An array of neon signs of nonsense words on a wall. Image generated by the Stable Diffusion AI algorithm.

Polyglot in the Machine: AI for Language Learners

AI is the order of the day lately. Have you seen how many fantasy photos have been filling up Instagram lately? Thanks to the now wide availability of open source AI algorithms, some powerful computing power is in the hands of users courtesy of apps like Dawn AI and Lensa. Type in a few words, and the computer does the painting.

It’s new tech, opening new possibilities alongside new ethical challenges that users are gradually becoming sensitive to. But the benefit to individual language learners here is apparent very little imagination stretch. First and foremost, these algorithms parse human language. So why not, for instance, type in some target language – say, ein Hund mit grünen Augen (a dog with green eyes) – and see if the picture matches what you meant to say? It should act as a kind of machine validation that the language you produce makes sense.

It already works to a point with some languages. Models like Dall-E (seen at work below in the web-based Craiyon.com) cope reasonably well with non-complex, non-English prompts.

A screenshot from Craiyon.com, a web-based AI image generator built on DALL-E Mini.

It can be hit and miss, but Craiyon understood my German for the most part!

So it works – up to a point. The current stumbling block is linguistic and cultural bias. For a start, models like Stable Diffusion were initially developed and trained with English input. And as one web experimenter shows, non-English results can leave a lot to be desired, with a definite advantage for Western European languages. This isn’t surprising, given that the technique samples from pre-existing web content; the predominance of certain languages means there is a lot more of that to learn from.

Ai Work In Progress

It’s clear these techniques are nascent and emerging, as most casual users will admit. Even if English is your target learning language, for example, images can frequently be so off the mark that you may question whether it understood a single word of your prompt.

Things are improving, though, especially with regular updates to the Stable Diffusion model. There are even a couple of language augmentation projects floating around in beta, including. one that adds ‘Japanglish’ capabilities to the current algorithm, overcoming one particular cultural blindspot.

And, if you have the skills, you can add to many ongoing open source projects to extend and finesse the capabilities of AI algorithms. I’m sad to report that that isn’t in my skillset, but it’ll be interesting to follow how this develops over the coming months!

One thought on “Polyglot in the Machine: AI for Language Learners

Leave a Reply

Your email address will not be published. Required fields are marked *

This site uses Akismet to reduce spam. Learn how your comment data is processed.