Named-Entity Recognition (NER) identifies specific entities within a text, such as names of people, places, organizations, and institutions. While the technology can identify unique names, it faces challenges in domain-specific applications, and accurately targeting low-resource languages is even more challenging. The Kazakh language is an agglutinative and morphologically rich low-resource language. This study addresses the issue of data sparsity resulting from the language’s multi-dimensional features and the challenges posed by the unclear boundaries and ambiguity in the tourism domain. In this paper, we propose a low-resource Kazakh NER model based on small samples that benefits from word and stem-level embedding and dependencies across adjacent labels. This model uses words and stems as linguistic features for word embedding, enhances semantic features, and trains word vectors with the Skip-gram model. Our model includes Convolutional Neural Networks (CNN) and Bidirectional Long Short-Term Memory (Bi-LSTM) with a Conditional Random Field (CRF). It explores the integration of Kazakh linguistic features with deep learning and attention mechanisms, validating the proposed model for two script datasets on Arabic and Cyrillic letters of Kazakh with promising results.