A synthetic training pipeline and a 5-stream Tesseract ensemble voted under a lexicon-anchored, ROVER-style scheme adapted for a low-resource setting, and results on a 422-paragraph benchmark against a fine-tuned-Tesseract baseline of character error rate (CER) are reported.
The results show that multilingual transfer is the dominant factor in extremely low-resource Bantu translation while eliminating the need for heuristic proxy selection, and all systems fail to preserve tonal diacritics, highlighting an open challenge.
Samiratu Ntohsi, Neza David Tuyishimire, Anesu Kafesu et al.· 0 citations
MameLoshnLM, the first open-source 8B-parameter language model built specifically for Yiddish, is presented, providing both a foundation for Yiddish NLP and a practical template for language model development in historically rich but digitally underrepresented languages.
Uri Katz, Omer Goldman, Tomasz Limisiewicz et al.· 0 citations
Control studies on model depth, learning-rate magnitude and scheduling, the interaction between learning-rate decay and data quality, and the effect of including instruction data during pre-training are reported, and an unexplained surge in benchmark performance midway through the constant-learning-rate phase is documented.
Lusheng Zhang, Shien He, Tianxing Yan et al.· 0 citations
This paper tests whether prompting the same (frozen) SLM in three typologically diverse languages and aggregating the outputs can improve classification without retraining or translation, and suggests that cross-lingual diversity rather than surface-level input variation drives the gain.
MTEB-PT is presented, a Portuguese benchmark constructed from a subset of MMTEB, comprising 14 existing datasets across Semantic Textual Similarity (STS), classification, retrieval, and reranking, and shows that language-specific fine-tuning still improves model performance in Portuguese, especially on task types that match the adaptation data most closely.
Lucas H.T. Okamura, Alexandre Alcoforado, A. H. R. Costa· 1 citation
Part-of-speech (POS) tagging for low-resource languages is limited more by the expense of generating labeled training data than by algorithmic constraints. This study evaluates the extent to which a length-stratified, minimally-supervised pipeline can recover tagging accuracy for the Assamese language using a labeled seed of only 3 to 5 sentences per distinct sentence length (142 to 228 sentences in total), compared to a weak-supervision classifier trained on a conventional 80% corpus split (2,840 sentences). An evidence-gated cascade is constructed, comprising Brown-style word clustering, a seed-ambiguity audit with context-based disambiguation, confidence-gated cluster labeling, and empirically validated rule fallbacks. This approach achieves 94.9% to 95.0% accuracy but only 19.8% to 20.0% token coverage from its minimal seed. To address the coverage gap, every remaining token is backfilled using a classifier trained on the pipeline’s accumulated evidence, resulting in 100% coverage at 84.7% to 86.0% accuracy (depending on seed size), as evaluated against a fixed, shared test set for direct comparison with the larger-budget baseline (90.8% accuracy, same coverage, same test set). Three negative results are reported alongside the positive findings: Brown-style clustering provides negligible benefit at this scale, Viterbi sequence decoding does not transfer to the backfill classifier from the larger-budget setting, and naive self-training fails due to confirmation bias. Additionally, through five replicated random seed draws per condition, it is demonstrated that the residual seed-size effect (3 versus 5 sentences per length, closing approximately 21% of the accuracy gap to the larger-budget baseline) is statistically significant rather than the result of a single favorable draw.
B. Sarma, Rupam Baruah, Diganta Baishya· International journal of com...· 0 citations