How Machine Translation Works
Antonio Toral
Panel 3: Literary translation and AI (CEATL Strasbourg Conference), 4 October 2024
Intro: Opening the Black Box
Comment fonctionne la traduction automatique
MT
System
How Machine Translation Works
1
Table of contents
1. NMT 101: Quality
2. LLMs 101: Flexibility
3. Post-editing
4. Technology for Inspiration
2
NMT 101: Quality
NMT training: predict the translation
Output sentence
Parallel MT Monolingual
Data System Data
(sentence pairs) (sentences)
Input sentence
3
NMT key 1: words as concepts
spain = {2, 1, 6}
Forcada (2017)
4
NMT key 2: all words interconnected
ENC-layer 2
ENC-layer 1
Input word
embedding
machine translation is fun
ENCODER
Slide by A. Bisazza. Machine Translation Couse (BA in Information Science at U. Groningen) 5
NMT key 2: all words interconnected
la traduzione
DEC-layer 2
DEC-layer 1
Output word
embedding
ENC-layer 2 [BEGIN] la
DECODER
ENC-layer 1
Input word
embedding
machine translation is fun
ENCODER
Slide by A. Bisazza. Machine Translation Couse (BA in Information Science at U. Groningen) 6
NMT vs SMT
7
NMT vs SMT
EN-DE | EN-FR
Type text here:
The Budapest Prosecutor’s Office has initiated an investigation on the accident.
Translation:
Die Budapester Staatsanwaltschaft hat ihre Ermittlungen zum Vorfall eingeleitet.
Slide by A. Bisazza. Machine Translation Couse (BA in Information Science at U. Groningen) 7
NMT vs SMT
EN-DE | EN-FR
Type text here:
The Budapest Prosecutor’s Office has initiated an investigation on the accident.
Translation:
Die Budapester Staatsanwaltschaft hat ihre Ermittlungen zum Vorfall eingeleitet.
Slide by A. Bisazza. Machine Translation Couse (BA in Information Science at U. Groningen) 8
LLMs 101: Flexibility
LLM Training 1: predict the next word
9
LLM Training 1: predict the next word
Output text (continuation)
Textual LLM
Data
Input text
9
LLM Training 1: predict the next word
Input: teach me how to bake bread
A likely output:
10
LLM Training 1: predict the next word
Input: teach me how to bake bread
A likely output: in a home oven
10
LLMs Training 2: instruction tuning
https://ogre51.medium.com/instruction-fine-tuning-of-llms-a-comprehensive-guide-e2f197e19c36
11
LLMs vs NMT for translation
Much more flexible
Less literal translations (Raunak et al., 2023)
Tower (Alves et al., 2024), a LLM then specialised on translation tasks,
seems the current state-of-the-art (top rank at WMT 2024)
12
Post-editing
Post-editing
Post-editing (PE) compared to human translation
Less creative translations (Guerberof and Toral, 2022)
Lower lexical variety and higher source language interference (Toral, 2019)
Translator’s voice partially lost (Kenny and Winters, 2020)
Less enjoyable task, demoted (Moorkens et al., 2018)
13
Post-editing
Post-editing (PE) compared to human translation
Less creative translations (Guerberof and Toral, 2022)
Lower lexical variety and higher source language interference (Toral, 2019)
Translator’s voice partially lost (Kenny and Winters, 2020)
Less enjoyable task, demoted (Moorkens et al., 2018)
But, in terms of errors (Koponen, 2016), post-editing
comparable human translation. E.g., Garcia (2010)
or even better. E.g. Plitt and Masselot (2010)
13
Technology for Inspiration
Inspiration. Example 1 (Youdale, 2019)
Use of technology to support distant reading
Corpus tools to complement close reading
Measure style: sentence length, repetitions...
Better informed translations
14
Inspiration. Example 2 (Kolb and Miller, 2022)
PunCAT: a tool to assist with the translation of puns
Aims: facilitate brainstorming and provide inspiration
Automatically translate each sense of the pun
Allow user to explore the semantic fields of these translations
15
Inspiration. What about MT? (Guerberof and Toral, 2022)
16
Inspiration. What about MT? (Guerberof and Toral, 2022)
The [MT] proposal gave me an
adjective, I wouldn't have
thought of.
The problem when you translate is
that you have lots of words in your
head but you don't know how to
reach them.
Carlota Gurt
literary translator and writer
16
Inspiration. What about MT? (Farrell, 2022)
Survey: 86% of translators that use MT, use it for inspiration
Example
“I translate passages or sentences myself and then use the MT on the source
text to see what it comes up with, and I may adjust my translation on that
basis or indeed completely ignore the MT text. The MT never takes the lead
but can sometimes be useful as a supplement.”
17
Merci de votre attention!
Antonio Toral
a.toral.ruiz@rug.nl
https://antoniotor.al/ceatl24.pdf
17
References i
References
D. M. Alves, J. Pombal, N. M. Guerreiro, P. H. Martins, J. Alves, A. Farajian,
B. Peters, R. Rei, P. Fernandes, S. Agrawal, et al. Tower: An open
multilingual large language model for translation-related tasks. arXiv preprint
arXiv:2402.17733, 2024.
M. Farrell. Do translators use machine translation and if so, how? results of a
survey held among professional translator. In Proceedings of the 44th
Conference Translating and the Computer. Tradulex, 2022.
References ii
M. L. Forcada. Making sense of neural machine translation. Translation spaces,
6(2):291–309, 2017.
I. Garcia. Is machine translation ready yet? Target. International Journal of
Translation Studies, 22(1):7–21, 2010.
A. Guerberof and A. Toral. Creativity in translation: Machine translation as a
constraint for literary texts. Translation Spaces, 11(2):184–212, 2022.
D. Kenny and M. Winters. Machine translation, ethics and the literary
translator’s voice. Translation Spaces, 9(1):123–149, 2020.
References iii
W. Kolb and T. Miller. Human–computer interaction in pun translation. In
J. Hadley, K. Taivalkoski-Shilov, C. S. C. Teixeira, and A. Toral, editors, Using
Technologies for Creative-Text Translation. Routledge, 2022. To appear.
M. Koponen. Is machine translation post-editing worth the effort? A survey of
research into post-editing and effort. Journal of Specialised Translation, 25
(25):131–148, 2016. ISSN 0169-2607. URL
https://sites.google.com/site/wptp2015/.
J. Moorkens, A. Toral, S. Castilho, and A. Way. Translators’ perceptions of
literary post-editing using statistical and neural machine translation.
Translation Spaces, 7(2):240–262, 2018.
References iv
M. Plitt and F. Masselot. A productivity test of statistical machine translation
post-editing in a typical localisation context. The Prague bulletin of
mathematical linguistics, 93:7–16, 2010.
V. Raunak, A. Menezes, M. Post, and H. Hassan. Do gpts produce less literal
translations? In The 61st Annual Meeting Of The Association For
Computational Linguistics, 2023.
A. Toral. Post-editese: an exacerbated translationese. In Proceedings of the XVII
Machine Translation Summit, Dublin, Ireland, August 2019.
R. Youdale. Using computers in the translation of literary style: challenges and
opportunities. Advances in Translation and Interpreting. Routledge, United
Kingdom, June 2019. ISBN 9780367141233.
