The Death of Language Learning?

Machines are gaining linguistic abilities at an ever-increasing pace. Given that, will we humans continue to learn second languages? If so, in what ways? For how much longer? For what reasons?

Our starting point is to ask why people learn languages now. For some, it is to communicate. For others, it is to gain access to the culture. For yet others, it is to amuse or challenge themselves. Finally, there are those for whom it’s a way to differentiate themselves.

Today’s LLM technology can already handle cross-linguistic content astonishingly well, be it in the form of words, images, or sounds. We can place a phone running some app between ourselves and our conversation partner and carry on a meaningful dialog. This in itself reduces or eliminates the need for someone to learn Italian in preparation for their trip to Italy, for example.

The technology will certainly be integrated in the near future into various devices which will provide the same basic functionality but in a more integrated and less intrusive way. We might have AI pins, and AI glasses, and AI earplugs. This will further weaken the need to learn languages for social or even work situations.

Won’t latency be a problem? In other words, won’t we spend twice as much time communicating because the machine will need to wait for the original utterance to finish, then give its translation? Actually, this is a problem which is likely to be solved in the relatively near future by various optimizations and augmentations to the language models and surrounding systems. In the worst case, there might be brief pauses before and during the machine translating.

The above focused on real-time conversations. But of course communication also occurs in the form of written documents, or videos. However, there is nothing about that form of communication that the new generation of language tools cannot handle at least as well, or probably better, not to mention more easily, than either translating them by hand, or learning another language to be able to read them.

These are some cursory, initial thoughts about how language technology might impact the need for language learning to communicate in social and work situations. Bottom line: it will reduce the need to zero over time. But there are other reasons for language learning. How will machines affect those use cases? There is too much here to discuss in a single article. We’ll continue next month in July with a discussion of how language learning for other purposes—such as as a personal project or pastime, or to savor novels in their original language, or to connect heart-to-heart—might be affected by LLMs.

How not to predict the future of LLM language abilities

The problem of latency we mentioned above is a good example of how, in imagining how the future will unfold, we must remember that technology will continue to advance in ways and at a speed that we find it hard to imagine, both in the short and long term.

This is the fatal oversight—actually, one of many—in the recent position paper published by the Japanese Association of Translators called “Statement on the Public and Private Sector Initiative to use AI for High-Volume Translation and Export of Manga”. The initiative in question is the “manga-tech” startup named Orange, which recently received funding from Shogakukan and others.

JAT is not happy about this on a number of fronts:

First, in its current form, AI translation has yet to demonstrate the level of quality required to adequately portray nuance, cultural background, or character traits.

Leaving aside the question of whether today’s human translations consistently demonstrate such level of quality, this just makes us want to ask: well, then, what about in its future form? Why issue such a high-profile statement pissing on LLM manga translation when another six months or two years of technology development will change its conclusions? And the whole point of Orange and other manga-tech companies is to develop and customize and tailor and extend technology for manga, so we’re just at the starting point. Studies have already yielded promising results on using LLMs to translate manga. So JAT talking about AI translation in “its current form” is a pure red herring.

I guess it’s too much to expect that an organization like JAT, whose mission after all is to support and protect flesh-and-blood translators, would mount a last-ditch effort to maintain a dying historical process where only humans are supposed to translate manga, even when machines can already or will soon be able to translate higher volumes, at higher quality, at lower cost, and thereby actually stimulate the popularity around the world of manga in translation, which seems like should be considered a laudable goal (currently only 2% of Japanese manga are translated, it is said).

The JAT statement misses the mark in other ways. They seem likely to have been misled by the abysmal performance of historical MT systems on manga, But that’s completely unsurprising given the limitations of that technology, the uniquely visual nature of manga, and the lack of bilingual resources for training. They also probably don’t understand the ability for LLMs to accept prompts and examples that can greatly improve the quality of their output. They probably don’t understand that LLM-based translation can be an iterative process, where humans step up to the role of editors and reviewers and commentators—a perfectly natural evolution of the historical role of manga translator, one that the JAT should be embracing, not denigrating. They probably don’t understand the capability of LLMs to produce translation in different styles and registers for different audiences, which could greatly increase the world-wide popularity of manga. This is more than just technophobia; it’s Luddite ignorance.

Full disclosure: I recall having made a presentation at a JAT meaning many years ago. I also know several members of that fine organization, including manga translators who do astonishingly good work and have built great careers around their skills. I also once translated a manga called 童夢 (Domu), not very well I’m afraid. I wonder if that translation is still floating around somewhere.

Our point here is not about translation of manga per se—we’re focused on language learning—but rather the tendency on the part of observers and forecasters and commenters to ignore (in some cases it would seem willfully) the eminently predictable, exponentially accelerating advances that are inevitably going to occur in the coming years.

And one final point about manga and translating them. The JAT statement makes the implicit assumption that manga are and always will be written by humans, but in reality a greater and greater proportion of the work, include story design, graphics, and narrative is likely to be done by computers. This trend is already underway. So the question arises, if the machines are writing the manga, with the human role shifting to being to being one of guiding and directing and refining their work, why couldn’t the machines also produce English versions of their works right from the beginning, bypassing the entire notion of “translation” as we conceive of it now? And that is all the more so when one imagines a world of “manga-on-demand” where smart machines generate personalized manga directly in the language of the consumer. And come to think of it, if JAT really has the interests of its manga translators in mind, wouldn’t it be pushing for a model that does not involve people sitting for hours at a computer laboriously churning out English manga?

In the next article, we will take up the topic of how other motivations for language learning—such as pillow talk, or cultural understanding, or showing off—could be affected by AI.