The paper introduces QwenSub-MT, a system that is context-aware and constraint-controlled and has been developed for the WMT 2026 Video Subtitle Translation shared task. It carried out the translation of Simplified Chinese subtitles into 6 different languages. QwenSub-MT used Qwen3.5-Instruct together with a multistage 4-bit QLoRA adaptation and treated adjacent subtitle cues as contextual blocks. In order to enhance consistency and disambiguation, the system incorporated video synopsis information, previous translations, terminology memory and selective visual grounding. A revision stage was carried out to correct semantic, terminology and formatting errors, while duration-aware length control, line-break optimization and deterministic SRT validation were used to make sure that the outputs were readable and structurally valid. 500 subtitle files were submitted in five language directions and obtained a macro-average score of 58.459, placing sixth in the overall ranking. The results show that contextual modelling and structural control are useful, but candidate diversity, independent quality estimation and stronger adaptation to the target language still needs improvement.

This paper presents HEMANT framework, the BSVLP submission to the WMT 2026 Low-Resource Arabic–Asian Machine Translation Shared Task. The system addresses ten bidirectional translation directions involving Arabic, English, Hindi, Bangla, Indonesian, and Urdu. HEMANT integrates language-specific Unicode normalization, spelling correction, named-entity recognition, knowledge-base-assisted entity translation, transliteration, multilingual transfer learning, and parameter-efficient adaptation of the NLLB-200 model using LoRA. Two directional multilingual models were trained for many-to-Arabic and Arabic-to-many translations. The official results showed considerable variation across language pairs. Arabic–Hindi produced the strongest relative performance, while English–Arabic remained the most challenging pair. In several directions, COMET scores were more competitive than BLEU and TER, suggesting that semantic adequacy was often better preserved than lexical overlap. The findings highlight both the potential and limitations of entity-aware multilingual adaptation for low-resource Arabic–Asian translation.

This paper presents a two-stage neural machine translation framework for English and selected low-resource Northeastern Indian languages. The approach incorporated transfer learning, bidirectional fine-tuning, back-translation, and LLM-assisted synthetic data augmentation. In Stage 1, the NMT model was fine-tuned on cleaned gold-standard parallel data. In Stage 2, additional synthetic sentence pairs were generated using back-translation and Qwen2.5-Instruct, followed by filtering based on language identification, length ratio, duplicate removal, named-entity preservation, semantic similarity, and round-trip consistency. The framework was evaluated on English–Assamese, English–Mizo, English–Manipuri, English–Bodo, English–Khasi, and Nagamese–English translation directions. Results show stronger performance for several Northeastern-language-to-English directions, while the contrastive system notably improved Manipuri (Bengali)–English translation from 26.56 to 30.59 BLEU. The findings indicate that controlled synthetic augmentation can improve low-resource MT, although its effectiveness remains language- and direction-dependent.

This paper describes our submission to the Indic MT 2025 shared task, where we trained machine translation systems for five low-resource language pairs: English–Manipuri, Manipuri–English, English–Bodo, English–Assamese, and Assamese–English. To address the challenge of out-of-vocabulary errors, we introduced a Named Entity Translation module that automatically identified named entities and either translated or transliterated them into the target language. The augmented corpus produced by this module was used to fine-tune a Transformer-based neural machine translation system. Our approach, termed HEMANT (Highly Efficient Machine-Assisted Natural Translation), demonstrated consistent improvements, particularly in reducing named entity errors and improving fluency for Assamese–English and Manipuri–English. Official shared task evaluation results show that the system achieved competitive performance across all five language pairs, underscoring the effectiveness of linguistically informed preprocessing for low-resource Indic MT.

This paper presents our machine translation system that was developed for the WAT2024 MultiInidc MT shared task. We built our system for the Sindhi-English language pair. We developed two MT systems. The first system was our baseline system where Sindhi was translated into English. In the second system we used Hindi as a pivot for the translation of text. In both the cases we had identified the name entities and translated them into English as a preprocessing step. Once this was done, the standard NMT process was followed to train and generate MT outputs for the task. The systems were tested on the hidden dataset of the shared task