Great Go, Glitchy Grammar
If there are sentient beings on other planets, then they play Go.—Emanuel Lasker, chess world champion
|
| ACM Prize in Computing source |
David Silver, the 2019 ACM Computing Prize winner, is highlighted this month in the main journal of the ACM—the Communications of the ACM. He is featured in the “Last Byte” article at the end.
Today we discuss one aspect of the article in regard to machine assistance.
The article highlights his work on computer programs that play Go, under the title “Playing With, and Against, Computers.” Silver led a Google DeepMind project that used learning methods to make the program AlphaGo, which beat the many-times world Go champion Lee Sedol in March 2016.
The victory was only the beginning; subsequent iterations of the algorithm have been able to learn without any human data or prior knowledge except the rules of the game and, eventually, without even knowing the rules. Very impressive. The versions that Silver refers to include AlphaGo Zero and finally AlphaZero. We covered the 2016 match here and discussed AlphaZero here. See also this explanation of AlphaZero.
The Last Article
Each month the last article is special. It starts on the last page 
The article is also always in the form of an interview: a series of questions and answers by the person being highlighted. This month the author, Leah Hoffmann, asks Silver questions about his work on Go playing programs.
One Question
About midway through the interview, Hoffmann poses the following:
Eventually, you built a system that learned to play Go on a smaller, nine-by-nine sized board.
Silver answered:
We had some successes in the early days on the small-sized boards. Our system did learn, through these very principled trial-and-error reinforcement learning techniques, to associate different patterns with whether they would lead to winning or losing the game. Then I started collaborating Sylvain Gelly at the University of Paris on a project called MoGo, which became the first nine-by-nine Go championship program.
What is wrong with his answer? What is interesting about this answer, from a Google point of view? Any thoughts before you turn the page? Well, posts do not have pages on the blog, so we’ll use a section heading as divider.
Our Answer
The answer is that the response is not correct grammar. Do you see? It should be:
Then I started collaborating with Sylvain Gelly at the University of Paris on a project
We find it fun to see an article about robotic methods that can play the world’s best Go, yet misses a chance to discover a simple grammatical error. There are lots of programs, free ones even, that catch this one. (Microsoft Word 2016, however, seems not to.)
A Google Fix?
This gave us the idea of seeing whether existing Google tools could fix it automatically.
One available tool is Google Translate, and this recalled the idea of our old post titled, “Can We Translate English To English?” That idea, which I (Ken writing this part) occasionally use when communicating with chess officials in many countries, is to try repeated translations into the target language until you get one that translates back to English the way you want it. Here, we have freedom to use any language L as the intermediary.
Using just the last sentence in the quote, here is what I get when I translate into L and then back. People used to write in Latin even for fellow speakers of their first language, so I tried Latin:
Latin inserted a missing preposition, but not the right one. Thinking of a language with similar richness of grammar to Latin but more native speakers who are currently alive (so that Google gains more data year over year), I tried Russian:
This got the main part right, except for ‘y’ coming back as ‘i’, but wobbled on the second half—which is frankly a mouthful in English. To multiply the number of native speakers by 10, I went to Chinese. Google offers the choice of “Traditional” or “Simplified” Chinese, but coming back is the same:
Going via Chinese got “with” but lost the specificity of “collaborating” and also messed up the second half. So I tried completely the opposite idea: an artificial language with zero native speakers and a limited vocabulary—Esperanto.
Suddenly it is perfect. What is curious is that the Esperanto puts no preposition between “kunlabori” and Sylvain Gelly, who looks like a direct object the same as in the mistaken English original. But Google knows to insert the extra “with” on coming back.
We’ll invite you to have fun trying other languages. Among ones I tried, only Italian and Portuguese matched the exactness of the correction via Esperanto. But all of them fixed the issue of “with.”
Open Problems
Does this show a commonality among human languages in their entirety that can help us improve our experiences with any one of them?
In particular, when I am explaining a technical point to a native speaker of another language, I sometimes check how GT renders what I write into that language, to make sure it works. Once, in Russian, I edited the Russian to make it come back to English as I wanted—and communicated that paragraph dually in Russian. Another time, while communicating with a chess arbiter born in Hungary, I composed a poem in Hungarian that way:
Az éjszakai bagoly mindig ácsorog
De a zümmögése soha nem nyugszik
és a szeme látja az összes sarkot.
A számok örök folyása
Megvillantja azt, ami sötét számunkra
és megmutatja a silókban elrejtett igazságot.
For those who don’t know Hungarian: The poem is grammatically correct, but does not make that much sense, and has subtleties that a native would not use – like a bagoly would gubbaszt, and not ácsorog.
Also, I have found the puzzle to find “what is wrong with the answer” really unfair when your post is full of other grammatical typos, like “Go a on smaller”.
And just to contradict myself, maybe what bothered me was not as much ácsorog~perched used with the bagoly=owl, but that it had no location specified. Similarly in English I don’t think you can simple say that “The bird perched.” without adding where.
The “Go a on smaller” typo is also in the original article. We mouse-copied the text and did not notice. My English original for the owl poem was:
The night owl is always perched
But his hooting never stops
and the machinery grinds on.
The eternal flow of numbers
Flashes what is dark to us
and shows the truth hidden in the silos.
The context was that the arbiter unexpectedly got a quick turnaround from me after 1am my time; my statistical report showed some issue and the poem was self-reference.
I see, in that case, I suppose I just don’t get poetry… It is quite interesting that “machinery grinds on” got translated to “his eyes see all the corners” – I wonder if it was to get a rhyme.
Whoops—while revising the poem, I did change it in English to “his eyes see all the corners”—and it was indeed to get a (“weak”) rhyme.
Using deepl.com, and going to French and back to English, I get the following which seems fairly good!
“In the first few days, we had some success on the smaller boards. Our system learned, through these trial-and-error reinforcement learning techniques, to associate different patterns with winning or losing the game. Then I started collaborating with Sylvain Gelly at the University of Paris on a project called MoGo, which became the first nine-by-nine Go championship program.”
Reading the Chinese, I think the main issue is that “championship program” got interpreted as “championship project.” That’s an understandable mistake since “program” can mean “project” in some contexts.
Regarding the Portuguese translation, I did some iterations EN-PT-EN using Google Translate for the complete paragraph.
In the first translation to Portuguese, the sentence ‘small-sized boards’ is translated to ‘placas de pequeno porte’ (small-sized plates). There are different Portuguese words for ‘board’ that depend on the context and, in this case, ‘boards’ should have been translated to ‘tabuleiros’ instead of ‘placas’; the translation would be correct if the text was about circuit boards (in fact, it is the only scenario I can think of). Curiously, Google has selected this very specific case. Maybe it is due to my recent search history for Arduino boards…
The second iteration translated ‘small boards’ (‘-sized’ was gone) to ‘pequenas pranchas’, which would be correct if Go was played on surfboards. The third iteration translated ‘little boards’ to ‘pranchas’ – the adjective suddenly disappears. In the fourth iteration and on, ‘boards’ is constantly translated to the Portuguese word ‘placas’, ending in a repeating cycle for the translation of the full paragraph.
This method (of translating in another language and then back) may result in a meaningful and grammatically correct text in English, but it does not guarantee that all parts of the text have the intended meaning in the foreign language, as shown in the example above.
I think translation algorithms can be improved by identifying other words in the text that can hint at the context (e.g. Go, game, losing, winning), and then use this information to pick the proper translated word.
why is the blog writing of ml when its forte is complexity and knows nothing about ml?