an investigation into the efficiency of translation dictation, by dr. masaru yamada, kansai...
TRANSCRIPT
An Investigation into the Efficiency of Translation
Dictation
Masaru Yamada, Kansai University, OsakaMichael Carl, Copenhagen Business School and
National Institute of Informatics, Tokyo
Goals of the study
● ENJA15: compare translation dictation (Aug. 2015 – Jan. 2016):
– with other translation modes (from-scratch, post-editing)
– between different experienced users (students, professionals)
● JTD16: investigate translation dictation (Feb. 2016 – March 2016):
– learning effects in 6 successive translation sessions
– different behaviour experienced users
● Compare with translation behaviour in other language pairs:
– 150 hours TP data in the TPR-DB
– EN → DA, DE, ES, HI, JA, ZH
ENJA15: Translation Experiment
● Ambition:Investigate variations of human translation processes across different translator profiles and translation modes: translation, post-editing, dictation
● Method:
➔ Translate six short English source texts (110-160 words) from English into Japanese under controlled conditions
➔ Record translation activities: transcribed speech, keystrokes and gazing
The ENJA15 ExperimentAugust 2015 to January 2016
● Each translator translated total 6 texts English-to-Japanese:1. from-scratch translation (T)2. translation dictation (D)3. MT post-editing (P)
● 39 native Japanese Translators ● 14 translators: 10+ years experience ● 17 translation students < 2 years experience● No experience in translation dictation
● 55 hours of English-to-Japanese user activity
● Data is publicly available free of charge
Number of Insertions per PU (1000ms)
● Longer chunks of coherent insertions in translation dictation (D)
Average Production InEfficiency
● InEff = insertions+deletions/length of final translation
● D involves more deletions per translation than T
P02 P04 P18 P19 P22 P24 P280.000
0.500
1.000
1.500
2.000
2.500
T
D
Average Translation Duration per Texts
● Text 3: most difficult across all translation modes
● Less variance in Post-editing
Multilingual Translation Experiment (MLE)
> 150 hours of translation process data (T,P,E,D,C)774 translation sessions 108053 ST tokens, 122323 TT tokens
Average Translation Duration per Language
● EN-to-{DA,ES,DE} easier than EN-to-{JA,ZH,HI}
● Text 3: more difficult independent from TL
JTD16: Longitudinal Translation Study
● ENJA15: Translators had no experience with ASR and TD
– Some liked it others did not
● Can TD be learned within a few translation sessions and will TD translation speed increase?
● 7 translators (from the ENJA15 experiment) get used to TD over six successive days
● Translate 2 texts each day on 6 successive days within 3 weeks
– Text type: company mission statements
– Amounts to 7 * 12 = 84 translated texts
Post-experimental questionnaire
● With your experience from today's translation session: is TD easier than HT (regular ‘typing’)?
1: Yes, it is a lot easier,2: It is slightly easier3: about the same4: It is a lot more difficult than HT
● Translators did not perceive significant improvement over the 6 sessions.
Day1 Day2 Day3 Day4 Day5 Day60
1
2
3
4
5
P2
P4
P18
P19
P22
P24
P28
Conclusion
● Translation Dictation and ASR:
– novel method of Human-Machine Interaction in Translation
– quicker than from-scratch, approximately as efficient as PE
– no learning effect over 6 days
● Why do translators produce longer chunks in TD:
– concurrent reading ST and typing easier than TT speaking?
– speaking interferes with phonological loop to a larger extent than writing?
● Why is there no learning effect
– better (more professional) training required?
– maybe more fundamental things going on?
● Which texts are suited for dictation, post-editing, translation
– How are processes different/similar?