Evaluation Campaign
>> Data Downloads

Username/Password

You will need a username and password to download the data. Please register here.

Overview

This webpage offers the following data:

• Training data: This is the same for the translation of the manual transcription and the translation of ASR lattices and n-Best list.
• Development data: Different development sets for the translation of the manual transcription and the translation of ASR lattices.
• Test data: Will be available here August 16th for both tracks.

The training data is the same in both cases.

The development sets are different. The development sets for the source language are:

• manual transcriptions if only manual transcriptions are translated.
• n-best lists and lattices from a Speech Recognizer if ASR output is translated.

Training Data

Contents

Every data package contains 2 files with 20,000 lines each for source and target language. Languages other than English are segmented based on the ASR output.
Please note: The punctuation marks in the English files are not separated. You can certainly separate those punctuation marks in all tracks and conditions.
The Korean Data was segmented using the ATR KOMA-HanTag software. This software is not publicly available

Data Packages

• Training Package Chinese → English
  (Training-CE.tar.gz)
  
  
• Training Package English → Chinese
  (Training-EC.tar.gz)
  
  
• Training Package Japanese → English
  (Training-JE.tar.gz)
  
  
• Training Package Korean → English
  (Training-KE.tar.gz)
  
  
• Training Package Arabic → English
  (Training-AE.tar.gz)
  

Development Sets for Translation of Manual Transcription

Contents

Every data package contains 2 development sets. The test sets of CSTAR 2003 and IWSLT 2004 in source and target language as plain text. Languages other than English are segmented based on the ASR output.

Data Packages

• Development sets(text) Chinese → English
  (Development-Text-CE.tar.gz)
  
  
• Development sets(text) English → Chinese
  (Development-Text-EC.tar.gz)
  
  
• Development sets(text) Japanese → English
  (Development-Text-JE.tar.gz)
  
  
• Development sets(text) Korean → English
  (Development-Text-KE.tar.gz)
  (this development set only contains the CSTAR03 test set. The IWSLT04 test set is not available in Korean)
  
  
• Development sets(text) Arabic → English (Development-Text-AE.tar.gz)

Development Sets for Translation of Speech Recognizer output

Contents

Every data package contains 2 development sets. The test sets of CSTAR 2003 and IWSLT 2004 in source and target language. The source side is provided as Speech Recognizer output (lattice and n-best list). Languages other than English are segmented based on the ASR output.

• No Korean lattices or n-best lists will be provided.

Data Packages

• Development sets(lattices & n-best lists) Chinese → English
  (Development-lattice-nbest-CE.tar.gz)
  This set contains lattices and n-best lists for Chinese generated with Speech recognition systems of different accuracy.
  
  
• Development sets(lattices & n-best lists) English → Chinese
  (Development-lattice-nbest-EC.tar.gz)
  
  
• Development sets(lattices & n-best lists) Japanese → English
  (Development-lattice-nbest-JE.tar.gz)
  
  
• Development sets(lattices & n-best lists) Arabic → English
  (Development-lattice-nbest-AE.tar.gz)

Speech Data

Contents

Chinese Speech Data

• Development set 1 - CSTAR 2003 (CSTAR-2003_chinese.tar.gz)
  
  
• Development set 2 - IWSLT 2004 (IWSLT-2004_chinese.tar.gz)
  
  
• Test set - IWSLT 2005 (IWSLT-2005_chinese.tar.gz)
  

Japanese Speech Data

• Development set 1 - CSTAR 2003 (CSTAR-2003_japanese.tar.gz)
  
  
• Development set 2 - IWSLT 2004 (IWSLT-2004_japanese.tar.gz)
  
  
• Test set - IWSLT 2005 (IWSLT-2005_japanese.tar.gz)