To support the development of automatic means for evaluating translation quality, the LDC was sponsored to solicit 11 sets of human translations for a single set of Mandarin Chinese source materials. The LDC was also asked to produce translations from various commercial Machine Translation (MT) systems as well as from MT systems available on the Internet.
The Multiple-Translation Chinese (MTC) Corpus, produced by Linguistic Data Consortium (LDC) catalog number LDC2002T01 and ISBN 1-58563-217-1 contains 17 human and MT translations of a target corpus.
Three sources of journalistic Mandarin Chinese text were selected from existing LDC corpora:
- Xinhua News Service: 52 news stories
- Zaobao News Service: 27 news stories
- Voice of America Mandarin broadcast transcripts: 26 news stories
for a total of 105 stories.
For an example of Mandarin Chinese Text please click on this example. For an example of English translation please click on this example.
The Xinhua data were drawn from the "Chinese Treebank Corpus" (LDC2001T11); the file names and "doc_id" attributes assigned to these stories match the file names used in the Chinese Treebank release.
The Zaobao and VOA data were both drawn from the "TDT3 Multilanguage Text Corpus" (LDC2001T58); their file names and "doc_id" attributes match the "DOCNO" tags assigned to these stories in the TDT3 release.
Selection of stories from the two newswire collections was controlled by story length: all selected stories contain between about 340 and 400 Chinese characters. The selection from VOA broadcasts varied more widely, between 100 and 1,000 characters per story.
The VOA Mandarin transcripts in TDT3 were created manually by a professional transcription service, but with limited editorial quality control -- while generally quite complete, these transcripts were not expected to exceed the quality or accuracy of closed-caption text in television broadcasts.
Zaobao is a news portal from Singapore and many of its news stories are translations from other news agencies' releases.
There are a total of 1,890 data files.
February 23, 2002: Possibly due to segmentation errors, two of the Machine Translation (MT) systems, tb2 and tb3, have output seven instances of obscenities, in files VOM19981114_0900_1588.sgm and chtb_177.sgm
Portions: 1994-1998 Xinhua, 1998 Zaobao