Unverified50% confidenceBenchmarkExact time
On Odia and Swahili, the back-translation method, using only 1000 parallel sentence pairs, outperforms continued pre-training by 7.7 to 16.4 XCOMET-XXL points in in-domain evaluation.
1
Sources
50%
Confidence
Long-term
Relevance
10/14/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedIn experiments with 85 million non-embedding parameters and matched training token budgets, unconstrained fitting produced higher held-out average bits-per-byte across all 12 languages under a shared unconstrained decoder.71% similarUnverifiedOn all low-resource languages, back-translation, compared with fine-tuning on parallel data alone, improves the XCOMET-XXL score by 16% to 173% in the direction of translation into the low-resource language.70% similarUnverifiedThe advantage of continued pre-training is that it can exploit massive unlabeled monolingual data, but the cost is that it requires a large amount of compute, and what is learned is only the distribution at the level of language form, not directly targeted at translation, the cross-lingual mapping task.68% similarUnverified手语翻译的端到端方法受限于平行语料稀缺性，目前性能仍落后于级联方案67% similarUnverifiedTraditional n-gram models and neural language models trained on a single language assign extremely low probability scores to cross-lingual word sequences, causing the decoder to tend to forcibly map foreign-language words to phonetically similar words in the target language.66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1028053API
curl https://kongchang.com/api/v1/knowledge/claims/1028053MCP
get_claim(id=1028053)