The document summarizes an analysis of post-edited translations compared to human translations in terms of translationese principles. The analysis uses several datasets containing human translations, machine translations, and post-edits. Experiments measure lexical variety, lexical density, length ratio, and perplexity on parts-of-speech sequences to analyze differences between human translations, machine translations, and post-edits. The results generally show that post-edits exhibit characteristics that are between those of human translations and machine translations, indicating that post-editing does not fully remove the "footprint" of machine translation.
7. The Reader
Does PE affect the reading experience?
→ Are PE translations = HT?
3
8. PE vs HT. Theory and Practice
In theory, PE=HT
1. Translator is primed by MT output while post-editing (Green et al., 2013)
2. PE should contain the footprint of MT
3. HT should be preferred over PE
4
9. PE vs HT. Theory and Practice
In practice, quality of PE
• comparable to that of HT. E.g., Garcia (2010)
• or even better. E.g. Plitt and Masselot (2010)
5
10. PE vs HT. Theory and Practice
In practice, quality of PE
• comparable to that of HT. E.g., Garcia (2010)
• or even better. E.g. Plitt and Masselot (2010)
But... Quality is typically measured as number of errors (Koponen, 2016)
5
11. PE vs HT. Beyond Number of Errors
Characteristics of PE vs HT
• Czulo and Nitzke (2016) Terminology in PE closer to MT than HT
• Daems et al. (2017) Discrimination between PE and HT not possible
• Farrell (2018) lexical variability in PE<HT
6
12. PE vs HT. Translationese
Research has proven the existence of translationese: HT=original text
• Normalisation
• Simplification
• Interference
• Explicitation
7
13. PE vs HT. Translationese
Research has proven the existence of translationese: HT=original text
• Normalisation
• Simplification
• Interference
• Explicitation
This paper: quantitative analysis of PE vs HT in terms of translationese principles
7
30. Lexical Density
lexical density =
number of content words
number of total words
(2)
Content words: adverbs, adjectives, nouns and verbs (UDPipe)
12
31. Lexical Density
lexical density =
number of content words
number of total words
(2)
Content words: adverbs, adjectives, nouns and verbs (UDPipe)
→ Simplification principle
12
32. Lexical Density Results (Taraxu)
ende deen esde
0.4800
0.4900
0.5000
0.5100
0.5200
0.5300
0.5400
0.5500
0.5600
HT
MT
PE
LexicalDensity
13
33. Lexical Density Results (Taraxu)
ende deen esde
0.4800
0.4900
0.5000
0.5100
0.5200
0.5300
0.5400
0.5500
0.5600
HT
MT
PE
LexicalDensity
HT > PE MT
13
34. Lexical Density Results (all)
Translation Dataset and translation direction
Type Tarax¨u IWSLT MS
de→en en→de es→de en→de en→fr zh→en
HT 0.55 0.53 0.53 0.48 0.46 0.59
PE -1.00% -2.48% -4.31% -3.46% -1.24% -0.46%
MT -0.81% -0.69% -4.53% -5.14% -0.94% -2.37%
14
35. Lexical Density Results (all)
Translation Dataset and translation direction
Type Tarax¨u IWSLT MS
de→en en→de es→de en→de en→fr zh→en
HT 0.55 0.53 0.53 0.48 0.46 0.59
PE -1.00% -2.48% -4.31% -3.46% -1.24% -0.46%
MT -0.81% -0.69% -4.53% -5.14% -0.94% -2.37%
PE-NMT -3.88% -1.47% -0.46%
PE-SMT -0.54% -2.87% -4.78% -3.04% -1.09%
PE-RBMT -1.46% -2.09% -3.84%
14
38. Length Ratio
length ratio =
|lengthST − lengthTT |
lengthST
(3)
Hypothesis: compared to HT, PE translations are closer in length to source text
15
39. Length Ratio
length ratio =
|lengthST − lengthTT |
lengthST
(3)
Hypothesis: compared to HT, PE translations are closer in length to source text
→ Normalisation principle
15
40. Length Ratio Results (Taraxu)
ende deen esde dees
0.000
0.050
0.100
0.150
0.200
0.250
ht
pesmt1
pesmt2
perbmt1
perbmt2
lengthratio
16
41. Length Ratio Results (Taraxu)
ende deen esde dees
0.000
0.050
0.100
0.150
0.200
0.250
ht
pesmt1
pesmt2
perbmt1
perbmt2
lengthratio
HT > PESMT ≥ PERBMT
16
42. Length Ratio Results (all)
Dataset Direction
Length ratio
HT PE MT
Tarax¨u
de→en 0.16 -38.5% -36.9%
en→de 0.22 -33.4% -38.5%
es→de 0.17 -25.2% -21.0%
IWSLT
en→de 0.17 -3.4% -18.8%
en→fr 0.18 6.7% -10.9%
MS zh→en 1.41 -9.9% -9.1%
17
43. Length Ratio Results (all)
Dataset Direction
Length ratio
HT PE MT
Tarax¨u
de→en 0.16 -38.5% -36.9%
en→de 0.22 -33.4% -38.5%
es→de 0.17 -25.2% -21.0%
IWSLT
en→de 0.17 -3.4% -18.8%
en→fr 0.18 6.7% -10.9%
MS zh→en 1.41 -9.9% -9.1%
Competence missmatch. PE=prof, HT=anyone
17
44. Perplexity on PoS Sequences
Process:
1. PoS tag monolingual corpora (Universal Dependencies tag set) for source
and target languages
2. Build language models on PoS tagged data
3. PoS tag each translation (MT, PE and HT) and calculate:
18
45. Perplexity on PoS Sequences
Process:
1. PoS tag monolingual corpora (Universal Dependencies tag set) for source
and target languages
2. Build language models on PoS tagged data
3. PoS tag each translation (MT, PE and HT) and calculate:
PP diff = PP(translation, LMsource) − PP(translation, LMtarget) (4)
18
46. Perplexity on PoS Sequences
Process:
1. PoS tag monolingual corpora (Universal Dependencies tag set) for source
and target languages
2. Build language models on PoS tagged data
3. PoS tag each translation (MT, PE and HT) and calculate:
PP diff = PP(translation, LMsource) − PP(translation, LMtarget) (4)
Hypothesis: PP diffPE < PP diffHT
18
47. Perplexity on PoS Sequences
Process:
1. PoS tag monolingual corpora (Universal Dependencies tag set) for source
and target languages
2. Build language models on PoS tagged data
3. PoS tag each translation (MT, PE and HT) and calculate:
PP diff = PP(translation, LMsource) − PP(translation, LMtarget) (4)
Hypothesis: PP diffPE < PP diffHT
→ Interference principle
18
53. Conclusions
PE=HT. PEs:
• Are simpler (lexical variety and density)
• Are more normalised (length ratio)
• Have more interference from the source language (PoS sequences)
21
54. Conclusions
PE=HT. PEs:
• Are simpler (lexical variety and density)
• Are more normalised (length ratio)
• Have more interference from the source language (PoS sequences)
MT paradigms
• (PE)SMT better than (PE)NMT in lexical variety and density
• (PE)NMT has less interference than (PE)SMT
21
56. Discussion
1. Does PE contribute to the impoverishment of the target language?
2. In this study HT better than PE. But number of errors HT≥PE
22
57. Discussion
1. Does PE contribute to the impoverishment of the target language?
2. In this study HT better than PE. But number of errors HT≥PE
• PE may be better suited than HT for some domains, e.g. technical
22
58. Discussion
1. Does PE contribute to the impoverishment of the target language?
2. In this study HT better than PE. But number of errors HT≥PE
• PE may be better suited than HT for some domains, e.g. technical
3. It’s not the fault of the post-editing process per se... but of MT
22
59. Discussion
1. Does PE contribute to the impoverishment of the target language?
2. In this study HT better than PE. But number of errors HT≥PE
• PE may be better suited than HT for some domains, e.g. technical
3. It’s not the fault of the post-editing process per se... but of MT
• PEs should be better if MT is. E.g. interference in PE-NMT<PE-SMT
because interference in NMT<SMT
22
60. Future
• Effect of PE guidelines, translator’s expertise, etc.
• Measures with deeper linguistic information
• Automatic discrimination between PE and HT
• More data (industry?)
Data and code available: https://bit.ly/2zeKf0b
23
61. Thanks: L. Bentivogli, S. Castilho, J. Daems, M. Farrell, L. Macken, L. Marg
and M. Popovi´c
23
62. Thanks: L. Bentivogli, S. Castilho, J. Daems, M. Farrell, L. Macken, L. Marg
and M. Popovi´c
Go raibh maith agaibh!
Ceisteanna?
Antonio Toral
@ atoral
23
63. References i
References
L. Bowker and J. Buitrago Ciro. Investigating the usefulness of machine
translation for newcomers at the public library. Translation and Interpreting
Studies, 10(2):165–186, 2015. ISSN 1932-2798. doi: 10.1075/tis.10.2.01bow.
URL http://www.jbe-platform.com/content/journals/10.1075/tis.
10.2.01bow.
64. References ii
O. Czulo and J. Nitzke. Patterns of terminological variation in post-editing and
of cognate use in machine translation in contrast to human translation. In
Proceedings of the 19th Annual Conference of the European Association for
Machine Translation, EAMT 2017, Riga, Latvia, May 30 - June 1, 2016, pages
106–114. European Association for Machine Translation, 2016. URL
https://aclanthology.info/papers/W16-3401/w16-3401.
J. Daems, O. De Clercq, and L. Macken. Translationese and post-editese : how
comparable is comparable quality? LINGUISTICA ANTVERPIENSIA NEW
SERIES-THEMES IN TRANSLATION STUDIES, 16:89–103, 2017. ISSN
0304-2294. URL https://lans-tts.uantwerpen.be/index.php/
LANS-TTS/article/view/434/409.
65. References iii
M. Farrell. Machine Translation Markers in Post-Edited Machine Translation
Output. In Proceedings of the 40th Conference Translating and the Computer,
pages 50–59, 2018.
R. Fiederer and S. O’Brien. Quality and machine translation: A realistic
objective. The Journal of Specialised Translation, 11:52–74, 2009.
I. Garcia. Is machine translation ready yet? Target. International Journal of
Translation Studies, 22(1):7–21, 2010.
S. Green, J. Heer, and C. D. Manning. The efficacy of human post-editing for
language translation. Chi 2013, pages 439–448, 2013. doi:
10.1145/2470654.2470718. URL
http://vis.stanford.edu/papers/post-editing.
66. References iv
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/.
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.