Active Curation of Bi-Text Resources in Commercial Localization Workflows
1 like391 views
The document discusses the active curation of bi-text resources within commercial localization workflows, particularly leveraging W3C semantic web standards to enhance translation processes through linked data. It emphasizes the importance of human validation in improving automatic term extraction and machine translation efficiency. Furthermore, it highlights tools like Babelfy and BabelNet to provide contextual translation suggestions and facilitate a decentralized localization web of multilingual resources.
Active Curation of Bi-Text Resources in Commercial Localization Workflows
- 1. Active Curation of Bi-Text Resources in Commercial Localization Workflows Dave Lewis TCD, Andrzej Zydroń XTM International
- 2. • Open Data on the Web: W3C Semantic Web standards allow data to be published on Web – Fine-grained URI-based inter-linking – Extensible meta-data – Standard Query APIs • Enables a Localization Web – Terms and translations become linkable resources – Meta-data from L10n workflows adds value – Leverage in training Machine Translation and Automatic Term Extraction The Localization Web The Localization Web = Decentralised Annotated Global Translation Memory and Term Base
- 3. Web of Multilingual Content
- 5. • Rich word and phrase resources to assist translators Babelfy: Public Lexical Resources
- 6. • Translation suggestions can be fed into MT for more reliable translation Links to BabelNet offer suggestions for Definitions and Translations
- 7. Babelfy & Babelnet offer more term suggestions
- 8. • Public resources may not always yield the right definitions or translations for the context • Need to track human validation/ rejection to train automatic term extraction
- 9. Active Curation of Linked Language Resources The company has also reduced its production capacity by ceasing manufacture of chest freezers and freestanding microwave ovens Extraction & Segmentation production capacity capacité de production ✔ ✔ Annotation with Existing Terms chest freezer microwave oven réfrigérateur four à micro-onde ? ? ? ? Auto suggestion from Babelfy/Babelnet D'autre part, la société a réduit sa capacité de production en arrêtant la production de réfrigérateur et de fours micro-onde pose-libre Machine Translate with Term Translations MT Vendor? D'autre part, la société a réduit sa capacité de production en arrêtant la production de congélateurs coffres et do fours micro-ondes pose-libre ✗ congélateurs coffres fours micro-ondes ✔ Postedit and capture terms in context ✔✔ ✔ ✔ ✔ ✔ PE PE PE PE PE PE PE ✗ PE ✔
- 10. • CSV of the Web: tables and JSON meta- data • JSON-Linked Data • Provenance Vocabulary • Data Catalogue • Open Annotation • ITS2.0 Vocabulary • Also: – Provenance Plan – Open Data Rights Language Linked Data Based on W3C Standards
- 12. Parallel Text & Term base Posteditors Machine Translation Active Curation: Dynamic MT Retraining • Tighten curation cycle: from projects to segments – Prioritise postedits for retraining • Prioritise Term Identification by posteditors • Assemble MT-ready, lexically-rich term bases
- 13. • TermWeb/XTM/DCU • Introducing Next Gen Machine Translation • Massive scale bilingual dictionaries • BabelNet • Automatic Term Extraction: forced decoding • Dynamic retraining • Optimal segment translation route • L3Data curation, sharing Next Generation Machine Translation
- 14. Data Management Lifecycles Publish Correct & refine Lex- concept lifecycleCorrect & refine Discover & use Discover & use Correct & refine Bitext lifecycle Discover data (Re)train- MT Revise and annotate Publish Content lifecycle Publish I18n & source QA Trans QA Post- edit Automated translation Consume Create
- 15. • Better in-context postediting: – XTM-Easyling • Feeding term suggestions from posteditor to Terminology Management – XTM-Interverbum • Dynamic Retraining – XTM-DCU • Bilingual Dictionary SMT improvements – XTM-DCU • NER, terminology enforcements, forced decoding – XTM-Interverbum-DCU • Postediting prioritisation and term flagging – TCD-DCU-XTM • Publishing interlinks of parallel text, lexically rich term bases – TCD: DG-T TM, EurVoc, Snomed-CT, LEMON, BabelNet • Closing the loop – operational instrumentation of postediting – XTM Systems Integration