International Scientific Journals
of Scientific Technical Union of Mechanical Engineering "Industry 4.0"

Fundamental tasks in Natural Language Processing include Part-of-Speech tagging, lemmatization, and dependency parsing, which provide crucial linguistic information essential for a wide range of NLP applications. POS tagging assigns each word in a sentence its grammatical category. Lemmatization identifies the dictionary form of each word, considering its contextual usage within the sentence. Dependency parsing, on the other hand, determines the structural relationships between words, producing dependency trees that capture the grammatical organization of sentences. In this paper, we introduce ALB-Stanza, a neural pipeline parser designed for sentence segmentation, tokenization, POS tagging, morphological feature annotation, lemmatization, and dependency parsing specifically for the Albanian language. To train the ALB-Stanza parser model, we used our own corpora annotated according to the Universal Dependencies schema and the Stanza neural pipeline. The model was evaluated on unseen data, demonstrating its effectiveness in accurately predicting POS and morphological feature tags, lemmas, and dependency relations for text in Albanian language.

OCR technology is widely used in various applications, including document digitization, data extraction, and document management systems. The OCR technology has seen significant advancements in recent years, especially with the integration of machine learning and artificial intelligence techniques. These advancements have led to substantial improvements in accuracy, particularly for standard fonts and clear document images. However, challenges still exist, especially when dealing with low-quality images, noisy images, handwritten text, or documents with unusual fonts. Some documents like ID cards, driving licenses, etc. use some security features like deliberate errors, OVI (Optical Variable Ink), Rainbow print, Guilloche pattern, fine line, and microprint to protect the documents from being counterfeited. These security elements also generate noise on the image to perform the OCR. In this paper, we present a way to enhance the OCR accuracy for ID-1 documents with security features through machine learning-driven image optimization. Albanian driving license images on the personalization process are used as a dataset to train the model. During the training process, the model of the ID-1 card is presented with a dataset containing input features (such as images, texts, or numerical data) along with corresponding labels or outcomes. After training, the model and the implemented algorithm to optimize the image for the OCR process are implemented in real-life application.