The Shallowness of Google Translate (by Douglas Hofstadter) | Hacker News Reader

The Shallowness of Google Translate (by Douglas Hofstadter)

theatlantic.com

54 pointsjsomers24 comments

Previous comments: https://news.ycombinator.com/item?id=16267363

Superb reminder of the difficulty of true NLP. Google Translate is an amazing and very useful tool, but machines have a long way to go before they really understand language.

Example sentence used by article:

> In their house, everything comes in pairs. There’s his car and her car, his towels and her towels, and his library and hers

I agree with the thrust of the article, that Google Translate is fundamentally limited (as most systems are) by its inputs and ostensible goals. But what does it mean to "really understand language"? What is the recognizable benchmark for "true NLP"?

At some point, how "good" comprehension/translation is very subjective, because it intersects so heavily with culture and history. Think of how debates still rage over interpreting Shakespeare among Western readers and academics, even though there isn't a language barrier.

I notice you quoted "really understand language." That is not a quote from the essay. Is it an indicator that you are taking the conversation in a different direction?

Hofstadter has written a lot about the question of what it means to "really understand a language", and the related more relevant question of what it means to translate from one language to another. Several of his essays in "Metamagical Themas" in the 1980s covered those topics, and they are quite interesting.

He gives an example "Mary was sick yesterday", with different levels of 'understanding'. Call it "sentence M" (quoting from https://archive.org/stream/MetamagicalThemas/Metamagical%20T... , including apparent typos)

> 1. Sentence M contains twenty characters.

> 2. Sentence M contains four English words.

> 3. Sentence M contains one proper noun, one verb, one one adverb, in that order.

> 4. Sentence M contains one human's name, one linking verb, one adjective describing a potential health state of a living being, and one temporal adverb, in that order.

> 5. The subject of Sentence M is a pointer to an individual named 'Mary', the predicate is an ascription of ill health to the individual so indicated, on the day preceding the statement's utterance.

> 6. Sentence M asserts that the health of an individual named 'Mary' was not good the day before today.

> 7. Sentence M says that Mary was sick yesterday and

> Just where is the boundary line that says, "You can't do that much processing!"? A machine that could go as far as version 7 would have actually understood-at least in some rudimentary sense-the content of Sentence M.

Elsewhere in the same book he (like in this Atlantic essay) talks about the problems of translating gender across languages.

I think it was in Godel, Escher, Bach where he talks about the problem of translating a line from Russian, along the line of "He lived on B____ street." It's possible from the story, which takes place in a real city, to figure out which street that was. Let's say it was "Main Street". Does the English translation keep the original Russian word and initial? Does it translate to "Main Street" and replace the "B" with an "M"?

Or consider his book "Le Ton beau de Marot: In Praise of the Music of Language", which contains 88 different translations of a 16th-century French poem.

I think it would be difficult to incorporate all of those thousands of pages of writing on the topic into a single essay for a lay public magazine, and that your expectations are too high.

While it's true that "good" is subjective, this is a solved problem in the Turing test or Chinese room sense. We judge professional translators, like those who work at the UN. We judge students learning a foreign language. There's no reason to believe that we can't apply similar techniques to judge machine translation.

Indeed, he gives many concrete examples of a minimum level of translation competency which should be expected for a good-enough system.

I agree that the OP gives good and reasonable examples of what should be doable by a competent translator. I think he is entirely on point about how people fail to realize how shallow Google Translate in terms of what we understand as "understanding" language. My quote about "really understand language" is from the GP -- I was asking him what he meant by that phrase.

I don't think I disagree with the OP much at all. But I was confused because much of his essay shows how Google Translate is getting the most basic things wrong. But then he ends with discussions about what it means for a computer to have true understanding, the type of understanding that can't be achieved with just more data.

It seems to me that the obvious screw ups that Google Translate is demonstrated to make could be alleviated through better algorithm design and data -- I.e. without achieving what Hofstadler argues is true understanding. Just like a self-driving car could be very safe despite having no more deep understanding of driving than Google Translate does of words.

My apologizes, I missed up on the thread context.

This sort of conversation of "really understand" is Searle's Chinese Room thought experiment. Hofstadter and many others have written a lot on the topic.

The classic joke in the '80s was that a systran project meant to translate Russian into English wasn't too good outside of purely technical work - The proverb 'The spirit is willing but the flesh is weak' became 'The vodka is good but the meat has gone bad.'

The joke dates from the 1950s. See https://www.google.com/search?tbm=bks&q=%22The+whiskey+is+ag... .

For example, https://books.google.com/books?id=90Y8AAAAMAAJ&q=%22The+whis... show the snippet from a 1958 publication:

"Univac cannot anthologize, though it has all sorts of language tricks, and can translate "The spirit is willing but the flesh is weak" to "The whiskey is agreeable but the ..."

and here's a 1965 reference of the same joke: https://archive.org/stream/journalofkentuck6319kent#page/272... .

wowzers - cool - I learned it in the early '80s from a guy who taught me - remembering some weird computers that usta be lying about, not at all surprised to see the barb directed at univac then - I believe it was aimed at Sperry in my day :-)

Seems Google has update their translation model?

For the Chinese one, at least:

>> After one year of working in Tsinghua University, Zhong Shu was transferred to Mao's translation committee to live in the city and back to school on weekends. He still holds the post of graduate student.

The leader of the Mao Selected Translation Committee is Comrade Xu Yonglian. Introducing Zhong Shu to do this job is Tsinghua classmate Qiao Guanghua.

On the appointed day, after dinner, an old friend hired a rickshaw to come from the city to congratulate. After the guests go, Zhong book said to me in fear:

He thought I had to do a "Southern study walk." This is not a good thing to do.

Now they correctly singles out person's name, "锺书(Zhong Shu)", as comparing to transliterate it as Book(meaning of the character Shu). Even with that 南书房行走，IMO, it did a not bad job, at least knowing it is its own entity, not to break into parts then translate.

As a native speaker, the style of the example text provided is quite elegant and old-school. 南书房行走 is a very confusing phrase, it looks like a verb but used as a noun phrase, and without context, it is hard for me to tell the meaning of it.

The updated version of the translation is pretty serviceable. Google Translate works best with functional text, like news/report, etc. Not quite there with literature, which is well known, probably on purpose. As someone works on MT project, this quality is pretty amazing. I won't necessarily say it is shallow, TBH.

>As someone works on MT project, this quality is pretty amazing. I won't necessarily say it is shallow, TBH.

Kind of exactly what this article is saying. To the people who work on this stuff, it's fascinatingly accurate and servicable, but to regular people it's a very poor substitute for actual bilingual understanding.

I wonder about those samples. The Chinese I would expect to be kind of garbage going off BLEUs from published papers, but the French and German should be great. Am I imagining it, or do they just not sound like RNN translations? The RNN translations in my experience tend to flow well and are comprehensible, albeit sometimes just wrong because they picked a wrong semantic interpretation. But these samples sound like the older phrase-based translations, where they immediately descend into word salad. As people noted at the time, the jump in quality for French/German was so large that people knew when the new translations went live without Google publishing anything... (I assume Hofstadter at least was careful enough to do his sampling after the new RNN translations were opened up.)

Google doesn't discuss how it rolls out the translation upgrades, exactly, and it's an uneven deployment. Can any Googlers comment on the possibility Hofstadter was using the old translations? Or can any NMT researchers compare and contrast his examples with current SOTA models?

“Am I imagining it, or do they just not sound like RNN translations?”

You’re imagining it. Google translate is about as bad as he’s saying. The translations are usually servicable for short phrases, but quickly fall apart for anything with depth or nuance.

The Google paper was an interesing technical description, but pretty weak in terms of evaluation: it had a single experiment, based on simple, short phrases, evaluated in a subjective manner by “experts”. It shouldn’t be surprising at all to find out that the method was a bit overhyped.

It’s quite possible that Google did greatly improve their performance...it just wasn’t very good to begin with.

The French translation in the article matches what Google returns today (ie garbage), albeit for one difference (it now translates "in pairs" to "par deux" instead of "en paires" (which is a slight improvement, but doesn't address the fundamental errors of the translation).

The garbage it gives is consistent with the garbage statistical approaches tend to give; particularly when translating from a language without gendered pronouns to a language with them.

My favorite, most concise example to demonstrate this is the sentence "my cousin and her wife". Anyone with basic understanding of English grammar would infer that my cousin is a woman married to a woman; Google Translate gives me back a French sentence where suddenly my cousin has become a man.

This is a great example of something that a rules based translation system would never get wrong (of course, rules based translation systems have plenty of other shortcomings) and that statistical approaches have a hard time dealing with.