Google Translate fails on simple sentences
Ernie Davis
Note: Google Translate is
here.
Systran is
here.
Bing is
here.
DeepL is
here
This collection should under no circumstances be taken as any kind of
serious benchmark. Not only is it very small, but, as the title indicates,
it was
originally formed in 2016 as a collection where Google Translate, specifically,
failed; hence, those versions of this page were
inherently unfair to Google Translate. In this latest
pass, I have particularly aimed at DeepL, which is the acknowledged leader of
the pack for the languages it covers; hence, it is systematically unfair to
DeepL.
It is just
an unsystematic collection of translation problems that seem like they
should be easy, but trip up Google Translate, DeepL, and the other translation
programs, during the time period I have been doing this.
I check these intermittently to see
what progress has been made, and update the page accordingly. As you
can see by comparing with the earlier versions, the state of the art has
improved steadily, but there are still plenty of "easy" sentences that
trip up the best publicly available systems.
Earlier versions:
September - December 2016
December 2016
May 2017
August 2018
March 2019
June 2018
Note: I don't actually speak French, so I am grateful to Pascal Amsili for
his corrections on a number of items.
Updated 12/27/2019.
Some general comments on this pass:
-
The MT technology continues to advance. Programs now get right quite a
few of the examples that they got wrong six months ago.
-
Mostly, with a little experimentation, I was able to find a variant that
broke them, even DeepL. Not always. Their ability to find long-distance
connections most of the time is really getting pretty impressive.
-
I now have more German examples and fewer French ones than in previous
iterations, because DeepL has gotten really quite hard to beat in French,
especially since I don't actually speak French.
-
There were two examples where a program had deteriorated since last
checking. In June, DeepL correctly translated "My neighbor is a woman"
as "Ma voisine est une femme"; now it gives "Mon voisin est une femme" (#12).
In June, GT translated "Les clés sont à ma mère," as "The keys are to
my mother" which isn't good, but now it gives "The keys are mine"
which is worse (#7).
-
There are still some fairly weird mistranslations. DeepL still translates
"I love you" as "Je t'aime. Je t'aime." and similar doubling in a
bunch of other languages (#15). Bing and DeepL are still translating "Elles"
in French as "Ellos" in Spanish (#5. GT and Bing now get this right.)
Examples
-
En: The plumber who we called to fix the dishwasher doesn't work any longer.
Fr: Le plombier que nous avons appelé pour réparer le lave-vaisselle ne
travaille plus.
GT, Bing, Systran: Le plombier que nous avons appelé pour réparer le
lave-vaisselle ne fonctionne plus.
DeepL: Le plombier qu'on a appelé pour réparer le lave-vaisselle ne travaille
plus. [correct]
The point here is that "works" in English can mean either "labors" or
"functions correctly".
Many other languages have separate words for these.
As of 12/27/2019,
GT, Bing, and Systran use the word for "functions
correctly" in translating into French. The programs have certainly improved
over time in making this distinction (see the previous versions of this page)
but can still be fooled.
GT and Systran make the corresponding
mistake in translating into Spanish, Italian, and German.
Bing gets it wrong in German and Italian, and makes a weird mistake in Spanish:
"El fontanero que llamamos para arreglar el lavaplatos doesn't trabajar más."
DeepL gets it right with German, but wrong with Spanish and Italian.
To my ear, the English original here sounds very slightly odd, but most of the
people I asked said it seemed fine to them. Many thanks for a number of
friends for discussions, particularly Christina Behme.
-
En: The clock stopped working.
Ge: Die Uhr funktioniert nicht mehr.
GT, Systran: Die Uhr hörte auf zu arbeiten.
Bing, DeepL: Die Uhr funktionierte nicht mehr. [Correct]
Hannah Bast approves the translations
"Die Uhr funktioniert nicht mehr" or
"Die Uhr geht nicht mehr". Christina Behme suggests
"Die Uhr ist stehengeblieben". The translations given by GT, Bing, and
Systran are characterized as understandable, but no one would ever say them.
DeepL's translation is in the past tense, where as the translation given by
Hannah Bast is in the present, but either is correct.
-
The clock that we gave to the electrician works.
Fr: L'horloge que nous avons avons donné à
l'électricien fonctionne.
GT: L'horloge que nous avons donnée aux travaux d'électricien.
Bing, DeepL: L'horloge que nous avons avons donné à
l'électricien fonctionne. [Correct]
Systran: L'horloge que nous avons avons donné à
l'électricien travaille.
If you change this to,
The clock that we gave to the busy German electrician works
then DeepL gets it wrong, though Bing still gets it right.
-
Fr: Je ne peux pas nager.
En: I cannot swim. I can't swim.
GT: I can not swim. [A small mistake, but an odd one for so simple a
sentence]
Bing, Systran, DeepL: I can't swim. [Correct]
-
Fr: Elles
Sp: Ellas
GT, Systran. Ellas. [Correct]
Bing, DeepL. Ellos.
-
En: Pierre called to Julia and her sister but they didn't hear him.
Fr: Pierre a appelé Julia et sa soeur mais elles ne l'ont pas entendu.
GT, Bing, Systran, DeepL: Pierre a appelé Julia et sa soeur mais
ils ne l'ont pas entendu.
-
Fr. Les clés sont à ma mère.
En. The keys belong to my mother.
GT: The keys are mine. [A worse result than in 6/2019]
Systran, The keys are to my mother.
Bing: The keys belong to my mother [Correct].
DeepL: The keys are my mother's [Correct].
-
En. Pierre's parents miss him.
Fr. Pierre manque à ses parents.
GT. Les parents de Pierre lui manquent.
Bing, Systran, DeepL: Il manque aux parents de Pierre. [OK]
-
En. The dog is pregnant.
Fr. La chienne est enceinte.
GT, Bing, Systran, DeepL: Le chien est enceinte.
DeepL. La chienne est enceinte. [Correct]
Adapted from a suggestion of Richard Socher.
-
En. Pierre's sister said to Jacques, "I will always be your friend."
Fr: La soeur de Pierre a dit à Jacques: "Je serai toujours ton amie."
GT, Systran, Bing, DeepL:
La soeur de Pierre a dit à Jacques: "Je serai toujours ton ami."
-
En. Pierre chatted with his neighbor and sent regards to her father.
Fr. Pierre a discuté avec sa voisine et a envoyé des
salutations
à son père.
Google: Pierre a discuté avec son voisin et a envoyé ses salutations à son père.
Systran: Pierre discuta avec son voisin et envoya des salutations à son père.
Bing: Pierre bavarda avec son voisin et envoya des salutations à son père.
DeepL: Pierre a bavardé avec sa voisine et a envoyé ses salutations à son père. [Correct]
-
En. My neighbor is a woman.
Fr. Ma voisine est une femme.
GT, DeepL: Mon voisin est une femme. (DeepL got this right in June 2019)
Bing, Systran: Ma voisine est une femme. [Correct].
-
En: I always tell them that they were adopted, so they hit me.
Ger: Ich sage ihnen immer, dass sie adoptiert wurden, also schalgen
sie mich.
GT, DeepL: Ich sage ihnen immer, dass sie adoptiert wurden, also haben sie
mich geschlagen.
Bing: Ich sage ihnen immer, dass sie adoptiert wurden, also haben sie
mich getroffen.
Systran: Ich sage ihnen immer, dass sie adoptiert wurden, also trafen sie
mich. [Correct]
-
Fr. Je découpe un avocat pour le déjeuner.
En. I cut up an avocado for lunch.
GT, DeepL: I'm cutting up an avocado for lunch. [Correct]
Bing: I'm cutting up a lawyer for lunch.
Systran: I cut a lawyer for lunch.
Thanks to Houda Bouamor for the suggestion.
-
Fr: Je t'aime.
DeepL English: I love you. I love you. (It does suggest plain "I love you"
as an alternative.)
Similarly doubled in all of DeepL's other languages.
Very strange. The other translation programs get it right.
-
En: The soup is hot because it contains jalapenos.
Sp: La sopa está picante porque contiene jalapeños.
GT, Systran, Bing, DeepL: La sopa está caliente porque contiene jalapeños.
(i.e. hot in temperature)
(Contributed by Robert Krovetz).
-
En: Your wife says you never buy her flowers.
Ger: Ihre Frau sagt, Sie kaufen ihr nie Blumen. [Correct]
GT: Ihre Frau sagt, Sie kaufen ihre Blumen nie.
(Taking "her" to be the
possessive pronoun rather than the indirect object --
that is, the flower that belong to her.)
Bing: Ihre Frau sagt, dass Sie nie ihre Blumen kaufen.
Systran: Deine Frau sagt, du kaufst nie ihre Blumen.
DeepL: Ihre Frau sagt, Sie kaufen ihr nie Blumen. [Correct]
Contributed by Christina Behme. Originally from a cartoon:
Marriage counsellor: Your wife says you never buy her flowers.
Clueless husband: To be honest, I never knew she sold flowers.
--- from "Sad and Useless Humor".
The Stanford parser
also gives the wrong parse for this sentence.
-
Ger: Johann bekam zwei Beagles im Tierheim, aber sie griffen seine Tochter an,
also brachte er sie zurück.
En: Johann got two beagles at the pound, but they attacked his daughter, so he
brought them back.
GT,
DeepL: Johann got two beagles at the shelter, but they attacked his daughter,
so he brought her back.
Bing: John got two beagles in the shelter, but they attacked his daughter,
so he brought her back.
Systran: Johann got two Beagles in the Tierheim
but they attacked his daughter, so he brought her back.
-
En: The workers at the canning factory wash fish, skin fish, bone fish, and
can fish.
Fr: Les travailleurs de la conserverie lavent les poissons,
dépouillent les poissons, désossent les poissons, et
mettent les poissons en conserve.
GT: Les ouvriers de la conserverie lavent le poisson, le poisson en peau,
le poisson en os et peuvent pêcher.
Bing: Les ouvriers à l'usine de conserve lont le poisson, le poisson de peau,
le poisson d'os, et peuvent pêcher.
Systran: Les ouvriers de l'usine de conserverie lavent les poissons,
les poissons de la peau, les os et peuvent pêcher.
DeepL: Les travailleurs de la conserverie lavent les poissons,
les dépouillent, les désossent et les mettent en conserve. [Correct!]
I'm really impressed that DeepL manages this. However, if you ask it to do
German, it's all at sea:
DeepL: Die Arbeiter in der Konservenfabrik waschen Fisch, Fischhaut,
Fischknochen und können fischen.
The Stanford parser also gets this wrong:
(ROOT
(S
(NP
(NP (DT The) (NNS workers))
(PP (IN at)
(NP (DT the) (NN canning) (NN factory))))
(VP (VBP wash)
(NP
(NP (NN fish))
(, ,)
(NP (NN skin) (NN fish))
(, ,)
(NP (NN bone) (NN fish))
(, ,)
(CC and)
(NP (NN can) (NN fish))))
(. .)))
Machine translation fails from Chinese
found by Yuling Gu, June 20, 2018.