This article investigates offline recognition of handwritten Kazakh text in the Latin script using a convolutional recurrent neural network. The relevance of the study is determined by the transition of the Kazakh language to the Latin alphabet and the need to automate the processing of handwritten documents. The proposed model consists of a convolutional neural network feature extractor, two bidirectional long short-term memory layers, and a Connectionist Temporal Classification decoder. The convolutional layers extract visual features from word images, the bidirectional recurrent layers model the sequential relationships between characters, and CTC enables end-to-end training without explicit character-level segmentation. A specialized dataset named KazEsim, containing 20,000 handwritten Kazakh name images, was created and divided into writer-independent training, validation, and test subsets. Experimental results showed a character accuracy rate of 96.5% and a word accuracy rate of 92.3%. Compared with a conventional CNN baseline, the proposed CRNN model improved character accuracy by 6.1 percentage points and word accuracy by 9.2 percentage points. The proposed model also outperformed the fine-tuned TrOCR-small comparative baseline while requiring fewer parameters and lower inference latency. These findings demonstrate the effectiveness of CNN–BiLSTM–CTC sequence modeling for offline recognition of handwritten Kazakh words in the Latin script.
A. Shormakova, M. Mansurova, Beibitkhan Yerkegul et al.· Computers· 0 citations
Large language models can emit fluent text while leaving intermediate semantic structure implicit. We study whether explicit event-role and logical-primitive side-state can improve a pretrained decoder without damaging its language behavior. We introduce PAT-ER, a decoder architecture with a normal token stream, an event-role register stream, and a primitive register stream. The primitive stream is motivated by the view that logical primitives answer characteristic semantic questions, such as what licenses a conclusion, what conflicts with it, or why evidence is insufficient. Across eight seeds on the same Qwen3-0.6B backbone, replacing token-pooled auxiliary heads with typed PAT-ER registers improves primitive macro-F1 by 0.209 (95% CI [0.182, 0.237]) and role-to-primitive macro-F1 by 0.091 (95% CI [0.074, 0.110]) with no language-model loss cost. A generic-register control shows that this is not merely the effect of adding latent registers: typed PAT-ER improves over generic registers by 0.116 primitive macro-F1 and 0.110 role-to-primitive macro-F1, with both confidence intervals excluding zero. A warm-started model then recovers pretrained language quality (LM loss 1.344 versus 2.555 for the frozen-backbone register model) while retaining most side-state behavior. Finally, a decoupled interface mode produces robust schema-grounded function calls on 242 held-out prompts (Hermes parse 0.952, exact arguments 0.981, JSON validity 1.000, IDK F1 1.000) while base-mode side-state metrics remain byte-identical to the warm-start baseline. The model is not a theorem prover and does not achieve perfect unseen tool-name copying; the contribution is a measured architecture signal and a usable, guarded interface.
N. Kadyrbek, Madina Mansurova· Machine Learning and Knowled...· 0 citations
