Datasets:
license: apache-2.0
task_categories:
- translation
language:
- ja
- en
This dataset contains document-length Japanese-English parallel texts from various sources. The intended use case is for translation tasks. The cumulative dataset is under an apache 2.0 license, but the sources differ on that account. If there are conflicts, assume the more restrictive clause takes priority (?)
I am unfamiliar with licenses in general, so if someone sees any issues with the licensing situation, please let me know.
Metadata meaning:
source: self-explanatory
specific to manual fanfic:
missed_lines: number of lines in the document where the index matched between the unaligned source and translation, but the content of the lines did not meet the match threshold. This number being high relative to the number of lines in the document indicates either a poor translation, or some other factor causing the source and translation to differ beyond the norm.
inserted_lines_src: number of lines that only appear in the source document, and matches are found surrounding them, indicating the line is some kind of insertion (like a note, afterword, title, etc). Is generally safe to ignore unless it's significant compared to the number of correctly aligned sentences, which would indicate a poor matchup between source and translation sources, for one reason or another.
inserted_lines_src: same as above, but for the target document instead
Dataset Sources
news_commentary (unknown license?)
iwslt2017 (cc-by-nc-nd-4.0)
https://www2.nict.go.jp/astrec-att/member/mutiyama/ (gpl v1.2)
manually scraped fanfiction and translations (apache 2.0)