gaze behavior directed the gripper towards the desired point in space, alongside with the head nodding and shaking behaviors. We designed a remote setup to experiment subjects interacting with the cobot remotely via Zoom teleconferencing. In a simple collaborative scenario, the efficacy of these behaviors was assessed in terms of their impact on the formation of trust between the robot and the user and task performance. 19 people participated in the experiment with varying ages and genders.
Makine öğrenmesi algoritmaları ile sosyal medyada Türkçe clickbait tespiti
Clickbait strategy, mostly used in headlines and teaser messages, aims to attract people's attention, and make them click on the link by using intriguing expressions with various text-related features. Clickbait, which has become very common especially in social media in recent years, is a major problem for the flow of information. Since the information promised in the clickbait headline is generally not included in the main text, clickbait headlines disappoint readers and is problematic for ethics of journalism. In this thesis, we constructed a Turkish dataset –ClickbaitTR– with 48,060 samples, including headlines of Turkish news sources extracted from Twitter, and made it publicly available. Various machine learning algorithms such as Artificial Neural Network (ANN), Logistic Regression (LR), Random Forest (RF), Long Short-Term Memory Network (LSTM), Bidirectional Long Short-Term Memory (BiLSTM), and Ensemble Classifier (EC) were applied on the dataset for detecting the clickbait headlines. The results show that the BiLSTM has the best performance in detecting clickbait headlines with 97% accuracy followed by the LSTM, the ANN, and the Ensemble Classifier with 93% accuracy. In addition to a successful clickbait detection performance, in this thesis, linguistic and psychological analysis of clickbait sentences were presented with a focus on psychological mechanisms such as curiosity and interest. This thesis contributes to clickbait detection studies with the largest clickbait dataset and best clickbait detection performance in Turkish.
Arayüzle etkileşime giren kullanıcının bilişsel yükünün bayes ağına dayalı tahmini
The complexity of human machine interfaces is increasing significantly in parallel with the development of technology and excessive data growth, but human cognitive capacity is limited. Therefore, measuring cognitive load is one of the most preferential and common ways to test the usability of user interfaces. There are many different physiological, behavioral and subjective methods to measure human performance and workload. Moreover, there are cognitive predictive models and many related applications based on these models to predict performance and human workload on computer based tasks. The purpose of this study is to estimate the cognitive load and performance of the person by evaluating multiple methods together based on Bayesian network. For this, we modeled a Bayesian network that both uses a cognitive predictive model, and learns and regulates it with subjective data collected from people. After modelling, we conducted experiments with the interfaces of two different defense projects to collect data. We used the adapted Bedford scale at the end of each task of an interface and the NASA TLX rating scale for the overall rating of the interface after all tasks were completed. We confirmed that the Bayesian network effectively estimated the user's workload and performance. Our findings reveal that this model performs cognitive load analyzes much more efficiently in a short time. This study also demonstrates the differences between tasks and users, providing the opportunity to detect the complexity of subtasks and perform personalized performance and cognitive load analysis for each user.
Majör depresif bozuklukta duygu düzenleme farklarına yönelik etkili bağlantı modeli: Dinamik nedensel modelleme analizi
Major Depressive Disorder (MDD) is a mental disorder and one of the most critical and prevalent disorders in the world. Unlike other diseases, the diagnosis of mental disorders does not easily conform to objective tests. Medical experts consider several indicators to be able to distinguish a depressed person from a normal individual, and finding robust markers to aid diagnosis is still an active area of research. The recent proliferation of neuroimaging methods has brought up new opportunities in that regard. This study aims to contribute to these efforts by investigating the utility of Dynamic Causal Modeling (DCM) based effective connectivity measures for distinguishing MDD patients and healthy controls based on their responses to emotional stimuli. The analysis was conducted over an open fMRI dataset, including the brain responses of MDD patients and healthy controls to an emotional musical stimuli task. The results of the DCM effective connectivity model reveal an increasing sgACC to Amygdala connectivity and reduced dlPFC to Amygdala connectivity in the healthy controls compared to MDD patients. Thus, the findings of this study suggest that such differences in effective connectivity patterns in response to emotional stimuli can be useful in distinguishing MDD cases from healthy subjects.