Jun 2026· arXiv.org· Vol abs/2606.31718· 0 citations· 16 references
Computer Science
TL;DR
The results show that Romanian incurs a 3 to 5 percentage point drop relative to English in prompt-only settings, that few-shot prompting provides marginal gains over zero-shot, and that QLoRA fine-tuning improves macro F1-Score by more than 22 percentage points in both languages while reducing the cross-lingual gap from 3.3 to 1.4pp.
Abstract
Relation extraction (RE) for low-resource languages is typically constrained by the lack of annotated corpora. We investigate the feasibility of cross-lingual RE for Romanian by combining automatic dataset translation with large language model (LLM) inference. We translate the SemEval-2010 Task 8 benchmark from English to Romanian using an LLM-based translation pipeline and evaluate Gemma 4 31B under zero-shot, few-shot, and QLoRA fine-tuned configurations, against four encoder baselines spanning 125M to 560M parameters: XLM- RoBERTa (base and large), Romanian BERT, and RoBERT- large. We assess two task formulations: relation classification with marked entities and end-to-end extraction. Our results show that Romanian incurs a 3 to 5 percentage point (pp) drop relative to English in prompt-only settings, that few-shot prompting provides marginal gains over zero-shot, and that QLoRA fine-tuning improves macro F1-Score by more than 22 percentage points in both languages while reducing the cross-lingual gap from 3.3 to 1.4pp. The encoder baselines come within 1-4pp of QLoRA Gemma on Romanian despite being 50-250 times smaller, with monolingual Romanian BERT at 125M parameters matching multilingual XLM-R at 278M. The case for using a 31B model for single-task RE on Romanian is therefore weak in deployment scenarios where compute matters. We release the translated dataset, evaluation code, and trained models.
Large language models (LLMs) are increasingly used as general-purpose translation systems, but their behavior is usually evaluated under a single prompt shape: translate one source sentence into one target language. In practice, users may ask for one target language, for several related languages at once, or for translations conditioned on examples. This paper studies prompt scope and demonstration selection as experimental variables for local LLM machine translation. We evaluate English-to-Romance and English-to-Germanic translation on the full FLORES devtest split for nine official European Union languages. We compare three local instruction-tuned LLMs, llama3.2:3b, mistral:latest, and qwen2.5:14b, against dedicated MT baselines from OPUS-MT and NLLB-200. We test zero-shot prompting and k=5 few-shot prompting with random, lexical-similarity, and embedding-similarity demonstration selection. We also compare single-target prompts with JSON-formatted family-scope prompts that request all languages in a family at once. Results show that dedicated MT systems remain strongest overall, especially for Germanic languages. Few-shot prompting helps mistral:latest and qwen2.5:14b, but hurts llama3.2:3b; embedding retrieval is best on average for the stronger LLMs, but its advantage over random and lexical examples is modest. Family-scope prompting is feasible for stronger local LLMs but exposes structured-output failures in smaller models. These findings motivate evaluating LLM translation not only by language pair and metric, but also by prompt scope, retrieval strategy, and multi-target compliance.
Discourse data are the primary empirical basis of grammar writing in field linguistics, but producing interlinearized text is notoriously expensive - on the order of one hour of work per minute of recording. For endangered languages, where the time remaining to verify analyses with native speakers is itself limited, automating parts of the interlinearization workflow has direct documentary value. We implement a full neural annotation pipeline (morpheme segmentation, POS tagging, glossing) for Irabu Ryukyuan using deliberately small, transparent BiLSTM-CRF models, and evaluate it under a realistic hard constraint: approximately one hour of fully annotated discourse as the entire supervised resource. Two factors of the annotation itself are manipulated: its richness (with or without a POS tier) and its quantity (training budgets from 6 to 47 minutes). Gold POS improves grammatical glossing by +4.4 (SD 0.7) points (significant in all 5 seeds), and the gain grows as data shrink (+11.6 points at a quarter of the data); a POS tier more than halves the amount of glossed data needed to reach a given accuracy. In a fully automatic pipeline this gain is not yet realized: the tagger still errs on 12% of morphemes, and an incorrect POS misleads the glossing model more than no POS at all. The value is latent rather than lost: degrading gold POS with controlled noise shows the gain returning as tagger accuracy rises, with break-even near our tagger's current 88% and +1.6 to +3.2 points recovered at 92-96%. We conclude with a concrete recommendation for documentation practice: annotate quadrilinearly - text, POS, gloss, translation.
General-purpose large language models (LLMs) may struggle in specialised machine translation, but the conditions under which fine-tuning improves translation performance remain unclear. This study compares full-parameter fine-tuning (FPFT) and parameter-efficient fine-tuning (PEFT) for Chinese-English political discourse translation using a purpose-built corpus and the Qwen3-14B model. Translation performance was assessed on three 50-item test sets using BLEU, ROUGE-L F1, METEOR, and BERTScore F1, together with BLEU pass-rate likelihood-ratio G2 tests, paired t-tests, and paired Cohen’s dz for item-level score differences. The results reveal a clear contrast between unseen in-domain evaluation, maximum-overlap benchmarking, and semantically related but non-fine-tuned evaluation. On Test Set A and Test Set C, neither fine-tuning strategy produced a statistically significant BLEU pass-rate advantage over the base model, and paired tests across the continuous metrics did not show consistent fine-tuning gains. On Test Set B, which was sampled from the fine-tuning corpus, both fine-tuned models substantially outperformed the base model across all four metrics, with FPFT achieving the highest scores and PEFT providing a more computationally efficient alternative. These findings indicate that textual overlap between training and deployment data, rather than broad domain similarity alone, strongly conditions the observed benefit of fine-tuning. The study offers an empirically grounded framework for selecting fine-tuning strategies in specialised machine translation.
This study investigates the English-to-Turkish clinical translation performance of three large language models (Aya-23, Seed-X-Instruct, Gemma-3) using zero/few-shot prompting and QLoRA fine-tuning. Our three-stage setup evaluates: (1) general translation after FLORES-devtest training, (2) medical domain adaptation on i2b2 via a mixed dataset, and (3) generalization capacity on the independent FLORES-test set after mixed-data fine-tuning, using SacreBLEU, chrF, and COMET. Results show that Seed-X-Instruct achieves the strongest fine-tuned performance, reaching 32.70 BLEU, 64.34 chrF, and 90.65 COMET on the general FLORES test set, and 22.40 BLEU, 54.61 chrF, and 79.70 COMET on the medical i2b2 set. After mixed-data training, Seed-X improves from 11.48 to 32.39 BLEU on the unseen FLORES test set, whereas Aya-23 drops from 19.23 to 10.83 and Gemma-3 largely preserves its strong zero-shot baseline. These findings indicate that QLoRA effects are model-dependent and that domain adaptation does not yield the same generalization behavior across models.
Muhammetalp Erdem, Ş. Bora, Oguz Dikenelli· Signal Processing and Commun...· 0 citations
Automatic question generation is one solution to help create test items that require a lot of time and effort. The state-of-the-art model for automatic question generation in Bahasa Indonesia, which uses idT5, has several drawbacks, including misuse of context, overuse of question words, and an answer target that must be exactly from the context, which makes it extractive. This research aims to address those problems by improving the performance of previous models through a new fine-tuning scheme. This research proposed a prefix-prepend format for fine-tuning with the idT5-base model on the Indonesian Stanford Question Answering Dataset (SQuAD) and the Typologically Diverse Question Answering (TyDiQA) dataset. We evaluate the model with Bilingual Evaluation Understudy (BLEU), Recall-Oriented Understudy for Gisting Evaluation (ROUGE), and Bidirectional Encoder Representations from Transformers (BERT) similarity metrics. The results show that prefix-prepend fine-tuning improved the performance of the baseline model. Our best model achieved 0.1643 BLEU, 0.4099 ROUGE-L, and 0.7177 BERT similarity score on SQuAD, and 0.1941 BLEU, 0.4301 ROUGE-L, and 0.7291 BERT similarity score on TyDiQA. The human evaluation using the Content Validation Index (CVI) and a paired t-test indicated that the proposed model performed better than the baseline. While the proposed model addresses many of the baseline’s shortcomings, it still struggles to handle questions that require understanding complex relationships between entities. Future studies can explore improvements for this case using external knowledge or other models.
The results indicate that a moderately sized, shared self-attention architecture can deliver production-quality multilin-gual translation within the resource constraints of an academic de-ployment, while surfacing clear directions – low-resource language coverage, domain adaptation, and speech-based extension – for con-tinued development.
Darshan Gowda D H and Dr. Kruti R· International Journal of Adv...· 0 citations