Chinese Transliteration
July 21, 2021 · View on GitHub
Background
Transliteration translates proper names and technical terms across languages that use different alphabets and sound systems.
Example input/output
Input:
约翰伍兹 (yue han wu zi)
Output:
John Woods
Standard Metrics
- Word Accuracy in Top-1 (ACC) measures correctness of the first transliteration candidate in a candidate list produced by a transliteration system.
- Fuzziness in Top-1 (Mean F-score).
- Mean Reciprocal Rank (MRR).
- MAP measures precision in the n-best candidates for i-th source name, for which reference transliterations are available.
NEWS 2018 Dataset_03.
Named Entity Workshop (NEWS) is a long-running transliteration evaluation campaign. Chinese/English is one of the most popular NEWS language pairs. For NEWS 2018:
- The shared-task report is in http://aclweb.org/anthology/W18-2409.
- The data is described http://workshop.colips.org/news2018/dataset.html.
| Test set name | Source | Target | Test set size (phrase pairs) |
|---|---|---|---|
| NEWS 2018 Dataset_03 T-EnCh | English | Chinese | 1000 |
| NEWS 2018 Dataset_03 B-ChEn | Chinese | English | 1000 |
Results
English-Chinese
| ACC | F-score | MRR | MAP | |
|---|---|---|---|---|
| He, Cohen (2020) | 0.299 | 0.6799 | ||
| EDI (University of Edinburgh) | 0.304 | 0.6791 | 0.4364 | 0.304 |
Chinese-English
| ACC | F-score | MRR | MAP | |
|---|---|---|---|---|
| UALB (University of Alberta) | 0.3 | 0.8 | 0.374 | 0.3 |
| EDI (University of Edinburgh) | 0.276 | 0.83 | 0.386 | 0.276 |
Resources
- NEWS 2018 also releases training and development data for English-Chinese and Chinese-English.
| Train set name | Source | Target | Train set size (phrase pairs) |
|---|---|---|---|
| NEWS 2018 Dataset_03 T-EnCh | English | Chinese | 41318 |
| NEWS 2018 Dataset_03 B-ChEn | Chinese | English | 32002 |
- TRANSLIT: A Large-scale Name Transliteration Resource
- Dataset for transliteration of person names and geolocations
- Contains about 1.6 million entries in more than 180 languages (including Chinese), and approx. 3 million name variations.
- Project link
- Paper
Suggestions? Changes? Please send email to chinesenlp.xyz@gmail.com