Byte Pair Encoding (BPE) is widely recognized as an effective approach for machine translation across multiple languages. However, in morphologically rich languages such as Korean, BPE can lead to excessive segmentation, which harms word semantics and creates semantic confusion during the training. This semantic confusion ultimately leads to an overall degradation in translation quality. Subword segmentation is an effective solution to the vocabulary problem in neural machine translation. This paper proposes a method to optimize the Korean subword vocabulary for neural machine translation, based on the fact that a Korean subword vocabulary created with the BPE training algorithm contains many compositional subwords. The optimized Korean subword vocabulary demonstrates experimentally stabilized translation performance by maintaining a balanced distribution while removing unnecessary compositional subwords.

The lack of parallel corpora remains challenging for multilingual neural machine translation (MNMT), particularly for low-resource languages. This paper presents an unsupervised framework to utilize pre-trained cross-lingual encoders (XLM-R) in an unsupervised way and generates high-quality translations using monolingual corpora and bilingual dictionaries. The proposed method constructs pseudo-parallel corpora by combining word-by-word translation using bilingual dictionaries with contextual refinement via masked language modeling (MLM). To improve alignment quality, we propose a two-tier representation strategy: (1) word-level alignment that combines VecMap with FastText embeddings to address out-of-vocabulary (OOV) terms and capture morphological variations. (2) Sentence-level alignment using Adversarial Contrastive Learning (ACL) enhanced with Hard Negative Mining (HNM) to build semantically robust and discriminative sentence embeddings. Experimental results on the FLORES-101 dataset demonstrate that the proposed model outperforms the existing state-of-the-art models, with an average of +1.2 BLEU in bilingual settings and +0.9 BLEU in multilingual settings. Furthermore, the proposed model is evaluated on 4 low-resource Indian languages (e.g., Hindi, Urdu, Telugu, and Bengali), and it outperforms the state-of-the-art models with an average of +0.7 in bilingual and multilingual settings. Finally, evaluation in zero-shot and few-shot settings confirms the proposed approach’s robustness and generalization, demonstrating an effective solution for multilingual translation without using parallel corpora.

This study presents the application of free, open-source artificial intelligence (AI) techniques to advance machine translation for low-resource Turkic languages such as Kazakh, Azerbaijani, Kyrgyz, Turkish, Turkmen, and Uzbek. This machine translation problem for Turkic languages is part of a project to generate meeting minutes from speech transcripts. Due to limited parallel corpora and underdeveloped linguistic tools for these languages, traditional machine translation approaches often underperform. The goal is to reduce digital inequality for these languages and to support scalability. We investigate the effectiveness of free open-source pre-trained specialized and general-purpose AI models for morphologically rich state Turkic languages. This research includes developing parallel corpora for six Turkic languages, fine-tuning, and performance evaluation using BLEU, WER, TER, and chrF metrics. The parallel corpora for five pair languages, each of 300,000 and 500,000 sentences, were generated and cleaned. The results for corpora 500,000 parallel sentences show significant improvements compared with baseline NLLB-200 1.3B on average: BLEU increased by 23.81 points, chrF increased by 26.05 points, and WER and TER decreased by 0.36 and 33.95, respectively, after cleaning and fine-tuning. Six Turkic-language multilingual parallel corpora of 3 885 542 sentences were developed and the fine-tuning of NLLB-200 1.3B shows the following, compared with the results for 500,000 cleaned corpus: BLEU increased by 4.3 points, chrF increased by 1.7 points, and WER and TER decreased by 0.1 and 4.75, respectively These results demonstrate the high efficiency of corpus cleaning and synthetic data generation to improve the quality of machine translation for low-resource Turkic languages using AI models. These results were confirmed by external evaluation on the FLORES 200 dataset and human evaluation. The scientific contribution of this article is the development of a methodology for generating parallel corpora using a specialized AI model of machine translation and fine-tuning the specialized AI model on the created corpora, creating new multilingual parallel corpora of Azerbaijan–Kazakh, Kyrgyz–Kazakh, Turkish–Kazakh, Turkmen–Kazakh, and Uzbek–Kazakh pairs using the proposed methodology, cleaning them, and conducting fine-tuning experiments.

Domain-specific multimodal neural machine translation (DMNMT) aims to translate source language domain sentences into target language by incorporating images as additional contextual information. However, domain-specific multimodal scenarios frequently suffer from visual imbalance issues, such as one sentence corresponding to multiple images or even no images at all. Effectively integrating visual information into text to enhance domain sentence translation performance under visual imbalance issues is one of the critical challenges for DMNMT, especially for domain-related terms. To tackle these domain-specific visual imbalance problems, this article introduces a virtual domain distillation-enhanced multimodal fusion with the awareness of multiview correlations to enhance the robustness and performance of domain machine translation across various multimodal domain scenarios. We first adopt a multiview correlation-aware cross-modal distillation strategy to generate virtual domain visual scenes by extracting visual correlations among all images through multikernel representations. Subsequently, we integrate pseudo-domain visual scenes into text to improve the performance of domain-specific machine translation. Our proposed approach has the ability to capture domain visual representations across different scenarios, and contributing to more effective domain-specific translation. We conduct expensive experiments on three domain-specific and general-domain benchmark datasets. Experimental results demonstrate that our proposed approach achieves state-of-the-art (SOTA) machine translation scores on most test sets. The in-depth analysis demonstrates the effectiveness and robustness of our proposed approach for domain machine translation.

The rapid expansion of global communication, digital education, and cross-cultural collaboration has created a strong need for fast, accessible, and user-friendly translation tools. Traditional translation systems often require significant computing resources or rely heavily on cloud-based solutions. This research focuses on the development of a lightweight, client-side web-based language translator created using HTML, CSS, and JavaScript. The system is designed to support real-time multilingual translation through an external translation API, providing simplicity and accessibility across devices and platforms. This paper examines existing literature on machine translation approaches, including rule-based, statistical, and neural machine translation systems, and evaluates how these technologies have influenced modern translation APIs. A detailed explanation of the system’s structure, design methodology, implementation strategy, and performance evaluation is presented. The study demonstrates that a browser-based translator can meet the needs of everyday users without requiring a backend server. Limitations such as API dependency, translation accuracy issues for low-resource languages, and the lack of ofline support are discussed. Finally, the paper highlights potential improvements, such as integrating ofline AI models, speech support, and enhanced user personalization.

Teaching translation through the Grammar Translation Method (GTM) has been a usual practice throughout the world. In the contemporary period, especially after the AI revolution in the world, machine-based translations are also being rendered at a large scale. This study has explored the phenomenon of the use of Machine Translation in teaching translation from Urdu to English and vice versa. A questionnaire was developed for data collection, other than the execution of pre-test and post-test. There were 60 female students at a private college who were selected as respondents through purposive sampling. Students were divided into experimental groups 1& 2. Each group was taught translation through different methods: Machine Translation (MT) and juxtaposition of both MT and GTM. Data was analyzed both quantitatively and qualitatively. It was found that the combo of GTM & MT is more beneficial for teaching interlingual translations. However, students from semi-rural areas were found to be more comfortable with GTM.

Multilingual Information Retrieval (MLIR) systems have become essential tools in a digitally integrated economy. Users require pertinent information in various languages and across linguistic frontiers. A technique rooted in linear algebra and statistical semantics known as Latent Semantic Analysis (LSA) offers a solution for revealing patterns buried within the data, which may cut across languages. In this paper, we investigate the efficiency of LSA in MLIR tasks with various language pairs compared to traditional vector space models and the machine translation approach. Using the Europarl and CLEF corpora, and employing mean average precision (MAP), precision at 10 (P@10), and normalized Discounted Cumulative Gain (DCG), we demonstrate that LSA facilitates reasonable cross-lingual alignment under specific conditions. Moreover, we assess the model's performance considering changes in the number of latent dimensions and various preprocessing techniques applied before the central processing.

Recent advances in large language models have highlighted the need for high-quality multilingual medical datasets. Although Japan is a global leader in computed tomography (CT) scanner deployment and use, the absence of large-scale Japanese radiology datasets has hindered the development of specialized language models for medical imaging analysis. Despite the emergence of multilingual models and language-specific adaptations, the development of Japanese-specific medical language models has been constrained by a lack of comprehensive datasets, particularly in radiology. This study aims to address this critical gap in Japanese medical natural language processing resources, for which a comprehensive Japanese CT report dataset was developed through machine translation, to establish a specialized language model for structured classification. In addition, a rigorously validated evaluation dataset was created through expert radiologist refinement to ensure a reliable assessment of model performance. We translated the CT-RATE dataset (24,283 CT reports from 21,304 patients) into Japanese using GPT-4o mini. The training dataset consisted of 22,778 machine-translated reports, and the validation dataset included 150 reports carefully revised by radiologists. We developed CT-BERT-JPN, a specialized Bidirectional Encoder Representations from Transformers (BERT) model for Japanese radiology text, based on the "tohoku-nlp/bert-base-japanese-v3" architecture, to extract 18 structured findings from reports. Translation quality was assessed with Bilingual Evaluation Understudy (BLEU) and Recall-Oriented Understudy for Gisting Evaluation (ROUGE) scores and further evaluated by radiologists in a dedicated human-in-the-loop experiment. In that experiment, each of a randomly selected subset of reports was independently reviewed by 2 radiologists-1 senior (postgraduate year [PGY] 6-11) and 1 junior (PGY 4-5)-using a 5-point Likert scale to rate: (1) grammatical correctness, (2) medical terminology accuracy, and (3) overall readability. Inter-rater reliability was measured via quadratic weighted kappa (QWK). Model performance was benchmarked against GPT-4o using accuracy, precision, recall, F1-score, ROC (receiver operating characteristic)-AUC (area under the curve), and average precision. General text structure was preserved (BLEU: 0.731 findings, 0.690 impression; ROUGE: 0.770-0.876 findings, 0.748-0.857 impression), though expert review identified 3 categories of necessary refinements-contextual adjustment of technical terms, completion of incomplete translations, and localization of Japanese medical terminology. The radiologist-revised translations scored significantly higher than raw machine translations across all dimensions, and all improvements were statistically significant (P<.001). CT-BERT-JPN outperformed GPT-4o on 11 of 18 findings (61%), achieving perfect F1-scores for 4 conditions and F1-score >0.95 for 14 conditions, despite varied sample sizes (7-82 cases). Our study established a robust Japanese CT report dataset and demonstrated the effectiveness of a specialized language model in structured classification of findings. This hybrid approach of machine translation and expert validation enabled the creation of large-scale datasets while maintaining high-quality standards. This study provides essential resources for advancing medical artificial intelligence research in Japanese health care settings, using datasets and models publicly available for research to facilitate further advancement in the field.

La langue dialectale arabe est récemment passée d'une forme purement orale à une forme écrite informelle sur les réseaux sociaux en raison de l'expansion de l'accès à Internet. Ainsi, les internautes transmettent d’énormes quantités de données dialectales très bruitées et non structurées dans leurs dialectes natifs. Cette situation entrave la compréhension mutuelle entre les utilisateurs et présente des défis dans l'utilisation des approches traditionnelles de traitement du langage , car la plupart des ressources sont conçues pour les langages formels. Dans ce contexte, nous nous concentrons dans cette étude sur traduire des textes d'un réseau social écrits en dialecte tunisien (TD) vers l'anglais (EN), une langue formelle comprise dans le monde entier. Cependant, les caractéristiques grammaticales uniques du dialecte tunisien, telles que l'absence d'un corpus représentatif, rendent la traduction automatique utilisant la méthode Deep Learning difficile. Pour cela, notre objectif premier est de traduire le dialecte tunisien vers l'anglais en développant un corpus parallèle et un modèle de traduction automatique approfondi. Pour ce faire, nous évaluons plusieurs architectures neuronales et étudions les effets de l’utilisation de la langue standard arabe comme langue pivot pour traduire TD. Nous avons obtenu 67,89 % en tant que score BLEU en utilisant l'architecture Transformer.

The purpose of this study is a comprehensive quadrilingual comparative analysis of machine and human translation approaches to culturally loaded information in contemporary Chinese literature. The object of analysis is the translation of culturally loaded linguistic units from Yu Hua’s novel “Brothers” (兄弟) from Chinese into English, French, and German. This study compares human translations by professional translators with machine translations generated by ChatGPT and DeepL artificial intelligence systems across all three target languages. The research methodology employs systematic comparative analysis of material culture terms, stylistic units, and socioculturemes across four languages. Through comprehensive quadrilingual comparative analysis, this research concludes that cultural transfer strategies vary significantly between target languages and cultural contexts, while AI translation systems demonstrate consistent limitations across all target languages, particularly when dealing with emotionally nuanced content containing extensive background cultural and historical information. The study contributes to translation studies by providing systematic empirical evidence for the continuing necessity of human cultural expertise in literary translation while identifying specific areas where AI systems might serve as auxiliary tools.

In the field of neural machine translation (NMT), improving translation quality remains a significant challenge, especially for low-resource language pairs. Inspired by the prompt-based method, we propose a novel approach called prompt-enhanced pseudo-dynamic smoothing (PEPDS). This method leverages prompts to optimize the attention mechanism and enhance the model’s learning capacity during the training process. Our approach provides additional contextual information through the use of prompts and enhances the model’s ability to handle long-distance dependencies and complex grammatical structures. The PEPDS framework utilize linguistic insights to create effective contextual cues. We focus on four parts of speech in English sentences: nouns, adjectives, adverbs and verbs, to generate prompts that enrich the contextual information available to the model. This method can be viewed as a form of inside knowledge enhancement, particularly beneficial for low-resource scenarios and long sentences. We apply the prompt method during the inference phase, further enhancing the model’s translation performance. We evaluate PEPDS across multiple language pairs, demonstrating significant improvements in translation quality, especially for low-resource language pairs and complex sentences. Experimental results, attention pattern studies and translation case studies all demonstrate the effectiveness of our method.

Statistical machine translation (SMT) used parallel corpora and statistical models, to identify translation patterns and probabilities. Although this method had advantages, it had trouble with idiomatic expressions, context-specific subtleties, and intricate linguistic structures. The subsequent introduction of deep neural networks such as recurrent neural networks (RNNs), long short-term memory (LSTMs), transformers with attention mechanisms, and the emergence of large language model (LLM) frameworks has marked a paradigm shift in machine translation in recent years and has entirely replaced the traditional statistical approaches. The LLMs are able to capture complex language patterns, semantics, and context because they have been trained on enormous volumes of text data. Our study summarizes the most significant contributions in the literature related to LLM prompting, fine-tuning, retrieval augmented generation, improved transformer variants for faster translation, multilingual LLMs, and quality estimation with LLMs. This new research direction guides the development of more efficient and innovative solutions to address the current challenges of LLMs, including hallucinations, translation bias, information leakage, and inaccuracy due to language inconsistencies.

Purpose This study aimed to analyze the readability of Google Translate (GT) and Yandex Translate (YT) translation results on dialogue texts containing cultural terms from the book Antologi Cerita Anak Indonesia (ACAI). This study evaluated the effectiveness of the Neural Machine Translation (NMT) approach in GT and the Hybrid Machine Translation (HMT) approach in YT in conveying text meanings clearly and comprehensibly to readers.Method This research employed a cloze test involving 28 participants aged 18-24 years, along with a questionnaire to assess user preferences regarding GT and YT translation results. Text readability was analyzed using the Flesch-Kincaid Grade Level and Gunning Fog Index to measure the linguistic complexity of the translations.Results/Findings The study results show that GT's readability reaches 81.1%, while YT's readability is 74.5%, both categorized as the independent level according to Rankin &amp; Culhane's (1969) criteria. Additionally, 80% of the 20 questionnaire respondents stated that GT's translations were clearer than those of YT. Analysis using the Flesch-Kincaid Grade Level and Gunning Fog Index shows that the readability level of GT and YT translations is classified as advanced suitable for readers with a minimum education level equivalent to a bachelor's degree.Conclusion This study showed that GT has a higher readability level than YT, which might be because of its use of NMT, producing more natural sentence structures. Meanwhile, YT, which also relied on SMT, translates based on statistical patterns, making its translations more rigid. Although both systems could produce comprehensible translations, they still struggled with accurately translating cultural terms without additional context. Therefore, human involvement remained essential to improving accuracy and contextual appropriateness in machine translation.

The goal of natural language processing (NLP), which has recently gained popularity, is to improve the capacity of computers to comprehend and interact with human language. Consequently, to converse using natural language, it is crucial that spoken language be grammatically correct, especially for Geéz language. Geéz language sentences must follow certain norms of agreement in terms of number, person, gender, tense, and other factors to be considered grammatically correct. If the input sentence in Geéz language is improper, then it can have problems with subject-verb agreement, object-verb agreement, adjective-noun agreement, and adverb-verb agreement. The goal of the proposed work is to provide a neural machine translation approach for detecting and correcting grammar errors in Geéz sentences. We have prepared manually 11,490 Geéz parallel corpuses (Geéz language grammatically incorrect and grammatically correct sentences). After we have prepared a parallel Geéz sentence, we have used normalization, tokenization, padding, and one hot encoding as preprocesses. We have used two deep learning algorithms, including a bidirectional long short-term memory encoder-decoder and a long short-term memory encoder-decoder, for training the proposed model. Keras and TensorFlow were used for importing the required libraries, and we used the Python 3.7 environment for implementation. Two test cases are used for the evaluation technique. The first one is for the long short-term memory encoder-decoder model, and the second one is for the bidirectional long short-term memory encoder-decoder model. Finally, the bidirectional long short-term memory encoder-decoder model achieved best results with an accuracy of 82%, recall of 82%, precision of 85%, and F1-measure of 83% with balanced error type classes.

Cross-lingual sentiment analysis (CLSA) has become increasingly important in natural language processing and machine learning, enabling the understanding of opinions across diverse linguistic communities, particularly in low-resource languages (LRLs). Despite growing attention, persistent challenges such as limited annotated data, semantic misalignment, and cultural variation in sentiment expression continue to hinder progress. This systematic literature review (SLR) examines recent developments by analyzing the tasks, methods, and challenges reported in CLSA studies focused on LRLs. Following the PRISMA 2020 framework, a comprehensive search was conducted across major databases, including Scopus, IEEE Xplore, SpringerLink, Elsevier, and Google Scholar, covering studies published between 2021 and 2025. After applying inclusion and exclusion criteria, 27 studies were selected for analysis. The findings reveal that while polarity detection remains the dominant sentiment analysis task, emerging directions such as aspect-based sentiment analysis (ABSA), emotion detection, and hate speech recognition are gaining traction. Methodologically, most studies rely on multilingual pre-trained language models (PLMs), supplemented by machine translation, transfer learning, few-shot learning, and hybrid approaches. However, key challenges remain, including the scarcity of high-quality datasets, instability of few-shot performance, difficulties in handling dialectal variation, bias in PLMs, and the lack of standardized evaluation benchmarks. This review concludes by emphasizing the need for more culturally grounded tasks, adaptive hybrid frameworks, and fairness-aware evaluation practices to build robust cross-lingual frameworks and richer linguistic resources for underrepresented languages.

Neural machine translation has transformed automated translation, surpassing traditional methods with its significant accuracy improvements. However, despite its successes, NMT still encounters several challenges, such as handling low-resource languages, maintaining contextual coherence, and addressing ambiguities in translation. This research presents a novel hybrid NMT model to overcome these limitations. It combines the strengths of traditional translation methods with modern deep learning approaches. We conduct a comprehensive comparative analysis of our hybrid model against existing machine translation approaches, including rule-based machine translation (RBMT), SMT, and state-of-the-art NMT systems. Evaluation metrics BLEU is utilized to assess the performance across English-Hindi,English-Marathi,English-Bengali language pairs and domains. Our results demonstrate that the hybrid NMT model achieves superior accuracy and fluency in translation tasks, particularly for low-resource languages and complex sentence structures. This research highlights the potential of combining different machine translation approaches and findings suggest that integrating these methods can significantly improve translation quality. The findings offer valuable insights for future research and development of more robust and versatile translation systems. Our results demonstrate that the hybrid model offers significant improvements in translation accuracy, making it a promising approach for multilingual machine translation tasks. NMT surpasses both RBMT and SMT with a BLEU score of 35.6, highlighting its effectiveness in managing context and semantics. Qualitative assessments suggest that the hybrid model effectively minimizes common translation errors, making it a robust solution for multilingual machine translation tasks. Hybrid Neural Machine Translation (NMT) models are increasingly being applied in real-world applications where the combination of rule-based, statistical, and neural approaches offers distinct advantages.

Neural machine translation (NMT) approaches driven by artificial intelligence (AI) has gained more and more attention in recent years, mainly due to their simplicity yet state-of-the-art performance. Despite NMT models with attention mechanism relying heavily on the accessibility of substantial parallel corpora, they have demonstrated efficacy even for languages with limited linguistic resources. The convolutional neural network (CNN) is frequently employed in tasks involving visual and speech recognition. Implementing CNN for MT is still challenging compared to the predominant approaches. Recent research has shown that the CNN-based NMT model cannot capture long-term dependencies present in the source sentence. The CNN-based model can only capture the word dependencies within the width of its filters. This unnatural character often causes a worse performance for CNN-based NMT than the RNN-based NMT models. This study introduces a simple method to improve neural translation of a low-resource language, specifically Urdu-English (UR-EN). In this paper, we use a Fully Convolutional Neural Network (FConv-NN) based NMT architecture to create a powerful MT encoder for UR-EN translation that can capture the long dependency of words in a sentence. Although the model is quite simple, it yields strong empirical results. Experimental results show that the FConv-NN model consistently outperforms the traditional CNN-based model with filters. On the Urdu-English Dataset, the FConv-NN model produces translation with a gain of 18.42 BLEU points. Moreover, the quantitative and comparative analysis shows that in a low-resource setting, FConv-NN-based NMT outperforms conventional CNN-based NMT models.

Effective communication across languages remains a critical challenge, particularly in low-resource settings where conventional machine translation approaches falter due to sparse data and limited quality feedback. This paper presents a holistic framework to enhance reinforcement learning (RL) based machine translation systems tailored for such environments. We address the trifecta of challenges: sparse feedback on translation quality, ethical implications in algorithmic decision-making, and the imperative to adapt models to nuanced linguistic domains. This approach integrates advanced techniques in sparse reward handling, ensuring RL models learn efficiently despite limited feedback. Ethical considerations drive our methodology, emphasizing fairness, bias mitigation, and cultural sensitivity to uphold ethical standards in AI-driven translations. Additionally, domain-specific adaptation strategies are explored to tailor models to diverse linguistic contexts, from technical jargon to colloquialisms, enhancing translation accuracy and relevance. Through a rigorous experimental framework, including evaluation metrics like BLEU score and user feedback, we demonstrate substantial improvements in translation quality and ethical compliance compared to traditional methods. This research contributes to the evolution of robust, inclusive translation technologies pivotal for fostering global understanding and equitable access to information. This paper not only addresses current challenges but also sets a precedent for future research in AI ethics and machine learning applications, advocating for responsible innovation in crosscultural communication technologies

LLMs-based machine translation for E-commerce

Current multimodal neural machine translation (MNMT) approaches primarily focus on ensuring consistency between visual annotations and the source language, often overlooking the broader aspect of multimodal coherence, including target–visual and bilingual–visual alignment. In this paper, we propose a novel approach that effectively leverages target–visual consistency (TVC) and bilingual–visual consistency (BiVC) to improve MNMT performance. Our method leverages visual annotations depicting concepts across bilingual parallel sentences to enhance multimodal coherence in translation. We exploit target–visual harmony by extracting contextual cues from visual annotations during auto-regressive decoding, incorporating vital future context to improve target sentence representation. Additionally, we introduce a consistency loss promoting semantic congruence between bilingual sentence pairs and their visual annotations, fostering a tighter integration of textual and visual modalities. Extensive experiments on diverse multimodal translation datasets empirically demonstrate our approach’s effectiveness. This visually aware, data-driven framework opens exciting opportunities for intelligent learning, adaptive control, and robust distributed optimization of multi-agent systems in uncertain, complex environments. By seamlessly fusing multimodal data and machine learning, our method paves the way for novel control paradigms capable of effectively handling the dynamics and constraints of real-world multi-agent applications.