2 sonuçlar
Arama Sonuçları
Listeleniyor 1 - 2 / 2
Yayın All-words word sense disambiguation in Turkish(Işık Üniversitesi, 2019-09-06) Akçakaya, Sinan; Yıldız, Olcay Taner; Işık Üniversitesi, Fen Bilimleri Enstitüsü, Bilgisayar Mühendisliği Yüksek Lisans ProgramıWord sense disambiguation (WSD) is the identi cation of the meaning of words in context in a computational manner. The main subject of this study is to implement and compare the WSD results of various supervised classi ers (Naive Bayes, K Nearest Neighbor, Rocchio and C4.5) in all-words setting. To this end, we have constructed an all-words sense annotated Turkish corpus, using traditional method of manual tagging. During the annotation, a pre-built parallel treebank (aligned from Penn Treebank) has been tagged with the senses of Turkish Language Institutions dictionary. The approach of annotating a treebank allowed us to generate a full-coverage resource, in which syntactic and semantic information merged. In the WSD evaluations, three distinct experiments have been organized to determine the efect of using different feature sets on the disambiguation performance. First experiment has been conducted with a simple feature set that includes the fundamental local features. In the second experiment, the initial feature set has been augmented with several effective morphological features, and in the third one, the feature set has further been extended with the syntactic features. Our test results show that all classi ers have achieved better results in parallel to growing feature set. Additionally, integration of syntactic features has proved to be useful for WSD.Yayın Fingertip electrocardiogram and speech signal based biometric recognition system(Işık Üniversitesi, 2021-12-27) Güven, Gökhan; Güz, Ümit; Gürkan, Hakan; Işık Üniversitesi, Lisansüstü Eğitim Enstitüsü, Elektronik Mühendisliği Doktora ProgramıFingertip electrocardiogram and speech signal based biometric recognition system In this research work, we presented a one-dimensional CNN-based person identification system which depends on the combination of both speech and ECG modalities to improve the overall performance compared to traditional systems. The proposed method has two approach: one is to develop combination of textindependent speech and fingertip ECG fusion system, the other one is to develop a robust rejection algorithm to prevent unauthorized access to the fusion system. In addition to the system robustness, we have developed an ECG spike and inconsistent beats removing algorithm, which detect and remove the problems caused by either portable fingertip ECG devices or movements of the patients. First approach has been tested on 30, 45, 60, 75 and 90 people which were taken from LibriSpeech Corpus database and combination of both CYBHi and our private fingertip ECG database. The 3-fold cross validation test setup has been conducted while system working time was set to 10 seconds. In the first experiment, we achieved 90.22% accuracy rate for 90 people for ECG based system. For the speech based system, 97.94% accuracy rate has achieved for 90 people. For the combination of both system, 99.92% accuracy rate has been achieved. For the second approach, 90 people for ECG and Speech database were being used as genuine class, 26 people as imposter class, and after the performance evaluation in optimum rejection thresholds, 71.08% accuracy rate for imposters rejection and 71.05% accuracy rate for genuine recognition has achieved for ECG based system. For the speech based system, imposter class were 87.82% accurately rejected while genuine classes were 86.48% accurately identified. The combination of both system has achieved 91.68% accuracy for genuine identification rate whereas 96.05% accuracy for imposter rejection.
