rakhine-myanmar-parallel-corpus
收藏资源简介:
Rakhine-Myanmar平行语料库是一个专门为若开语与缅甸语(标准缅语)之间的机器翻译和自然语言处理研究设计的开源数据集。该数据集旨在解决若开语数字语言资源有限的问题,支持若开语的数字保存、研究和人工智能发展。数据内容由若开语和缅甸语之间的平行句对构成,以CSV格式组织,包含train.csv、dev.csv和test.csv等标准分割文件。每个CSV文件包含rakhine和myanmar两列,分别对应对齐的若开语句子和缅甸语句子。原始数据以纯文本文件形式提供,并通过脚本进行句子对齐和清理。数据集适用于多种任务,包括神经机器翻译、语言建模、聊天机器人训练、方言分析(若开语与标准缅语比较)、语言保护以及NLP基准开发。项目未来计划增加英语翻译层、语音数据集,并利用语义相似性模型改进对齐,目标将数据集扩展到10,000+句子对。当前数据集使用基于规则的句子对齐(逐行对齐),翻译质量依赖于人工整理,在规模和覆盖范围上有限。数据集采用知识共享署名4.0国际许可协议,鼓励通过添加新句对、改进翻译、修复对齐问题等方式进行贡献。
The Rakhine-Myanmar Parallel Corpus is an open-source dataset specifically designed for machine translation and natural language processing (NLP) research between Rakhine and Standard Burmese. This dataset aims to address the scarcity of digital language resources for the Rakhine language, supporting its digital preservation, academic research, and artificial intelligence (AI) development. The corpus consists of parallel sentence pairs between Rakhine and Standard Burmese, organized in CSV format, including standard split files such as train.csv, dev.csv, and test.csv. Each CSV file contains two columns: `rakhine` and `myanmar`, which correspond to the aligned Rakhine and Standard Burmese sentences respectively. The raw data is provided in plain text files, with sentence alignment and cleaning conducted via custom scripts. This dataset supports a wide range of tasks, including neural machine translation (NMT), language modeling, chatbot training, dialect analysis (comparison between Rakhine and Standard Burmese), language preservation, and NLP benchmark development. In the future, the project plans to add English translation layers and speech datasets, and improve alignment quality using semantic similarity models, with the ultimate goal of expanding the corpus to over 10,000 sentence pairs. Currently, the dataset uses rule-based sentence alignment (line-by-line alignment), and its translation quality depends on manual curation, resulting in limitations in terms of scale and coverage. The dataset is licensed under the Creative Commons Attribution 4.0 International License, and contributions such as adding new sentence pairs, improving existing translations, and fixing alignment issues are highly encouraged.
数据集概述
数据集名称:Rakhine-Myanmar Parallel Corpus
语言:缅甸语(my)
许可证:Creative Commons Attribution 4.0 International (CC BY 4.0)
任务类别:翻译
用途:机器翻译、语言建模、NLP研究、方言分析(若开语与标准缅甸语)、语言保护
数据集目的
- 提供若开语与缅甸语之间的平行语料,支持机器翻译模型训练。
- 促进若开语的数字资源建设与AI语言技术发展。
数据集结构
rakhine-myanmar-parallel-corpus/ ├── data/ │ ├── train.csv │ ├── dev.csv │ └── test.csv ├── raw_data/ │ ├── rakhine_text.txt │ └── myanmar_text.txt ├── scripts/ │ ├── align_sentences.py │ └── clean_data.py ├── README.md ├── LICENSE └── .gitignore
数据格式
- CSV格式:包含两列,
rakhine(若开语句子)与myanmar(缅甸语句子),每行为一对平行句子。
数据集构建方法
- 将原始文本文件放入
raw_data/目录(rakhine_text.txt和myanmar_text.txt)。 - 运行对齐脚本
scripts/align_sentences.py。 - 生成
data/train.csv。
使用场景
- 神经机器翻译
- 聊天机器人训练
- 若开语-缅甸语语言研究
- 方言对比
- 语言保护
- NLP基准测试开发
未来改进计划
- 增加英语翻译层(若开语↔缅甸语↔英语)
- 添加语音数据集(音频+转录)
- 使用语义相似度模型改进对齐
- 扩展数据集至10,000+句子对
- 在Hugging Face Datasets上发布
- 创建基准评估数据集
局限性
- 初始数据集使用基于规则的行对齐方法。
- 翻译质量依赖人工校对。
- 大规模数据集建议采用语义对齐方法。
- 数据集目前规模和覆盖范围有限。




