Machine translation

Assembled from 30 blocks · 3 evidence · 27 related

1. Story
2. The first domain in which neural methods reached the threshold of ordinary usefulness, the domain the transformer was invented for, and the domain where "human parity" was claimed earliest and disproved most instructively. In 2018 Läubli, Sennrich and Volk re-ran a parity claim on the Chinese–English news task and found the claim survived only under sentence-level evaluation: when professional raters were shown whole documents, they preferred human translation clearly, because the errors that distinguish the two are errors of cohesion, reference and consistency that a sentence cannot expose. Coverage has since widened faster than quality has been established — Google added 110 languages in June 2024, taking Translate to roughly 243 and adding languages spoken by some 614 million people, using a general-purpose language model rather than paired training corpora. That is a real change in access, and it is not the same claim as a change in quality, because for most of those languages no evaluation of comparable rigour exists at all.
3. Knowledge
4. The transformer architecture
5. Machine translation
6. Labour-market effects of machine learning
7. Connections
8. Labour-market effects of machine learning
9. The transformer architecture
10. The transformer is presented at NeurIPS
11. A human-parity claim in translation fails under document-level evaluation
12. The second International AI Safety Report reports mixed labour findings
13. Machine translation
14. Machine-generated code
15. Measured adoption of AI by firms
16. Payroll data shows a relative decline in entry-level employment in AI-exposed occupations
17. A whole-labour-market analysis finds no economy-wide disruption yet
18. The second International AI Safety Report reports mixed labour findings
19. Radiologist workforce projections and employment outlook, against the 2016 prediction
20. The AI accelerator
21. Deep learning
22. Machine translation
23. Machine-generated code
24. Growth in frontier training compute
25. The transformer is presented at NeurIPS
26. GLUE is saturated and SuperGLUE is built to replace it
27. Evidence
28. Testing a human-parity claim on the WMT Chinese–English news task with alternative evaluation protocols, human raters assessing adequacy and fluency showed a stronger preference for human over machine translation when evaluating whole documents than when evaluating isolated sentences, indicating that errors decisive for quality are frequently invisible at sentence level. V55 verification basis: the paper was not fetched; search results carried the title, the ACL Anthology identifier D18-1512, the EMNLP 2018 Brussels venue and the substance of the finding and attributed them to this paper.
29. The transformer replaces recurrence and convolution with attention alone, is more parallelisable, and was shown superior in quality on two machine translation tasks while requiring significantly less time to train. V55 verification basis: the proceedings were not fetched; search results carried the author list, the venue, the Long Beach dates of 4–9 December 2017, the page range and the substance of the abstract and attributed them to the NeurIPS proceedings. The paper's arXiv identifier, its 2017 preprint date and its reported BLEU scores were explicitly NOT confirmed by retrieval and are not asserted anywhere in this pack.
30. United States employment of interpreters and translators is projected to grow 2% from 2024 to 2034, slower than the 4% average across occupations and revised down from a previous 3% projection, with about 7,500 openings a year projected mostly from replacement need; the reported median annual wage rose to USD 59,940 in 2024. Economic research summarised in the CEPR column reports slower growth in translator employment where machine translation use is higher, and reductions concentrated in routine rather than creative translation work. V55 verification basis: neither source was fetched. The BLS projection, the 4% comparison and the median wage were carried in search results attributing them to the Handbook, partly via industry reporting on the Handbook's revision; the elasticity and percentage-reduction figures appearing in retrieved summaries of the CEPR column were NOT traced to the underlying paper and are deliberately not quoted in this pack.