2 sonuçlar
Arama Sonuçları
Listeleniyor 1 - 2 / 2
Yayın Semantic relation extraction by enriching word embeddings exploiting Turkish morphology(Işık Üniversitesi, Lisansüstü Eğitim Enstitüsü, 2025-03-18) Ercan, Gökhan; Yıldız, Olcay Taner; Işık Üniversitesi, Lisansüstü Eğitim Enstitüsü, Bilgisayar Mühendisliği Doktora ProgramıDistributed representations (DR) are used to capture semantic and syntactic patterns in language by analyzing the distributional relationships of words within textual data. The modeling methods that produce DR are based on the assumption (distributional hypothesis) that "words that occur in the same context tend to have similar meanings," which is inherent to the nature of language. These modeling methods, due to their unsupervised nature, can be trained without human judgment input, allowing researchers to train large datasets at relatively low costs. Although word-based models perform effectively for languages with limited vocabularies, such as English, they exhibit considerable inefficiency when applied to morphologically rich languages with unlimited vocabularies, such as Turkish. We observed that n-gram and statistical segmentation methods, which are commonly used in subword modeling to address the issues of out-of-vocabulary and rare-words, are highly sensitive to orthographic similarity. Consequently, these methods struggle to distinguish between unrelated concepts (e.g., shrink - shrine). Moreover, we noted that the impact of morphological segmentation methods on these types of problems has shown inconsistent results in the literature. This thesis aims to make conceptual assumptions and improvements concerning different types of semantic relationships (e.g., relatedness and similarity), to model the role of language morphology as an input in subword DR models, and to develop the dataset generation methodologies and evaluation methods to measure this effect. Within the scope of the study, different models and segmentation methods were empirically tested, the AnlamVer and OSimUnr datasets were produced, and the task of relatedness classification and associated evaluation methods were proposed to measure the noise introduced by segmentation to the model. Our experiments demonstrate that morphological segmentation produces significantly less noise compared to n-gram-based methods and can lead to substantial performance improvements depending on the nature of the task.Yayın Grammar or crammer? the role of morphology in distinguishing orthographically similar but semantically unrelated words(Institute of Electrical and Electronics Engineers Inc., 2025) Ercan, Gökhan; Yıldız, Olcay TanerWe show that n-gram-based distributional models fail to distinguish unrelated words due to the noise in semantic spaces. This issue remains hidden in conventional benchmarks but becomes more pronounced when orthographic similarity is high. To highlight this problem, we introduce OSimUnr, a dataset of nearly one million English and Turkish word-pairs that are orthographically similar but semantically unrelated (e.g., grammar - crammer). These pairs are generated through a graph-based WordNet approach and morphological resources. We define two evaluation tasks - unrelatedness identification and relatedness classification - to test semantic models. Our experiments reveal that FastText, with default n-gram segmentation, performs poorly (below 5% accuracy) in identifying unrelated words. However, morphological segmentation overcomes this issue, boosting accuracy to 68% (English) and 71% (Turkish) without compromising performance on standard benchmarks (RareWords, MTurk771, MEN, AnlamVer). Furthermore, our results suggest that even state-of-the-art LLMs, including Llama 3.3 and GPT-4o-mini, may exhibit noise in their semantic spaces, particularly in highly synthetic languages such as Turkish. To ensure dataset quality, we leverage WordNet, MorphoLex, and NLTK, covering fully derivational morphology supporting atomic roots (e.g., '-co_here+ance+y' for 'coherency'), with 405 affixes in Turkish and 467 in English.
