Enhanced Infrastructure for Creation & Collection of Translation Resources

Zhiyi Song, Stephanie Strassel (speaker), Gary Krug, Kazuaki Maeda

Introduction

LDC develops large scale parallel text corpora for sponsored research programsManual creation of parallel text by human

translatorsHarvesting, aligning potential parallel

documents from known repositories and the web

Recent expansion in scope and variety Requiring improvements in quality, efficiency

and cost-effectiveness

Context for Resource Creation

Previous focus primarily Chinese, Arabic newswire (NW) Current focus on "unstructured" data

Broadcast News (BN) and Broadcast Conversation (BC) Weblogs, Newsgroups (WB) Handwritten document images of many types (VAR)

New linguistic varieties Eight language pairs in the LCTL program Colloquial Arabic varieties for some projects

New evaluation requirements Multiple human translations, adjudication of multiple translations Translation alternatives for ambiguous source text Translation post-editing

Recent translation efforts

Language PairApproximate Volume (Words) Genres Methodology

Arabic > English 100M + BN, BC, NW, WB, VARmanual translation, parallel text harvesting, acquisition of existing manual translations

Chinese > English 100M + BN, BC, NW, WBmanual translation, parallel text harvesting, acquisition of existing manual translations

English > Chinese 250K + BN, BC, NW, WB manual translation

English>Arabic 250K + BN, BC, NW, WB manual translation

Bengali>EnglishPashto>EnglishPunjabi>EnglishTagalog>EnglishTamil>EnglishThai>EnglishUrdu>EnglishUzbek>English

250-500K + per language pair

manual translation, parallel text harvestingNW, WB

Manual Translation Pipeline

datapool

selectaudio

selecttext

selectedweb data

segmentinto

sentence units convert

toreleaseformat

sourcetext

translatedtext

validate release package

convert to translator-friendly format

translation

QC

transcription and

segmentation

Manual Translation

Commercial agencies vetted, trained by LDC Required to use LDC's project-specific guidelines

Accuracy and fidelity over fluencyGeneral principles, language-specific requirementsRules for named entities, disfluencies, emoticons, etc.Requirements for formatting and validationMultiple examples of preferred translation

Separate guidelines for specialized tasksPost-editing machine translation outputTranslation alternativesTranslation of novel single sentencesTranslation of handwritten document images

Translation QC

All translations undergo additional QC at LDCTypically 10% of training data, 100% of evaluation data

reviewed Standardized QC rating system deducts points

for each type of errorQC report including score, examples sent to translatorsFailing score requires re-translation of full data set

QC process facilitated by customized TransQC GUI

QCTrans GUI

Translation Project Management

Translation database is core management tool Document ID, language, genre, token count, LDC file server path Data set information including project, phase, partition, restrictions Translator assignment, due date, status, QC score, payment info

Backend to LDC Translator Extranet Translators access and submit assignments, validate submissions,

view QC reports, generate invoices, check payment status

Queries support status tracking but also assignment generation, data selection, cross-project coordination What translation assignments are pending delivery this week? What is average QC score for this translator on Chinese BC? List Arabic NW files from 2007 that have never been released as

GALE training data and are not part of any project's eval set

LDC Translation Database

Parallel text harvesting

Manual translation supplemented by harvesting and alignment of potential parallel textHarvest text from multilingual sites

E.g. newswire providers

Standardize markup formatUse BITS document mapping module to find likely

parallel documentsUse Champollion to find sentence alignments

High yields in GALE program82,000 Arabic-English document pairs67,000 Chinese-English document pairs

Conclusion

Robust, flexible translation infrastructure to support multiple, distinct, concurrent projects

Much of this infrastructure freely available from LDCTask specifications, guidelines available for all projects

http://projects.ldc.upenn.edu/gale/Translation/

QCTrans GUI slated for free, open-source distribution

Many resulting parallel text corpora already in LDC Catalog

Newly emerging data sets to be added over time

Recent corpora

Catalog Number Title

LDC2007T23 GALE Phase 1 Chinese Broadcast News Parallel Text - Part 1

LDC2008T08 GALE Phase 1 Chinese Broadcast News Parallel Text - Part 2

LDC2008T18 GALE Phase 1 Chinese Broadcast News Parallel Text - Part 3

LDC2007T24 GALE Phase 1 Arabic Broadcast News Parallel Text - Part 1

LDC2008T09 GALE Phase 1 Arabic Broadcast News Parallel Text - Part 2

LDC2009T02 GALE Phase 1 Chinese Broadcast Conversation Parallel Text - Part 1

LDC2009T06 GALE Phase 1 Chinese Broadcast Conversation Parallel Text - Part 2

LDC2008T02 GALE Phase 1 Arabic Blog Parallel Text

LDC2008T06 GALE Phase 1 Chinese Blog Parallel Text

LDC2009T03 GALE Phase 1 Arabic Newsgroup Parallel Text - Part 1

LDC2009T09 GALE Phase 1 Arabic Newsgroup Parallel Text - Part 2

LDC2009T15 GALE Phase 1 Chinese Newsgroup Parallel Text - Part 1

LDC2010T03 GALE Phase 1 Chinese Newsgroup Parallel Text - Part 2

Acknowledgements

This work was supported in part by the Defense Advanced Research Projects Agency, GALE Program Grant No. HR0011-06-1-0003. The content of this paper does not necessarily reflect the position or the policy of the Government, and no official endorsement should be inferred.