cross-lingual-span-profile
收藏资源简介:
cross-lingual-span-profile 是一个基于 MACULA 词位(lexeme)的结构化配置文件数据集,用于统计跨语言对齐中每个词位的 span 长度和多词倾向。该数据集通过聚合多种语言对齐结果,生成一种语言无关的中间语(INTERLINGUA)信号,能够指示希伯来语或希腊语词位在目标语言翻译中通常对应单个词还是多词短语(例如复合地名如“Kadesh Barnea”、复合数字如“four thousand”)。其预测依据是其他语言的实际对齐情况,无需目标语言模型。数据集包含两个核心字段:n_langs 表示支持该词位的独立语言数量(同一语言的不同版本会先合并,以避免权重偏差),multiword_rate 和 span_mean 分别表示跨语言平均的多词率和 span 均值。数据置信度随 n_langs 增加而提高,并会随着更多语言加入对齐而更新。该数据集适用于跨语言对齐、圣经文本多语言处理、中间语表示构建等任务,采用 CC0-1.0 许可,仅包含派生对齐统计信息,不涉及源文本分发。
The cross-lingual-span-profile is a structured profile dataset based on MACULA lexemes, designed to statistically analyze the span length and multi-word propensity of each lexeme during cross-lingual alignment. By aggregating multiple language alignment results, this dataset generates a language-agnostic interlingua (INTERLINGUA) signal, which can indicate whether a Hebrew or Greek lexeme typically corresponds to a single word or multi-word phrase in target language translation (e.g., compound place names such as "Kadesh Barnea", compound numerals such as "four thousand"). Its prediction is based on actual alignment data from other languages, without requiring a target language model. The dataset includes two core fields: 'n_langs' refers to the number of independent languages that support this lexeme (different versions of the same language will be merged first to avoid weight bias), while 'multiword_rate' and 'span_mean' respectively represent the cross-lingual averaged multi-word rate and span mean. The data confidence increases as the value of 'n_langs' grows, and will be updated as more languages are added to the alignment. This dataset is applicable to tasks such as cross-lingual alignment, multilingual processing of biblical texts, and construction of interlingua representations. It is licensed under CC0-1.0, and only contains derived alignment statistical information without involving the distribution of source texts.




