ChatGPT was released a bit over two years ago, which seems like a good time to assess the impact of LLMs on the language business. On the core task of translation I'm more excited about the prospects now than back then:
1. LLMs are beating neural machine translation systems in open competition - see the report on the ninth conference on machine translation (WMT24).
2. Foundation models now include more multilingual content.
3. Several multilingual and translation-focused LLMs have been released in recent weeks: Cohere's Aya Expanse, Translated's Lara, Unbabel's Widn.ai and DeepL's LLM.
4. LLMs can deliver very accurate translations by using translation project assets like localization style guides, glossaries and translation memories via Retrieval Augmented Generation (RAG). RAG needs to be expertly implemented by TMS providers, which isn't always the case.
LLMs still require human quality evaluation via MQM, exposing gaps and flaws in translation project assets also requiring human preference input. I believe that in 2025 we will see further improvements in LLM-based translation with RAG and agentic systems, even without further breakthroughs in LLM technology.
85
Automated, local translations now in my favorite browser
mozilla.org/en-US/firefox/11…
68
DeepL is a unicorn. How good are its machine translations? Find out in this latest analysis distilled from the MT Decider Benchmark Q4/2022.
#deepl #evaluation #machinetranslation lnkd.in/d3PUeUEV
88
Can ChatGPT help with the core task of the language business - translation?
With the right prompts we can control ambiguity, bias, politeness level!
Controlling this via prompts is clunky and not something you'd want to use in day-to-day translations.…lnkd.in/gQtSVFAr
1
102
Happy New Year!
New year, new data - the Q4/2022 MT Decider benchmark is out with a new language pair, updated test data and tools, and quote-normalized BLEU for more accurate evaluation of linguistic MT quality. lnkd.in/dmYrZzXq
49
Developer post
mteval - an open source Python library to automate automatic machine translation evaluation
To efficiently evaluate many datasets for many languages with many MT services with multiple metrics we needed a library t…lnkd.in/dBUAiyB4 lnkd.in/dXsuTMgK
My pitch for the World-Readiness Contest at the TAUS Massively Multilingual Conference last month. Didn't quite make the cut for the win ... but minding the MT quality gap is still quite cool, I think. lnkd.in/gdTfD4r8
My presentation from the second LSP Tech Talk Online Event in late September. Thanks for the opportunity lingo systems | Technology and consulting for the language industry! lnkd.in/gzhFksJR
1
New blog post on the typography of quotation marks with surprising implication on machine translation quality evaluation ...
#machinetranslation #evaluation #typography lnkd.in/g7T_B_Vc
Good to be back to an in-person conference. Thanks for the opportunity TAUS! #TAUSMM2022 lnkd.in/ep7fwSWD
I'm looking forward to co-present the TAUS DeMT™ Evaluation Report 2022 with @aavaz and my views on how to scale machine translation evaluation for many languages in my "Mind the Gap" pitch
lnkd.in/dTMvSmER #machinetranslation #evaluation lnkd.in/dwnEAmE4
Interested in the importance of machine translation quality for LSPs and how to choose the best quality service? I'll be presenting at the online LSP Tech Talk this Wednesday web-eur.cvent.com/event/2089…
Are you getting the same quality from all online machine translation services? No, it differs a lot - for some language pairs more than others. Evaluated services: Amazon Translate, DeepL, Google Translate, Microsoft Translator. polyglot.technology/2022/09/…
Want to choose the best online machine translation services for your language pairs? Achieve an up to 54% quality improvement by selecting the best services with the new MT Decider Benchmark. polyglot.technology/2022/09/…
Going live in a few minutes ...
RESULTS: DEMT Evaluation Report (feat. @T21Century, @TAUS_Data, @AchimRuopp) #MT #machinetranslation #l10n #data teapawt.org/i/broadcasts/1eaJbNawk…
1
Achim Ruopp retweeted
RESULTS: DEMT Evaluation Report (feat. @T21Century, @TAUS_Data, @AchimRuopp) #MT #machinetranslation #l10n #data teapawt.org/i/broadcasts/1eaJbNawk…
Want to know which machine translation service performs best across many language pairs? The MT Decider Index has launched! Here is the index for the second quarter of 2022:
1. Google Translate
2. Microsoft Translator
3. DeepL
4. Amazon Translate
polyglot.technology/2022/07/…
I am looking forward to talk about the challenges and benefits of evaluating online MT systems in this upcoming webcast with @AnneMaj_TAUS, @amirkamran and @FatMotherTucker! lnkd.in/eEMxzbYT
Local machine translation engines are now available for Firefox blog.mozilla.org/en/mozilla/… via @BergamotProject @firefox
Is forecasting time series the same like predicting next words in language models? Not quite, but some of the ideas around trends, seasonality and serial dependence could be useful. Keep on learning. lnkd.in/dZzZY-sy





