As 2024 comes to a close, it is time to reflect on the most popular stories, trends, innovations, and themes that made the Slator headlines throughout the year, highlighting key developments in the language industry.
Here is a selection of stories that attracted the most attention and engagement from our readers around the world.
Will Large Language Models Edge Linguists Out of the Language Industry?
One of Slator’s most-read stories in 2024 detailed a May 2024 paper from the University of Zurich and Georgetown University that explored the role of linguists in the evolving field of machine translation (MT). The entrance of large language models (LLMs) has reduced the reliance on linguists for grammar and semantic coherence while designing a system.
However, the authors concluded, there are a number of points in the process where linguistic expertise is still essential. These include building parallel corpora for MT; developing technology for low-resource languages; and identifying linguistic phenomena that may present challenges for a system. Linguists can be especially helpful as humans and machines interface, for example, by designing effective human evaluations and reliably assessing advancements in the field.
Google Translate Ditches Tool for Detailed Human Feedback
Google retired its longstanding human feedback tool, Contribute, which allowed users to press a button and submit an alternative translation.
Slator reported in April 2024, Google’s announcement, in which the company acknowledged Contribute’s role in improving Google Translate, explained that since launching the tool in 2014, “our systems have significantly evolved, allowing us to phase out Contribute.”
Users can, however, still submit feedback by rating a given translation “good” or “poor,” and, for the latter, selecting a reason from a drop-down menu — a less involved process that speakers of low-resource languages worry might halt improvement of MT for their languages.
Live Speech-to-Speech AI Translation Goes Commercial
Just one month into 2024, an increasing number of language AI researchers — from academia to private companies — had already begun to focus on live speech-to-speech translation (S2ST).
This only accelerated the adoption of live S2ST across multiple commercial applications thanks to LLMs, which kicked off in mid-2023, with models such as Meta’s SeamlessM4T and Google’s AudioPaLM.
Slator’s rundown of real-world use cases included business meetings, where Microsoft Translator, integrated with the Teams meeting app, provides real-time speech translation in more than 30 languages through Azure AI services. KUDO and Interprefy specialize in real-time AI speech translation for live events and conferences.
Even the high-stakes world of healthcare presents an opportunity for expansion, especially for providers already offering voice technology for healthcare clients. Orion Labs, for instance, offers live speech translation via its Push-to-Talk 2.0 platform.
Introducing Revamped New Translation Quality ISO Standard 5060
Published in February 2024, ISO 5060 applies not only to language services providers (LSPs), but also to in-house translation departments and individual translators. While it specifically provides guidance for human evaluation of translation output, it can be used for workflows involving human and machine translation, with or without subsequent post-editing.
The International Organization for Standardization (ISO) established a framework based on “bilingual examination of target language content against source language content,” with the goal of standardizing evaluations so they do not differ significantly from rater to rater.
There are seven main categories of errors, which can be classified as critical, major, or minor: terminology, accuracy, linguistic conventions, style, locale conventions, audience appropriateness, and design and markup.
Translation AI Agency Lengoo Files for Bankruptcy
In March 2024, Lengoo filed for bankruptcy in a Berlin court, with German news sources pegging Lengoo’s accumulated losses between USD 8-16m.
Christopher Kränzler, Alexander Gigga, and Philipp Koch-Büttner founded Lengoo in 2014, originally as an online platform for automating project management and administrative tasks.
Starting in 2018, investors such as Redalpine, Creathor Ventures, Piton Capital, Inkef Capital, Techstars, and Polipo Ventures expressed confidence in Lengoo’s developing proprietary translation system, with Lengoo raising USD 34m by February 2021 — making it a long way for the LSP to fall.
Amazon Flags Risks of Training LLMs on Web-Scraped MT
Training LLMs at scale relies on massive amounts of training data scraped from the web. A January 2024 research paper from Amazon investigating the prevalence and quality of MT on the web found that a “shocking amount of the web is machine translated” into many languages.
And oftentimes, that MT output is low-quality, raising concerns about the quality of training data for LLMs. Researchers also noted a selection bias toward “shorter and more predictable sentences,” potentially from low-quality English content machine translated into many lower-resource languages.
The pervasiveness of low-quality MT in training data, the authors warned, could lead to less fluent models with more hallucinations, particularly for low-resource languages.
Translators by Any Other Name
Slator’s January 2024 roundup of five polls from 2023 was crowned by the most voted-on — and perhaps introspective — question: Will the term “translator” disappear in the next five years? Close to half of respondents said no, with just over 30% saying it will “definitely” or “possibly” disappear in that time period.
Inspired by a SlatorPod interview with ASAP-translation.com CEO Jakub Absolon, another poll asked whether readers agreed with Absolon, who suggested the term “full post-editing” should not be used, and should be priced as human translation.
More than 65% of readers agree that the term should not be used, while 18.7% want to keep using it. The remaining 16% are happy to use whatever term the client prefers.
Other polls touched on inflation, with nearly half of respondents reporting flat rates; ChatGPT, which 80% of readers reporting they do not use it for translation; and the beloved Microsoft Language Portal, used “often” by 46.5% of respondents.
Real-Time Speech Translation Stars in Biggest OpenAI Release Since ChatGPT
OpenAI has not slowed down since being credited with unleashing accessible AI to the masses. The company’s May 2024 release of GPT-4o offered a range of new or improved capabilities. The single new model was trained end-to-end across text, vision, and audio, with all inputs processed by the same neural network, reportedly with enhanced performance in around 50 languages.
A demo of GPT-4o featured a brief conversation with OpenAI CTO Mira Murati asking the system a question in Italian, to which GPT-4o responds in English. Cue the hot takes of ‘RIP translators’ and shares in language learning resource Duolingo dipping 3%. OpenAI planned to launch support for GPT-4o’s new audio and video capabilities to a small group of trusted partners in an API within a few weeks.
EU Parliament Issues a New 2024 Call for Tenders for Translation Services
A February 2024 notice posted for translation services would cover translation of single and multiple source language documents in 24 languages for four European institutions: the European Parliament’s Directorate-General for Translation; the European Court of Auditors; the Committee of the Regions of the European Union; and the European Economic and Social Committee.
While the notice did not mention MT, it did specify output metrics for source and target languages, and contracts — with one lot per language, assigned to a primary contractor and up to four secondary contractors — are estimated to last up to 60 months. No specific budget was listed. Once awarded, the contract will become effective January 1, 2025.
Bankrupt Dutch LSP, WCS Group, Quickly Bought by France’s Powerling
In a provisional January 2024 ruling, a Dutch judge suspended payments by LSP WCS Group to its creditors, appointing an administrator to negotiate until a later hearing a few months later. Of 14 companies under the WCS Group, only one was listed as in “suspension of payment” status; all others are listed under “bankruptcy” status. At the time, WCS Group’s website listed 3,247 active freelancers, whose next steps were unclear.
Just a few days later, French LSP Powerling acquired WCS Group for an undisclosed amount. Powerling, which already had a presence in the Netherlands — plus France, Hong Kong, and the US — said the move was in line with the company’s goal of clearing EUR 25m in revenues by the end of 2024 through acquisitions in Powerling’s main markets.
