This study tackles the WMT 2026 Low-Resource Indic Language Translation task, covering English → Assamese, Bodo, Mizo and Nagamese. Low-resource languages lack parallel corpora of sufficient size and quality to train machine translation models. To address this challenge, we propose Morphology-Aware Lexicon Filtering of Back-Translated data to generate good-quality parallel corpora. Initially, monolingual text from publicly available sources is back-translated to English with an NLLB-200-3.3B model fine-tuned on the official training data. To filter this synthetic data, we introduce MALM (Morphology-Aware Lexicon Matching), a filter that checks whether words from a bilingual lexicon are preserved in a translation, using Morfessor stems to handle rich morphology and edit-distance matching to handle borrowed words. For our primary method, training occurs in two stages: fine-tuning on filtered synthetic (back-translated) data followed by fine-tuning on the official parallel data. Whereas for the contrastive method, we are only fine-tuning on the official parallel data. We fine-tune the NLLB-200-3.3B model for Assamese, Mizo, Nagamese and the IndicTrans2-1B model for Bodo. Our primary systems ranked first among primary submissions on two Category-2 pairs viz. English → Bodo and English → Nagamese. On Nagamese, where only 2,000 official sentence pairs exist, of the two methods we used, the primary system improved by a BLEU score of 8.33.

This paper reports on the submissions of Webis to the two subtasks of ImageArg 2023. For the subtask of argumentative stance classification, we reached an F1 score of 0.84 using a BERT model for sequence classification. For the subtask of image persuasiveness classification, we reached an F1 score of 0.56 using CLIP embeddings and a neural network model, achieving the best performance for this subtask in the competition. Our analysis reveals that seemingly clear sentences (e.g., “I support gun control”) are still problematic for our otherwise competitive stance classifier and that ignoring the tweet text for image persuasiveness prediction leads to a model that is similarly effective to our top-performing model.