Discussion AI Translation

https://www.theregister.co.uk/2018/10/24/ai_translation_baidu/

So, how long would it be before we can just download an add-on for firefox/chrome and the ai translate chinese with quality like a good translator?

 

Less than 2 decades. Mark my word.

 

idk, but i can read google translate just fine

 

For Chinese? Probably decades. It's hard to translate idioms and made up words, yet CNs are full of them. Plus, progress is slow.

edit: After skimming the article, it introduces a technique for live speech translation. This means it offers nothing new to novel translations, which already have the entire text available. Instead, it is targeted towards interpreters too facilitate conversation. It won't replace language precise interpreters, such as contract translators, or novel translators. Web novels are too wacky to translate well.

 

Well it won't be an offline addon since AI to be efficient it needs different kind of processors. As for online addon, how would that be different from web translation like google translate?

That said, you can probably get an AI to be almost as good as a translator already. AI is based on how much you train it. And many of the translation services are generic by nature. If you translate 100 chapters and feed that data into an AI so that it gets all the terminology, context and style right, it can probably translate the rest of the novel with little issue.

 

That training dataset is too small. Plus, you would need top-tier translations and I don't think there are that many. Plus, terminology usually ventures by novel - heck, a MC might swap all his techniques between arcs.

 

It depends on the type of literature, a xianxia who constantly repeats itself is pretty easy to deal with. Techniques can be a problem, sure. But an editor can easily read through the chapters after the initial translation, then adjust the techniques and retranslate with all the techniques in. It might take a few passes, but it should be doable.

Aka, the draft re-visioning approach.

 

I think read somewhere ai looks at patterns and data to find the difference like how humans know what a cat and dog looks like but ai can not, they need lots of pictures to accurately tell if the one being shown is a cat or not. language may be used for ai cause words basically a meaning and words are influenced by race, religion, culture, traditions, and media. maybe in a decade, we may have some access to a good enough translation of each other's language. but then I fear that ai do not have the human "feeling" like for example the sentence "I will kill you" sounds threatening but if you are given context, like a maybe something like a dog ate your homework. the ai do not know that there are meanings even to our actions, face, body language, and interactions with each other based on our relationship. so not only must ai understand humans, it must understand our very existence, even translators sometimes have trouble translating what the author really means because of word plays, maybe a reference to something. but you know the good news, the very second the ai can understand "humans" is the very second it knows of its existence as it compares your life to itself. scary

 

Teach me master!

 

Supposing two swallows carrying it together?

 

Honestly, it's just a technological problem. Once we can get like 10-50k cores per cpu, we can actually start throwing out the optimization algorithms and just brute force the learning.

 

Not remotely as easy as you think. First what about new terms? New names? Change in writing style? What if author make shit up or take idioms but change characters in them to be clever? Unless you had few dozen of books written by one author then you probably be able to train an AI to translate his work pretty well assuming he doesn't change up his style too much.

There simply too many none standard writing in stories that make them terrible for machine translation. Like standard document translation is like almost perfect with just google translate now but novels are complete shit.

 

We are about 100 years far away from that, mostly because something like an "IA" doesnt exist yet. The so called current IA is just a character chain comparation with the recorded database. Something like....., this chain of characters is close to this one, so this might be the correct result, just the product of a math process, a math algorithm.

....but a human has something called abstract thinking wich is on a completelly different level and that's also used on the language.

I'm going to place some examples. The translation from english to spanish is more or less decent, because english is a relatively simplier language, but when you carve on the translation you can see many big mistakes. For example, something like "BOOM" used like a sould will be automatically translated like "success", why? Just because the machine doesn't have the knowledge to understand the text, that's something fitted into the abstract thinking.

I said that english to spanish is relatively decent, but what happend to the inverse thing? The spanish to english translation is a complete shit (even worst than my english ), because spanish (like all the latin languages) is an extremelly complex language. It has many conjugations, many gramatical artifacts, the same word may means hundred of different things depending of the context, etc (and you must add the different spanish variations depending of the country). That's something totally out of the posibilities of the current IA (and the near future ones)

It's the same case for the japanese to english. Speaked japan language is also relatively simple, but when we are talking about japanese writting we are on the opposite, it's something extremelly rich and complex. A combination of kanjis could mean something completelly different depending on the context. It explains why some translations from japanese are close to be decent and others are an absolute piece of crap.

If you are a translator don't be worried, you won't lose your job on this life

 

African or European?

 

It depends on how far the progress of machine learning. My bet normal mtl is like currently edited mtl on this forum in the future. Readable but lost in translation potential is bigger. I remember back then in translation corner someone mention word that translated as godfather, but in novel context it mean sugar daddy.

 

But, of course, African swallows are non-migratory. So they couldn't bring a coconut back anyway.

 

I'm afraid translation isn't as simple as many people like to believe. The whole idea of using information from other translated text for the machine to translate more text is actually what Google translate is based off of already, and it has been at it for years. It uses hundreds of millions of documents as reference. I think we all know how accurate it is. Machine translation may eventually get to a point where it's decent, but that's still a ways off.

The problem is in the nature of communication. It's a lot more complex than we usually think about. As you're reading this you are attaching a meaning to every single word. You are further referencing previous words to help in deciding their meaning. Even more, you are reading all of this in context to everything that was written before it, just for a little more meaning. A machine attaches no meaning to any of this. It's shapes. Ones and zeroes. If you ask a machine to explain "compassion", it could give you a dictionary definition, or maybe point you to a few documents or videos with a "compassion" tag to them, but it couldn't give you its own example. This is the problem we already face with making a machine that can communicate like a human would at all.

A similar problem applies to translating. Machines translate (this is a little oversimplified of course) word "A" = word "B", or "A" = one of "B", "C", or "D", and with grammar rules: "the verb goes in X part of a sentence". This runs into a lot of problems; we don't always follow grammar rules, and the accurately translated word might not be on the list, or the wrong word might be chosen. The computer has no idea. It has no way of knowing, because none of it means anything to the computer. Machine learning tries to get around this by only having the machine reference complete translated text documents and statistically guess what the most accurate translation would be. Unfortunately this has limitations too. The machine has no idea what the writer's intents are (again, it means nothing to the machine), and that can make a huge difference on how things should be translated. Also the texts it references might tend toward one field of interest or another, which would change the most statistically likely translation. The machine does not understand genre, or nuance, or even how writing changes over time. Heck, it can't even tell what makes for a natural sounding sentence. So it can't factor any of that in when translating.

In other words, there's a fundamental problem to work out before we're going to get "actual translator quality" translations from machines.

 

Idioms, puns, jokes, cultural references and names are all basically impossible for a machine to get to say the least. Human translation of novels isn't going anywhere in the next century, until people develop a real AI at least.

Official documents and stuff like that works fine though~

 

What the AI is currently unable to incorporate. is probably localized pidgin or slangs. It's a common sub-language in any culture.

So the references in those types of sayings probably wont be translated properly or will be literally word-by-word.

I mean there is Urban dictionary for a lot of English slangs, and it is so diverse, because its dependent on the local culture. A slang word in one part of the USA, may have a different variation in another part of the country.