首页 > AI前沿 > IsharaBART: An efficient BART-Based Framework for Low-Resource Bangla Sign Language Translation

IsharaBART: An efficient BART-Based Framework for Low-Resource Bangla Sign Language Translation

arXiv自然语言 2026-11-12 14:21 15 阅读 查看原文
用户现在需要把给定的新闻正文排版成语义化的HTML片段,规则是用p、h2/h3、列表、blockquote、code这些,保留内容不增删。首先看内容,先理结构:首先可以分小节对吧?比如第一部分是背景与研究提出,然后是数据集构建,然后是模型训练,然后是对比实验,然后是消融研究?不对,先看原文的段落。首先原文第一句是背景,然后介绍IsharaBART,然后数据集,然后模型训练,然后对比,然后消融。 首先,先设h2的标题?比如

研究背景与框架提出

,然后下面放

里面的内容:Millions of Bangla Sign Language (BdSL) users lack reliable automatic translation tools. Existing systems are limited by data scarcity, unidirectional design, and computationally expensive architectures that are not suitable for low-resource deployment. We introduce IsharaBART, a bidirectional generative framework for gloss-sentence translation in BdSL. IsharaBART uses Bidirectional and Auto-Regressive Transformers (BART)-based conditional language modeling with denoising learning objectives. 这个p可以,重要的地方比如IsharaBART可以加strong?对,规则说重要句子可用strong。 然后下一个h2:

数据集构建

,然后p里面的内容:We developed a parallel dataset of 1,000 professionally annotated gloss-sentence pairs, balanced across tenses and structures. We expanded the dataset to 4,831 pairs through four augmentation techniques: 然后后面四个技术用列表对吧?对,四个技术是规则的,所以后面接
  • rule-based tense transformations
  • masked token substitutions
  • back-translation
  • one-shot prompting
,哦不对,前面那句We expanded...through four augmentation techniques: 然后直接放列表,然后再有没有后续?哦,这句说完列表就结束?不对,原句是We expanded the dataset to 4,831 pairs through four augmentation techniques: rule-based tense transformations, masked token substitutions, back-translation, and one-shot prompting. 所以把冒号后面的四个做成列表,没问题。 然后下一个h2:

模型训练

,然后p:A compact denoising sequence-to-sequence model was pre-trained from scratch on monolingual Bangla text from six different corpora, then fine-tuned on the annotated and augmented datasets. 这个直接p包裹就行。 然后下一个h2:

实验对比结果

,然后p里面:We compare IsharaBART against four larger baseline models under identical conditions, where it achieves comparable or superior bidirectional performance. On the gloss-to-sentence (G2S) task, it attains Bilingual Evaluation Understudy (BLEU-4) score up to 51.7, and on the sentence-to-gloss (S2G) task, it achieves 61.40, outperforming multilingual models despite having fewer than 10% of the parameters. 这里IsharaBART可以加strong,还有关键的分数可以加strong吗?对,比如51.7、61.40这些重要的可以加strong。 然后下一个h2:

消融研究结论

,然后p:Ablation studies demonstrate that pre-training provides 62.5-90.4% faster convergence and 11.5-118% higher BLEU-4 than scratch. Low-rank adaptation (LoRA) reduces trainable parameters by 99.3%, but with 28.0-52.2% loss of BLEU-4 compared to full-parameter fine-tuning. 这里LoRA可以加strong?或者重要的数据加strong。 等下,有没有需要代码的?原文里没有代码,所以不用pre。引用也没有,所以不用blockquote。现在检查有没有增删内容:所有原文的内容都要保留对吧?对,每个部分都要包含。 现在调整一下结构,确保每个h2的标题合适,段落正确。等下,再核对原文的所有句子: 1. Millions of Bangla Sign Language (BdSL) users lack reliable automatic translation tools. Existing systems are limited by data scarcity, unidirectional design, and computationally expensive architectures that are not suitable for low-resource deployment. We introduce IsharaBART, a bidirectional generative framework for gloss-sentence translation in BdSL. IsharaBART uses Bidirectional and Auto-Regressive Transformers (BART)-based conditional language modeling with denoising learning objectives. 这部分放在第一个h2下,p包裹,IsharaBART用strong。 然后数据集部分:We developed a parallel dataset of 1,000 professionally annotated gloss-sentence pairs, balanced across tenses and structures. We expanded the dataset to 4,831 pairs through four augmentation techniques: 后面四个列表项,没错。 然后模型训练:A compact denoising sequence-to-sequence model was pre-trained from scratch on monolingual Bangla text from six different corpora, then fine-tuned on the annotated and augmented datasets. 直接p。 然后实验对比:We compare IsharaBART against four larger baseline models under identical conditions, where it achieves comparable or superior bidirectional performance. On the gloss-to-sentence (G2S) task, it attains Bilingual Evaluation Understudy (BLEU-4) score up to 51.7, and on the sentence-to-gloss (S2G) task, it achieves 61.40, outperforming multilingual models despite having fewer than 10% of the parameters. 这里IsharaBART、BLEU-4、51.7、61.40可以加strong。 然后消融:Ablation studies demonstrate that pre-training provides 62.5-90.4% faster convergence and 11.5-118% higher BLEU-4 than scratch. Low-rank adaptation (LoRA) reduces trainable parameters by 99.3%, but with 28.0-52.2% loss of BLEU-4 compared to full-parameter fine-tuning. 这里LoRA、BLEU-4、99.3%、28.0-52.2%可以加strong。 现在把这些组织成HTML片段,不要文档级标签。现在写出来:

研究背景与框架提出

Millions of Bangla Sign Language (BdSL) users lack reliable automatic translation tools. Existing systems are limited by data scarcity, unidirectional design, and computationally expensive architectures that are not suitable for low-resource deployment. We introduce IsharaBART, a bidirectional generative framework for gloss-sentence translation in BdSL. IsharaBART uses Bidirectional and Auto-Regressive Transformers (BART)-based conditional language modeling with denoising learning objectives.

数据集构建与扩充

We developed a parallel dataset of 1,000 professionally annotated gloss-sentence pairs, balanced across tenses and structures. We expanded the dataset to 4,831 pairs through four augmentation techniques:

  • rule-based tense transformations
  • masked token substitutions
  • back-translation
  • one-shot prompting

模型训练流程

A compact denoising sequence-to-sequence model was pre-trained from scratch on monolingual Bangla text from six different corpora, then fine-tuned on the annotated and augmented datasets.

基线对比实验结果

We compare IsharaBART against four larger baseline models under identical conditions, where it achieves comparable or superior bidirectional performance. On the gloss-to-sentence (G2S) task, it attains Bilingual Evaluation Understudy (BLEU-4) score up to 51.7, and on the sentence-to-gloss (S2G) task, it achieves 61.40, outperforming multilingual models despite having fewer than 10% of the parameters.

消融研究结论

Ablation studies demonstrate that pre-training provides 62.5-90.4% faster convergence and 11.5-118% higher BLEU-4