Hacker News Reader

Comment by chacham15 | Hacker News Reader

I remember learning in my machine learning class that translation is a solved problem. However, the reason it appears to be bad is because it is computationally expensive. There are different levels of translation, each of which are increasingly more difficult to compute. There is word for word translation. As you can imagine, this is the lowest level, easy to compute, but overall has poor quality. First, word for word translation, then syntax translation, semantic translation, and finally interligual semantic translation. I took that class a while ago, so forgive me if I'm not 100% accurate, but that is generally the issue: the higher level translation you want, the more expensive the computation.

View on HN

I don't think translation is even a solved problem for a human fluent in both languages. The ideal is to write what the original author would have written, if they were writing fluently in the target language. Sometimes that means they would have written something that doesn't have the same literal meaning. If a pun doesn't work in the target language, a fluent writer would have used a different pun, but how do you know which?

Even if you want to preserve literal meaning, an example that comes to mind, but which I can't quickly source: IIRC the original Metroid manual, in Japanese, referred to Samus with a gender-neutral pronoun. In English, you basically can't do that without calling attention to it. 'They' and 'he/she' are unusual enough to read awkwardly; 'it' and 'he' would be incorrect; and 'she' would give away the twist. Whatever you choose, you're losing something, and it's a judgment call as to what.

I believe that technically, in English, "he" is the correct gender neutral pronoun. It just feels wrong so people commonly try to find other options.

It's been surprisingly fluid over the history of English. One interesting bit to read about is the history of "singular they", i.e. the use of "they" in sentences such as "if a person wishes to gain access, they must enter the code". This was fairly common in pre-19th-century English, then fell out of favor in the 19th and early 20th centuries as grammarians considered it incorrect ("they" was deemed solely a plural pronoun), and now it's making a comeback as a gender-neutral singular pronoun. The construction "he or she" also has usage going back centuries, as a different approach to that. And as you note, the use of "he" as a stand-in pronoun is also traditional, but falling a bit out of favor lately. Overall I don't think there is one correct answer for how English deals with that situation; it varies across writers and eras.

As _delirium said, the singular "they" has been making a comeback and will likely earn its place in academic writing in a generation.

Until we get human level AI that's going to be impossible. But decent translation of "normal" writing is certainly possible and "good enough".

> The ideal is to write what the original author would have written, if they were writing fluently in the target language. [...] If a pun doesn't work in the target language, a fluent writer would have used a different pun, but how do you know which?

Umberto Eco, in Mouse Or Rat?[1], describes two different approaches to translation, (I'm paraphrasing here) source vs. target. Your example is a good one, but it actually only touches on grammatical differences. There are cultural differences too.

For example, if I'm translating a cricket (the sport)-themed story from Australian English into Japanese, do I retain the cricket references, and have the meaning be unintelligible to Japanese readers with no knowledge of cricket? This is what Eco refers to as source translation, which is closer to a literal translation.

Or I could change the cricket references to baseball, which is more appropriate in a Japanese context, and Eco refers to as target translation. The purpose here is to convey the feeling, more than remaining true to the literal meaning of the original text.

[1]: http://books.google.com.au/books/about/Mouse_Or_Rat.html?id=...

While it's true that more accurate translation is more computationally expensive, and we do have good knowledge of the structures and algorithms for machine translation, it's far from being a solved problem, mostly because it requires a relevant multilingual corpus for training (see [1]) That's also why translation performance varies greatly depending on which kind of prose you are translating (yes, there are attempts to translate poetry as well, see [2]).

There is a huge amount of potential information on the web but it's rarely well curated and the next big problem is how to extract it.

[1] See http://www.independent.co.uk/life-style/gadgets-and-tech/fea...

[2] http://googleresearch.blogspot.com/2010/10/poetic-machine-tr...

The lack of high-quality bilingual corpora is hardly the primary obstacle to considering automatic translation a solved problem. Even for language pairs for which plenty of "bitext" exists, the current approaches to machine translation still face major challenges with regards to ambiguities, idioms, finely nuances imagery and other semantic subtleties, all of which pose hard, but not insurmountable challenges to an expert human translator.

Machine translation is no where near solved.

Machine translation is still one of the most funded areas in NLP, and the quality is still incredibly bad for many language pairs. DARPA has been running one machine translation grant program or another for more than a decade.

It is true that some of the best systems take a long time to translate sentences (on the order of 1 cpu-minute or more). But their quality is still not anywhere near human level for most languages.

Translation is one of the hardest ML/NLP problem and it is far from being solved.

AFAIK, current state of the art translation systems are not working with a layered approach as you defined (morphology->syntax->semantics->interlingua?). First, they do not use interlingua. They use statistical models from example translations (translation model) and target language (language model). They started to include syntactic and semantic clues recently but it is mostly adding extra information to the models and alignment-search process.

So computational complexity is not the real issue when it comes to translation quality. For most (if not all) languages, systems simply cannot achieve human translators level yet. For languages that lack large amount of example translation texts, machine generated translations are still terrible.

Most of your details are correct, but your first sentence is wrong. Translation is nowhere near solved.

> translation is a solved problem

If that's the case, can you link me to the source code? I have a couple of extra computers I can run it on.

I think it depends what you mean by "solved" and your perspective on AI in general. If you think that human brains operate on the same rules as everything else in the universe, then it is theoretically "solved" in that we know the path to get there but just haven't been able to practically do it (yet).

Reply on news.ycombinator.com