Machine Translation The Translator’s Choice Heidi Düchting Sylke Krämer Johann Roturier
Dec 30, 2015
Commercial Imperatives
Effective– Time-critical documents in volume
Efficient– Translation process automation– Combining translation technologies
workflow TM, MT, and PE tools
Control– Loose writing guidelines vs. Controlled Language rules
Improved machine translatability
Commercial Systems
Combine technologies TM with previously machine-translated and post-edited
segments for look-up
TM systems with MT component Rule based and Example based Pre-translate phase Towards improved post-editing efficiency? Not available in all systems
MT systems with TM component 100 % match look-up
Challenges
Setting a threshold for TM matches– 100% matches only
suitable when the objective is to provide MT output for gisting (no post-editing)
suitable when the MT system is fully customized and CL environment is in place (no post-editing?)
Quick PE New sentences in which only one character changes are sent
to the MT engine– W32.Beagle.AB is a mass-mailing worm that neither propagates via
network shares nor deletes files– W32.Beagle.AC is a mass-mailing worm that neither propagates via
network shares nor deletes files
Solutions (1)
Two-tier process Leverage Trados TM repository Use MT system to translate unknown segments (Systran
Premium 5.0) Use MT output as TM input
Determine the export threshold Existing TM segments vs. new controlled segments
– Uncontrolled: Symantec announced a patch was available– CL: Symantec announced that a patch was available
Solutions (2)
TMX format obvious choice as the exchange format XLIFF not supported by all MT systems source and target segments
<tu usagecount="1" creationdate="20050301T122255Z" creationid="SUPER"><tuv lang="EN-US"><seg>Then the worm searches all local and network drives for .gif, .bmp, and .wav
files.</seg></tuv><tuv lang="DE-DE"><seg>Then the worm searches all local and network drives for .gif, .bmp, and .wav
files.</seg></tuv></tu>
Processing TMX
Technical issues TMX's various implementations can create discrepancies
during the exchange process Identical source and target segment XML parser and TMX header
Pre and post processing with a single macro Modules to remove and restitute sections Environment: VBA
Pre-translation Workflow
Step 6:Import
segments into TM
Step 6:Import
segments into TM
Step 5:Post-
processing module
Step 5:Post-
processing module
Step 4:Call to MT
system
Step 4:Call to MT
system
Step 3:Pre-
processing module
Step 3:Pre-
processing module
Step 2:Export
unmatched segments
Step 2:Export
unmatched segments
Step 1:Analyze
new document
Step 1:Analyze
new document
Effective pre-translation
Efficiency and robustness Refinable
Opportunity for modifications Target segments CL environment predictability Frequent errors
Ideal scenario Address problems that could not be fixed with CL rules
Towards Automated Post-Editing
Surface post-editing No linguistic analysis: no second MT Text processing Frequent errors due to default MT settings Remove drudgery from post-editing
Lexical Capitalization (folgende vs. Folgende) Incorrect spelling (neuzustarten vs. neu zu starten) Missing contractions (à le vs. au) Extra words (fichier de .bmp vs. fichier .bmp)
Towards Automated Post-Editing
Syntactic Word order: “Klicken auf Sie” vs. “Klicken Sie auf” Wrong structures (transfer or generation issue): neither…nor
(ni ne..ni ne)
Textual Formatting: trailing spaces after symbols (backslashes) Punctuation inconsistent with style guide: inverted commas
for German
Towards Automated Post-Editing
Suitability of the environmentRegular expressions supportRE are a ‘way to describe text through pattern matching’ (Stubblebine 2003: 1)Grouping and Capturing:
1.Match: ([Kk]licken) (auf) (Sie)
2.Replace: \1 \3 \2
Next steps
New environment– GMS integration
Centralized interface with content Transport layer MT as plug-in
– XLIFF format To machine translate unmatched segments
– PE replacements Fine-tune contextual replacements
Conclusions
Combining MT & TM is efficient leverage post-editing is not repeated increased throughput
Environment for avoiding errors facilitated when CL rules are introduced Scope of errors is reduced
New opportunities for translators Fine-tuning MT user dictionaries Refine automated PE tasks
Thank [email protected]