We need to talk about Gaelic and Machine Translation, specifically we need to talk about the use of unedited Machine Translation.
Machine translation (MT), most notably Google Translate has become increasingly common in recent years. Using MT at all raises many practical, ethical and philosophical questions and I could write a series of articles on this subject alone.
Today, however, I’m looking particularly at the use of unedited MT. This is quite simply when someone puts some text into Google Translate in one language and then uses the supplied translation without doing any editing or checking whatsoever on it.
If you don’t know a language and want to get the gist of what something is saying, Google Translate is a good and easy way to do it. However, it is highly likely to contain mistakes and hallucinations and isn’t a substitute for a real translation and a human opinion is always necessary for any important matter.
I’m raising the issue of unedited MT just now as I’ve seen it used a few times in real world situations in the last few months, including by public bodies which I find concerning.
I wrote a blog post last year about the lack of merchandise featuring Gaelic, where I believe the tourist and arts and crafts industries are missing a trick. At a couple of craft fairs and shops in the run up to Christmas, I saw a few products featuring Gaelic. I was delighted and was keen to buy some until I saw that the Gaelic was wrong, clearly an MT job by someone who doesn’t understand the language.
Then for Christmas, I was kindly given a Kobo e-reader as a present by my Mum. I was keen to see what Gaelic books were available. There were indeed many great Gaelic books available but also some books which were MT slop which were clearly English books fed into MT and put straight up on the internet for sale with cringeworthily bad Gaelic. Any Gaelic speaker would notice straight away but not perhaps a parent of children in Gaelic medium education or someone wanting to buy a book for a Gaelic speaking friend.
More recently, I wrote to ScotRail in Gaelic with a question about tickets (namely whether they planned to put RailBus tickets to Glasgow Airport on the ScotRail app). They said they don’t in fact reply to Gaelic correspondence in Gaelic but that I could translate their website content into Gaelic using an MT button on the website.
I’ve seen a few organisations suggest that the availability of an MT button on websites offering translation to many languages is provision for Gaelic and supporting the language. This shows a fundamental lack of understanding of the position of Gaelic which I’ll look into more below.
Most worryingly of all, a tech company which has supported Gaelic a lot in the past, using experienced translators has now moved some of its Gaelic provision from translators to unedited Machine Translation.
This has me very worried for the future of Gaelic reading and service provision.
Why? Let’s take a look.
Rupert, Rupert the bear!
OK – so where does Rupert the Bear come in?
My family are very big fans of Rupert the Bear. In fact, my parents actually had an ongoing argument as to what colour our plucky little ursine hero is – brown as shown on the cover of his books, or white as shown in the comic strips themselves.
Regardless of the colour of his fur, imagine that Rupert is having an adventure in Cornwall with some Cornish pixies or the like.
I am a big fan of the Cornish language and am impressed by the amazing progress made in the language revival over the last 100 years. However, there were some bumps in the road for the language, the most notable of these being a major fracture in the movement in the 1990s about the best written form to use. Cornish was largely revived from written sources and a great deal of the Middle Cornish available was written in verse.
Commentators in favour of placing more emphasis on the latest available written sources were concerned that putting too much emphasis on poetic language had a somewhat stilting event on the revived language.*
Linguistically, one of the key characteristic features of the Rupert adventures is that the stories are written in rhyming verse.
We’ll never have to recreate the English language based on Rupert the Bear (great fun though it would be) but imagine that we were basing machine translation into English on AI trained on Rupert comic strips.
“Machine translation is terribly new – And golly gosh it’s in poetry too!”
In a way, Gaelic machine translation is like that. Basically, what comes out in MT is dependent on what goes in as it depends on what is available in the target and source languages online.
For Gaelic, online sources are not medieval plays in verse or rhyming comic strips but are overwhelmingly translated official documents like annual reports, Gaelic language plans, educational resources, news reports and such like. It also contains a lot of old Gaelic books that have been digitised with OCD but never proofed and other “wild west web” type content. And even within the formal documents available, these are more likely to be documents about education or local service provision, rather than say relating to science or engineering issues. In terms of breadth, there isn’t a huge amount of literary, colloquial or lower register material available online (though social media has increased this to some degree).
This means that what is available is somewhat stilted to the formal and high register rather than more informal everyday formats. This isn’t ideal. It means that everyday subjects can sound more stilted than they would be in natural Gaelic.
“Move over rover, and let Jimi take over!” – MT and corpus
Public Domain, https://commons.wikimedia.org/w/index.php?curid=17011402
As a minority language, Gaelic isn’t as standardised as English.
Don’t get me wrong, Gaelic is very standardised by minority language standards and dialect difficulties are minor.
However, in the issue of terminology, Gaelic isn’t as standardised as it might be, particularly for new and specialist terminology. A lot of work is needed to make sure that terminology of this type is agreed, understood and used. This is a part of what is called Corpus Planning. The more text that is written in Gaelic, and the wider the range, the more important that Corpus Planning becomes.
How does MT relate to Gaelic corpus planning? It doesn’t. Not at all. In fact, it works against it.
If specialist and new words are uncommon and are therefore not all over the internet for AI to see, MT may not pick up on the words that translators and experts use and will simply invent its own terms which are not likely to be as suitable or as readily understood.
Another problem is that Gaelic often doesn’t make distinctions between things in specialist writing in the way that English does due to a historic lack of use in formal documents. In the example of a risk assessment, for example, the English might use risk, danger and hazard in the same document. In Gaelic, most people would use “cunnart” for all three in the first instance and in translation, careful thought would have to be given. MT would be simply likely to come out with cunnart, cunnart and cunnart!
Furthermore, MT is inconsistent. In their 1967 song “The Wind Cries Mary” the Jimi Hendrix Experienced asked:
“Will the wind ever remember, the names it has blown in the past?”
For Google Translate, the answer is simply no. It will not consistently use the same term for the same word. If you put a document into MT in a programme that breaks text into different fields or if you put material into MT separately, it will often come out with different words for the same thing.
“I feel fantastic” – hallucination / uncanny valley and context
Like AI in general, MT hallucinates. It particularly does this if there are typos or punctuation marks in the wrong places and as a translator, I can confirm that there are almost always small mistakes in the source text. If you are translating from Gaelic to English, the problem is worse as spelling isn’t often as consistent.
MT is prone to inventing place-names. You are fine for major placenames but again, for smaller placenames that aren’t all over the internet, it will invent English equivalents for Gaelic names (say small hills or glens in the Highlands where there is no English equivalent) or Gaelic equivalents for obscure English names which don’t have a Gaelic equivalent. I’ve found it is particularly keen on creating Gaelic names for Orkney and Shetland where it uses the Gaelicised Norse names in the Hebrides to invent similar Gaelic names for the Northern Isles! This undermines the good efforts of Ainmean-àite na h-Alba, lead national corpus planning group for Gaelic place-names.
Never forget, if in doubt, MT always prioritises sounding correct over being correct!
MT also creates an uncanny valley. It says things in Gaelic which are not wrong grammatically but are a bit strange and unnatural. If you were a language teacher, you’d say “well, that’s not wrong, but you just wouldn’t say it!”
Uncanny Valley – picture of Tara the creepy android who spawned many internet conspieracy tehries
Perhaps most importantly is context. Good translation requires context and often there isn’t enough in translations fed into MT to ensure that a correct translation is given.
This is particularly the case for websites, one of the areas where unedited MT is often recommended as a way of accessing Gaelic content by the website owners. A lot of web text, particularly buttons and headings, is short and without context.
This matters. Does “refuse” mean diùlt (deny) or does it mean sgudal (rubbish)? Does “open” mean open as an adjective or as a verb? I also recently saw some MT generated text where “charge” as in electrical charge was translated as financial charge and “current” was translated in a temporal rather than an electric sense. Without enough context, this issue will never be ironed out.
Idioms are an issue too. For common idioms, no matter how different Gaelic and English versions are, MT should be able to handle it to some extent. But for less common idioms, where they are not all over the internet in writing, this isn’t the case and this is never likely to be solved. I was wondering about what the Gaelic would be for “to hold a candle to” the other day. For “this doesn’t hold a candle to”, Google Translate suggested a literal translation: “chan eil seo a’ cumail coinneal ris” which is a heinous calque and just doesn’t make sense in Gaelic.
The moral of the story is:“Machine translation can be bad, use a human translation and you’ll be glad!”
Anna Frater and Steven King ** – Gaelic speakers and reading
When I started learning Gaelic at university back in 1991, the amazing Gaelic poet Anna Frater was my Gaelic tutor and I was also lucky enough to have speaking classes with Mina Smith.
As well as learning the language, I had a learning moment about Gaelic reading. I was sitting my Gaelic exam and Anna was an invigilator. She had a paperback with her which was in English.
The young, ignorant and somewhat patronising Alasdair asked himself “why does Anna have an English book when she could have a Gaelic book? If I was fluent in Gaelic, I’d read Gaelic books all the time”.
I’m really embarrassed looking back on this now. People should read whatever they like in whatever language they like and restricting reading of English on ideological grounds would just be silly.
More pertinently, what I didn’t realise then was that there were in fact almost no Gaelic novels in print at the time and that someone could probably read all the Gaelic books in existence in the span of a few years. Things have moved on a lot since then, of course, in terms of written Gaelic but the point remains. I was thinking with a monolingual mindset and didn’t understand that Gaelic speakers have a different relationship with reading Gaelic and reading English.
A monolingual mindset is often the problem here.
Very few Gaelic speakers are as confident in reading and writing Gaelic as they are in English. Many extremely fluent native speakers didn’t get a chance to learn to read Gaelic at school or aren’t confident in reading Gaelic. Many younger fluent speakers are L2 speakers of Gaelic or have more confidence in their English reading skills than in their Gaelic skills.
Often too, Gaelic speakers tend to read about Gaelic in a narrower range of registers or domains than they would in English due to the reading material available.
And of course, Gaelic speakers are bilingual and have access to a near-infinite amount of reading material in English whether it be for pleasure in the magical world of literature or for necessity in the world of work reports and gas bills!
For the reasons above, material in Gaelic needs to be good and easy to understand and as idiomatic and natural as possible or people simply won’t read it. Translation in Gaelic is about status as well as corpus. It is about creating good material in Gaelic that encourages Gaelic speakers to read and increase their confidence in reading. Using unedited MT undermines this.
Does anybody remember the happy days of the early 80s where you’d buy a TV, a new-fangled microwave oven or your first video recorder and the instructions would invariably be appallingly translated from Japanese or Chinese? We’d struggle with the English instructions because we had no alternative but we’d never choose to read material like that if we had a choice.
Remember the halcyon days of the early 80s when we were encouraged to use microwaves to make Sunday roasts?
Why then would anybody choose to use a button that translates a website or document from good English into flawed Gaelic that might well contain inaccuracies?
Using MT is great for working out the gist of something in a language you don’t understand (I use it for reading articles on Wikipedia about French railways for example) but quite simply, nobody would ever use it to translate material from a language they understand perfectly to a language which they are less confident in reading.
Another danger is that if anybody is monitoring the usage of Gaelic on automatic translation buttons on their website, it will give them the false impression that there is no demand for written material in Gaelic. What is actually means is that there is no demand for AI slop.
Poisoning the well
Using unedited Machine Translation is simply poisoning the well for written Gaelic.
I saw a good example of this recently when the Bluesky interface was translated into Gaelic. I saw somebody saying “I see that Bluesky has been translated into Gaelic but it will just be machine translation so I’m not going to bother looking at it”. It was in fact translated by a translator but the perception that it might be MT was enough to put this person off and I suspect that this is a feeling that many people have.
Quite simply, people won’t read things they know or suspect to be unedited Machine Translation. Even worse, readers might thing that MT generated slop has been created by an actual translator which might bring real, live translators into disrepute. Either way, the more unedited MT is used, the more people will be reluctant to read any Gaelic.
Economic and cultural effect
Using unedited MT has risks for translators by reducing the work available. There is also a bigger issue about the use of post editing of MT driving down the pay for linguists which I don’t have time to cover here. These are not just abstract figures, these are real people, people you’ve met at cèilidhs, people who taught you a song, a story, the genitive case or the dental blocking rule who are might now find themselves in the unenviable position of having to reconsider both the cost of living crisis and a potential career swerve.
Conclusion:
Unedited Machine Translation is bad for Gaelic and rather than creating a lot of Gaelic reading material at virtually no cost it can actually discourage Gaelic reading and impoverish the language and damage the Gaelic economy.
“Unedited MT is bad for us and for me, it’s time to throw it into the sea”
Alasdair
*I have no dog in the fight in the great Cornish orthography debate and am not an expert. I don’t know whether the argument about language written in verse is justified or not and I’m aware that the counter argument is a lack of later written material but I’m just using it here to illustrate a point. I am, however, delighted that there is a standard written form and that the language continues to grow and thrive. I think Gaelic has a lot to learn from Cornish.
**I don’t think the book was actually was Steven King but it makes good clickbait!
If you are still reading, don’t forget to check out the excellent punk version of the Rupert the Bear theme tune by the chronically overrated Sunderland band the Toy Dolls!
Powered by WPeMatico
Tadhail air Trèanaichean, tramaichean is tràilidhean
