Whose brain’s in the garbage bin?

(or Musings of an old translator about AI).

This morning I received a communication from an online group I’m in. The author wrote: “Well, it's clear to me that in the near future (too) many translators will rashly embrace the use of this new Stupid Robot, throwing their own brain in the garbage bin. I am leaving the group.”? That’s sad, I thought.? And surely a bit extreme.? ?To me such a comment implies a belief that the use of Large Language Models (LLMs) ?for translation and the application of acquired intellectual skills (i.e. the ability to translate) are somehow mutually exclusive or even incompatible. ?Was such a belief based on the assumption that the output of a Large Language Model ?is accepted “as is” and incorporated into the localization workflow without further review?? There are certainly use cases where the publication of unedited or LLM or MT output has been adopted as a corporate policy,? but is there any evidence that this is actually happening across the board?? Surely most output from automatic translation is channelled into some kind of CAT tool, or sophisticated translation management environment. ?Such CAT tools have provisions for review and the consideration of a range of proposals which can include 100% TM matches, fuzzy matches and the outputs of a number of MT tools.? Professional translators can reject them all, even the so-called 100% matches (which, of course, can be completely wrong), modify them in the light of their specialist knowledge, or accept them.? ??

At a meeting I recently attended,? some translators were? not opposed to doing this type of work so long as it did not result in a drop in their income. Employers and clients of translators undoubtedly have their own ideas of how much per hour they are prepared to pay translators for what has commonly become known as “post-editing work”.? How much time a translator will typically spend on a segment in a CAT tool will largely depend on the type of text being translated, the intended purpose of the translation and the perceived quality of the LLM translation output. This is a matter of negotation which is likely to be of more concern to practitioners than researchers and developers. ?

When I used to review automatic translations within the memoQ environment for a major international corporation, I would open every segment (our workflow never involved locked segments).? The documents translated were mostly technical specifications, the vocabulary of which was known to the MT system being used in those days, so I rarely had to check terminology. ?At sentences like “Unscrew the refill cap by turning it anti-clockwise”, I would simply click through to the next segment – does anyone know a better way of putting that? ?A sentence such as “Maintenance operations on the lifting gear must be carried out by a senior maintenance engineer and a fitter” would have given me pause for thought. Is that clunky? Can I make the meaning clearer? Is “senior maintenance engineer” right? ?Should I switch things around and make the agent the subject? ?Or should I just leave the sentence ?alone? Going back to the title of this post, my brain would certainly have been engaged and not sitting in the garbage bin! ?Did I feel devalued? Not a bit of it. ?Not everyone, however, seems to feel the way I did.?

But it’s not just in the field of translation that there is concern about the potential impact of AI on jobs and income.? A recent Goldman Sachs study?found that generative AI tools could, in fact, impact 300 million full-time jobs worldwide, which could lead to a "significant disruption" in the job market.? Media workers, financial analysts, paralegals and legal assistants were among the occupations highlighted as being at risk.? In medicine, too,? it looks like the need for human intervention is being reduced by systems incorporating Artificial Intelligence.? AI is being used in cancer rsearch and diagnosis, and an AI tool can accurately identify normal and abnormal chest X-rays in a clinical setting, according to a study published in the publication Radiology. ?This ability of AI to perform tasks formerly carried out by medical practitioners who express some concern about how their work will be valued in the future has led the National Institute for Healthcare and Care Excellence (NICE) to issue guidelines on the adoption of digital technlogy which are intended “to drive innovation into the hands of health and care professionals”. ?NICE takes the line that the aim in introducing AI-enabled healthcare technology is to support, not replace, healthcare professionals who will ideally come to understand and “own” these new systems. ??The form of such ownership will undoubtedly be determined in discussions between medical authorities and the radiological institutions over the coming years: ??will less radiologists be employed, or will the existing numbers of ?radiologists be deployed more usfeully?

According to Bard (yes, why not ask it?),? the expression "take ownership of a system" means to assume responsibility for the system and its performance. This includes identifying and addressing problems, making improvements, and ensuring that the system meets the needs of its users. ?Translators have a long history of being adopters of innovation. When I started working as a translator it didn’t take me long to realise I could make more money by dictating my translations on a cassette recorder and sending my tapes to a typist than I could by bashing them out on my own Adler, partly because of the time wasted correcting my typos with the dreaded Tippex. ?When the electronic typewriter came along offering the possibility of correcting typos and rephrasing sentences in memory before committing a translation to paper, I borrowed the money to buy one.? In 1985 my first desktop computer (or workstation) cost around ￡3000. I borrowed the money to buy that because it would open the door to the ￡33000 project which set me up in business.? My story is not unique. ?Many employed and self-employed translators would embrace new technology and often become the go-to-guys for anyone around them with computer problems. ?With a few exceptions, most translators are keen and informed users of translation memory software by which they do not feel demeaned or diminished.

In fact, translators have driven the development of CAT systems, and those of us who used the early Trados software which only worked from within MS Word will hardly recognise today’s sophisticated self-contained translation management systems like Trados Studio and memoQ . As we know, Generative AI encompasses far more than translation, but translation is what translators know about, translation is their gig. ?So, will translators be replaced by AI?? Well, what is likely is that translators who are conversant with the inner workings of AI systems and see them as an opportunity rather than a threat will be more employable than translators who consider AI an insult to their own intelligence. Here again we are talking about “ownership”. ?To cut to the chase, ?ownership means not restricting our involvement with AI to being mere passive (and perhaps angry) recipients of AI generated output. ?In business, knowledge is power when it can be harnessed to dispel the ignorance of others to our financial advantage. ?Translators have always progressed their careers and gained power by acquiring specialist? knowledge that bolsters their status as experts in their field, that makes them stand out as genuine professionals. ??We have all met and admired the Japanese-English translator who can understand and explain a complex chemical equation with the same ease as his scientist clients. He can turn up for a meeting with wild hair, stubbled chin and in torn jeans and we are still in awe of him, or rather his brain. ?As knowledge workers,? translators ?should consider AI as another specialism they need to master to enhance their power and authority in relation to employers or clients. ??

There’s no shortage of YouTube videos and blogs giving us an introduction to machine learning and AI. That’s a start, but translators would be well advised to master two specific areas of activity that will give them the confidence to take the onward march of AI in our stride. These are prompt engineering and fine tuning. Prompt engineering is the process of writing text that can be interpreted and understood by a generative AI model. A prompt can range from a simple instruction such as “Translate the source text from English into Spanish” to a sequence of detailed directions setting out the style, the tone and the register of the translation, e.g. “Translate the input text into Spanish, using the correct technical terms employed in the automotive industry.? Remember that the text is intended for publication in a popular daily newspaper in Argentina. ?Use simple language and ?employ appropriate Argentinian words and phrases”. ?The success of the translation generated by the LLM will be dictated by the level of detail and pointedness of the prompt.? The ability to write such prompts is a valuable and valued skill.

The other way translators can bring their expertise to bear on the performance of LLMs is by becoming involved in the fine-tuning process.?? Fine tuning is one of the ways an LLM can be adapted to fit a particular use case. ?Through this machine learning technique a pre-trained model is further trained on a new dataset to improve its performance on a specific task. ?For example, a general-purpose model will be customized to translate pharmacological texts. ?There is no point in using a dataset full of inappropriate terminology or translation errors which would debase rather than enhance the baseline model. Uncurated data is just noise. Translators will surely play a vital role in the compilation, curation and audit of specialist ?bilingual datasets;? they will ?be the go-to experts when it comes to providing assurance that LLMs are generating top-quality translations. Is there anything professionally demeaning in this? ?

The major providers of CAT solutions like RWS are already integrating ChatGPT and GPT4 into their product portfolios.? Custom.MT have a memoQ plug-in to link to these OpenAI models, and other developers are undoubtedly also working on harnessing the translation capability of other LLMs ?like Llama2 and Falcon within CAT tools.? The ChatGPT Translator Plus app which I developed has several ways of generating a CAT-tool-ready TMX file. ??These developments bring the translation output of LLMs right into the ?control centre of the translation process, which is what the translation management system has become.? A ?pilot will need considerable skill and experience to land a small single-engine aircraft on a dirt-track airfield during a storm, and may feel that’s what being a pilot is all about. Is the pilot who uses autopilot to land an airbus on smooth tarmac on a balmy afternoon any less a skilled pilot? ?Should they feel demeaned or insulted because a computerised system has taken control of much of the landing process?? Should a translator reject a translation proposal simply because it has been generated by an LLM? ?Is there any professional future for translators who won’t even consider changing their way of working?

As long as there are chip manufacturers and energy suppliers, artificial intelligence is here to stay. As a correspondent recently wrote “The genie is out of the bottle”. ?Top-notch transcreators and ultra-specialist translators are unlikely to find a gleaming new robot seated at their desk anytime now. In the broad, not clearly defined area of “general” translation LLMs are already generating output that is barely distinguishable from the work of human translators in what are called the “high-resource” languages. Many professional translators have already incorporated DeepL or ChatGPT into their workflow and regard its output as their “working draft” or as a means of brainstorming, or use it simply to conduct research. ?As I mentioned above, a few days ago I participated in an online meet-up on AI and the future of the translation profession organised by the Institute of Translators and Interpreters. ?Before the meeting I had a foreboding that I would be swamped with feelings of doom and gloom and yet more talk about brains going into the garbage bin. I could not have been more wrong. The translators in my group were keen to know more about LLMs and expressed a wish to explore how they could incorporate these new developments into their own career paths. There was concern, yes, but certainly no outright hostility. ?Most agreed that with their skillset professional translators would still be in demand although they may eventually have a different job title (“content reviewer”, “translation auditor”, ?“language consultant”?)? as? the “machine” is increasingly seen as the translator. ?But what’s in a name as long as the price is right?

Should anyone wish to see for themselves how ChatGPT translates and how translators can control the LLM’s output, ?my ChatGPT Translator Plus app is available at https://mydutchpal.com/shop.

?