A Bengali-Hindi-Telugu parallel corpus for enhanced literary machine translation.

Learning fair representation for fine-tuning pre-trained language models.

PADI-Location-AR-EN: A normalized Arabic-English spatial entity dataset for epidemiological surveillance.

Assessment Criteria of Machine Translation Quality Using Error Analysis in Literary Works: A Case Study of Khalil Gibran’s al-Arwāḥ al-Mutamarridah (The Rebellious Spirits) (2017)

Artificial intelligence (AI) is increasingly prevalent. Patients and clinicians may use AI-based tools in many different languages. To investigate AI translation tools for descriptions of genetic conditions and how AI identification of genetic conditions is affected by translations. We used Neural machine translation (NMT) and large language-model (LLM) translation to translate descriptions of 40 genetic conditions into 191 and 93 languages, respectively. Excluding translations retaining English medical terms verbatim, we respectively focused on 139 and 70 languages. After assessing translations, we assessed the ability of 3 proprietary and 3 open-weight general LLMs to identify conditions in the translations. We analyzed how accuracy was affected by the conditions' prevalence in the literature, and attributes of the languages (the script, language family, and prevalence of the language in training sources). We also investigated adaptive translation for select languages. We found significant differences in condition identification based on the translation method, condition, language, and prediction model. The accuracy of some models was more affected than others by factors like the conditions' literature prevalence, language script, family, and language prevalence. Adaptive translation for select languages did not improve translations or diagnostic accuracy with the 3 tested LLMs. However, further analysis with 1 language showed that this approach was more effective with smaller LLMs. AI-based translation has variable performance, which can affect the ability of AI models to recognize genetic conditions. These findings should inform safe medical AI use to support consistent performance in different languages.

Abstract This article focuses on three Catalan linguistic items or structures that were previously used in the verification of the gravitational pull hypothesis: the modal marker caldre , the imperfective-perfective aspect distinction, and a number of light verb constructions with the verb fer conveying emotional states. Since translational effects had been found to occur in connection with them for the English-Catalan language pair, they are taken to be good candidates to test the machine translationese hypothesis, according to which patterns of over- or underrepresentation in human translations (when compared to non-translations in the target language) tend to be exacerbated in machine translation. Two of the three hypotheses put forward in the article are borne out by the data. The study draws on four components of the COVALT corpus. It also throws light on other aspects of the items under scrutiny, such as source text trigger distribution. The findings are relevant in that they highlight concrete (as opposed to abstract) ways in which machine translation departs from idiomatic usage as reflected by distributional frequencies. This tends to happen in human translation too, but machine translation carries the tendency further.

Machine translation has increasingly shifted toward Neural Machine Translation (NMT) because of its ability to handle input and output sequences of varying lengths. The incorporation of attention mechanisms in NMT systems enables the model to focus on the most relevant parts of the source sentence, rather than relying solely on a fixed representation of the entire input. While NMT improves translation quality by addressing long-range dependencies and contextual understanding, it also requires a large parallel corpus for training, which is a challenge for languages with less resources. The main focus of this research is to give solution for the unique challenges of translating Ayurvedic texts using NMT. Ayurvedic texts have collection of special and scientific words related to medicines and treatments. This makes the translation process more complex and needs very efficient approach for accurate translations. Also, the content of ayurvedic text books is in the form shlokas which is formed using very complex and compound words. In order to simplify the translation process efficiently this work uses a sandhi splitter module and an Anvaya Generator/ word reordering module. In order to develop NMT system for low resource language pair Sanskrit-Malayalam, there is a need of developing a parallel corpus especially for Ayurvedic text books. Also, as the NMT model is proposed for translation it requires a minimum amount of parallel data in the corpus. So, a number of general domain Sanskrit text books with verses, called shlokas, were also considered for developing parallel corpora. The authors developed a parallel corpus for Anvaya Generator, sandhi splitter and translation. Mainly four NMT models were developed trained and tested especially for shlokas as input. The two models are basic transformer model with attention and an encoder-decoder model using Long-Short term Memory (LSTM) with attention. The other two are developed by adding two modules called Sandhi Splitter and Anvaya Generator in the pre-processing stages of the earlier models- Transformer based model and LSTM based model. The limitations of low resources and richness in grammatical structure of Sanskrit- Malayalam language pair are overcome by the concepts of deep learning and the additional modules used in preprocessing stages for developing the models. The models were tested with and without sandhi splitter and Anvaya Generator modules. The transformer-based model integrated with sandhi splitter and Anvaya Generator system achieved a higher average BLEU score of 73.11 and a uni-gram BLEU score of 76.93 for Sanskrit verses to Malayalam translation.

English plays a crucial role in higher education, as many academic resources are presented in English. However, many students still face difficulties in understanding English texts due to unfamiliar vocabulary and complex sentence structures. To overcome these challenges, students often use machine translation tools, particularly Google Translate. This study aims to investigate undergraduate students’ perceptions of using Google Translate in learning English within academic contexts. This study applies a mixed-method design using an explanatory sequential approach. Quantitative data were collected through a questionnaire consisting of 25 Likert-scale items distributed to 34 undergraduate students in Medan. Qualitative data were obtained from semi-structured interviews with five selected participants to provide deeper insights. The findings reveal that Google Translate is primarily used as a vocabulary support tool. A total of 73.53% of students use it to check word meanings, while 97.06% consider it useful for learning English. However, its use decreases in more complex tasks such as paragraph and essay writing. Students also recognize its limitations, especially in grammar and contextual accuracy. Despite this, students do not entirely rely on the tool. Instead, they use it critically by verifying the results with other sources and combining them with their own language knowledge. In conclusion, Google Translate serves as a supportive tool rather than a primary learning resource. Its effectiveness depends on strategic and critical use.

Enhancing end-to-end speech translation via multi-stage knowledge distillation.

• FFTGate: a novel activation function that integrates temporal and spectral activation histories. • FFT-based frequency-domain gating modulates neuron activations using spectral features. • Target-adaptive regularization of the activation’s trainable parameters improves generalization. • FFTGate improves generalization on image classification and neural machine translation tasks. • FFTGate outperforms 18 existing activation functions on benchmark datasets. Activation functions are crucial in determining how deep neural networks process information and adapt during training. Although widely adopted activation functions such as ReLU, GELU, and Swish demonstrate strong empirical performance, they rely primarily on pre-activation inputs, ignoring temporal activation histories that capture how neuron activations evolve over time and spectral features that capture frequency-domain patterns in activation histories. This can limit their adaptability, particularly in dynamic or time-varying tasks. To address this, we propose FFTGate , a novel activation function that integrates temporal and spectral activation information from neuron activation histories into the activation process. Specifically, FFTGate first captures per-channel activation histories using a sliding window of a defined length. Next, it applies the Fast Fourier Transform to extract spectral features from the captured activation histories, which are then used to form an FFT-based gating mechanism subsequently used to modulate each neuron’s pre-activation input. This enables networks to adapt more effectively to evolving inputs, thereby improving generalization. FFTGate is evaluated on both vision and language domains. Experimental results show that FFTGate consistently outperforms 18 established and recently proposed activation functions on small-scale image classification tasks (CIFAR-10, CIFAR-100; Top-1 accuracy gain of +0.74 %) and large-scale image classification task (ImageNet-1K; Top-1 accuracy gain of +0.36 %), demonstrating strong generalization on datasets of different scales. Furthermore, FFTGate achieves even greater improvements in neural machine translation, with BLEU score gains of +2.94 on Kaggle German–English, +1.08 on IWSLT’13 French–English, and +0.88 on IWSLT’13 English–French over the best-performing baselines. These results demonstrate that FFTGate enhances generalization and stability in both image-based and time-varying tasks. Code and results are available at: https://github.com/Ndupuechi/FFTGate .

Research challenges and future directions in transformer-based neural machine translation

Generative artificial intelligence does perform very well in machine translation, but in the legal scene of English-Chinese translation, controllability and professional reliability are not enough. Legal translation itself is characterized by intensive terminology, complex sentence patterns and strict logical requirements, so the requirements for accuracy and efficiency are extremely high. Because of this, the actual effect of this kind of AI in the legal field has not been fully verified, especially in consistency and coherence at the discourse level. In addition, there is a lack of in-depth research on the extent to which the project can improve the quality of translation. This study compares the performance of ChatGPT-5 and DeepL in English-Chinese legal translation, and examines the differences in the effects of different prompt strategies. Five prompt strategies are designed, from simple to complex, to deal with legal texts. There are three evaluation dimensions : lexical richness, semantic accuracy and discourse coherence, and then the statistical analysis method is used to find out the significant differences. The results show that the prompt design has a great influence on the translation quality. The more structured the prompt is, the more consistent and accurate the output is. In terms of semantic accuracy and coherence, GPT-5 is similar to DeepL. However, under the structured prompt, the lexical richness of GPT-5 is better. On the whole, this study promotes the development of legal machine translation, reveals the different effects of prompt engineering, and shows that structured prompts can improve lexical richness, while the two systems have their own advantages. Based on this, an evidence-based prompt optimization framework is proposed.

Given the rapid integration of digital technologies in language education and the unique linguistic complexity of Arabic, urgent research is needed to empirically determine how learners’ digital literacy competencies influence Arabic as a Foreign Language acquisition. This mixed-methods study investigates the relationship between digital literacy levels and language proficiency among 124 non-native Arabic learners enrolled in university-level AFL programs. Quantitative data were collected using a validated digital literacy scale adapted from the European Digital Competence Framework (DigComp) and a standardized Arabic proficiency test aligned with ACTFL benchmarks. Qualitative interviews and classroom observations explored learner perceptions and instructional practices. Results revealed statistically significant correlations between digital literacy skills-particularly information navigation, multimodal communication, and content creation-and reading comprehension, writing accuracy, and listening proficiency. Findings indicate that digitally literate learners demonstrate greater autonomy, increased exposure to authentic media, and more effective utilization of learning tools. Key challenges include unequal access to technology, limited pedagogical integration of digital tools, and learner dependency on machine translation. The study underscores the necessity of embedding digital literacy instruction within AFL curricula and provides recommendations for teacher training, curricular design, and future research

The paper presents a neural machine translation system with a context-sensitive mechanism. It uses a recursive self-encoder to learn the representations of English sentences, which are then contextualized by topic distribution. Furthermore, the parameters of the model are optimized by applying a deep reinforcement learning algorithm, which enhances the accuracy of the model further. Experiments are performed on different datasets to assess the performance of the proposed model. It was found that the model exhibits its highest level of translation performance at the cosine similarity threshold of 0.9, and that there are notable improvements in the quality of translations following fine-tuning of the parameters utilizing deep reinforcement learning. Bleu scores increase by 2.00 - 4.84 points, which indicates the efficiency of the fine-tuning stage. Besides, the incorporation of the context-sensitive bilingual-constrained recursive autocoder boosts the bleu scores up to 5.23 - 7.76 points over the baseline and variant models. On the whole, the addition of deep reinforcement learning and the context-sensitive method makes the model much more effective at producing English translations that are correct.

In the context of the continuous growth in cross-border digital communication and multilingual information processing demands, the modeling and difference identification of ASEAN English multilingual corpora have become an important topic in intelligent translation research. This paper constructs a technical framework that integrates corpus collection, language identification, sentence-level alignment, subword segmentation, Transformer modeling, and intelligent comparative analysis. It introduces multi-head self-attention, difference scoring function, vector representation learning, and feature fusion mechanisms, and uses joint loss to achieve collaborative optimization of translation generation and difference discrimination. Experiments show that the model in this paper achieves a BLEU score of 39.63, a TER of 0.347, a semantic similarity of 0.879, and a difference identification accuracy of 92.3%; compared with Transformer, BLEU increases by 7.3% and the accuracy improves by 3.6 percentage points. This method can effectively reveal the translation differences between ASEAN English and Thai, Vietnamese, Indonesian, and Malay, and has practical significance for regional English variant computing research and intelligent translation analysis. Povzetek: This study focuses on the intelligent comparison of multilingual corpora of ASEAN English, and constructs a neural machine translation framework that integrates Transformer semantic modeling, difference scoring, feature fusion, and joint optimization. Based on experiments with four types of corpora, the model's BLEU score reaches 39.63, TER drops to 0.347, and the accuracy of difference identification reaches 92.3%, which can well reveal the differences in cross-language vocabulary, structure, and semantics.

This research aims to conduct an exploratory study focused on analysing post-editing (PE) in scientific and technical texts. The specific objectives of the study are: 1) to analyse interference in the language of writing (Franco Aixelá, 2013); 2) to classify the characteristics of scientific and technical language (Terán Rueda, 2016); 3) to compare the performance of machine translation (MT), namely Google Translate and DeepL, into Spanish; and 4) to study the PE proposals made by learners and identify their mistakes and successes. To achieve the objectives we have set, we draw on real-world experience from the specialised translation classroom at the University of Lleida. We used a corpus composed of two types of medium-high difficulty English texts from a US toy manufacturer and a Chinese photovoltaic inverter manufacturer, which were provided by a translation agency. This dual choice responds to the concern about how MT responds to informative and instructional texts as opposed to more creative ones, where culturemes come into play. Finally, the results obtained allow us to conclude that DeepL is currently the best-performing tool for scientific and technical texts. This study sheds light on the importance that PE should begin to take on in curricula.

The article analyzes how artificial intelligence reflects regional linguistic variation in Spanish in machine translations and examines the findings and their implications for translator training. To do this, a translation task is defined that collects texts from trend reports in English related to technology, clothing, and food. The corpus is translated using DeepL, Google Translate, and ChatGPT in April 2025. A qualitative analysis is presented that identifies errors, linguistic choices, and regional patterns. The results show that, by offering only a single option for translating into Spanish, the comprehension and communicative effectiveness of the target text are compromised. The article reflects on the critical use of machine translation in translator training and its potential as a teaching resource.

In today’s increasingly globalised world, translators cannot avoid collaborating with technological tools that offer a wealth of possibilities. The advent and proliferation of generative AI and Large Language Models have brought several advantages to translators, particularly in terms of saving time. However, it is important not to overestimate these systems’ capabilities, bearing in mind that human intervention is still necessary to ensure the highest quality. This study investigates how two groups of students collaborated with each other and with different digital tools to understand the benefits that translators can derive from using traditional computer-assisted translation (CAT) tools, machine translation, and LLMs, and the most effective way to organise collaborative translation work. It also examines the most common challenges associated with using these tools, providing an initial guide for professionals and semi-professionals on how to approach translation tasks with them.

Large Language Models (LLMs) have revolutionized Natural Language Processing, including machine translation (MT), achieving unprecedented performance. However, this progress masks underlying asymmetries in training data and model architecture that impact multilingual translation quality. This paper introduces LingualX64, a novel dataset spanning 64 languages, designed to evaluate the extent to which these asymmetries affect LLM translation performance, particularly under zero-shot conditions. LingualX64 is constructed to minimize data overlap with existing LLM training corpora and to provide a balanced representation of diverse linguistic features, enabling a more robust assessment of cross-linguistic generalization. Our evaluation reveals significant performance disparities across languages, highlighting the impact of data scarcity and linguistic complexity on translation quality. These findings underscore the need for strategies to mitigate asymmetries in LLM training and model design to achieve more equitable and robust multilingual translation capabilities. LingualX64 provides a valuable benchmark for researchers and developers seeking to address these challenges and unlock the full potential of LLMs for global communication.

Multilingual Large Language Models (LLMs) struggle with cross-lingual tasks due to data imbalances between high-resource and low-resource languages and the monolingual bias in pre-training. Existing methods, such as bilingual fine-tuning and contrastive alignment, improve cross-lingual performance but often require extensive parallel data or suffer from instability. To address these challenges, we introduce a Cross-Lingual Mapping Task in the pre-training phase, which enhances cross-lingual alignment without compromising monolingual fluency. Our approach bi-directionally maps languages within the LLM’s embedding space, improving both language generation and comprehension. We further introduce a Language Alignment Coefficient to robustly quantify cross-lingual consistency, even in limited-data scenarios. Experimental results on machine translation (MT), cross-lingual natural language understanding (CLNLU), and cross-lingual question answering (CLQA) show that our model achieves up to 11.9 BLEU score gains in MT, an increase of 6.72 in CLQA BERTScore-Precision and more than a 5% increase in CLNLU accuracy over strong multilingual baselines. Our findings highlight the potential of embedding cross-lingual objectives into pre-training, improving multilingual LLMs.