SEARCH RESULT

Embedding models play a crucial role by capturing the semantics of contexts in information retrieval, complementing traditional lexical methods. With the rise of Retrieval-Augmented Generation (RAG) systems, information retrieval has become an essential component of text generation. Generative models tend to produce misleading answers when an answer requires an external source, when the model hallucinates, or when the response becomes outdated due to the temporal limitations of training datasets. Retrieval-augmented methods enhance response reliability and accuracy by leveraging relevant retrieved information. Therefore, improving retrieval performance is essential for ensuring more precise and dependable outputs. As a core component of RAG systems, embedding models can be multilingual, including up to a hundred languages. However, they have inadequacies with low-resource languages such as Turkish, because training data cannot provide equal coverage across all languages. In contrast, text-matching-based lexical search algorithms work independently of language. Nevertheless, they have shortcomings with mismatching cases due to suffixes and prefixes with agglutinative languages. While lemmatization and stemming provide partial solutions, these methods cannot be applied directly to morphologically more complex languages like Turkish. This study introduces Turkuaz-Embeddings, a novel model specifically optimized for Turkish. It consistently outperforms widely used multilingual and Turkish embedding models, achieving up to 20% and on average 9% improvements in retrieval performance across diverse benchmarks. In some scenarios, it ranks highest among both multilingual and Turkish embedding models. Moreover, compared to its baseline architecture, it shows up to 35% improvement, with an average gain of 20%. The model also excels in zero-shot retrieval tasks, demonstrating robust generalizability. This study also proposes an innovative token-based lexical retrieval method. By leveraging sub-word tokens from a Turkish tokenizer, without relying on lemmatization and stemming, this approach enhances traditional lexical search performance. It yields up to 10% improvement and consistently achieves an average gain of 5% across evaluation scenarios. Both advancements in this study contribute to improved semantic representation, more efficient lexical matching, better information retrieval, and eventually, more accurate and contextually relevant generations in Turkish RAG systems.

International Conference on Advanced Technologies, Computer Engineering and Science
ICATCES

Enes Sadi Uysal Mehmet Fatih AMASYALI