An AI experiment in Spanish translation

I compared results from Deepl, Google, and Bard, and got a surprise

You’re reading the My News Biz newsletter. My goal is to help digital media entrepreneurs find viable business models.

I’ve been writing blog posts in Spanish and English for 14 years. Although I’m fluent in Spanish, I’m not a native speaker, and I’ve always worried a bit about the finer points. I try to write in a conversational style, and I use a lot of metaphors, which don’t always translate well.

“He was out in left field” might work fine in American English, but in Spanish it just sounds weird. So does the Britishism “It was a sticky wicket.”

And in Spanish, if you say “soy una tumba” (literally, I’m a tomb), everyone knows you mean “mum’s the word, I’ll keep your secret.” In Mexican Spanish to say something is “muy padre” (literally, “very father”) means really good. And the cultural history behind all that requires a book to explain.

Image by Gino Crescoli from Pixabay

Sound and sense

And don’t get me started on the minefield of prepositions. So I’ve turned to native Spanish speakers over the years to give a final edit to my Spanish texts and save me embarrassment.

But that time may have passed. The tools of artificial intelligence (AI) have made translation much better. In the past few years, Google Translate’s Spanish tool has noticeably improved.

The reason could be that Google Translate and Deepl, another online translator, can draw on a larger database of documents in both Spanish and English than is available in some other languages by virtue of the number of global users, according to my AI course instructor, Sil Hamilton.

Results aren’t as good in some other languages I’ve used, such as German, French, and Italian. But there’s no question. AI has made translation tools better.

The test: Deepl vs. Google vs. Bard

I gave each of these translation tools a 1,600-word newsletter of mine to translate into Spanish: Notice good work, and praise it.

Neither Claude nor Bard nor ChatGPT seemed capable of comparing the translations either in pairs or together to extract and display all the differences. So I made my own line-by-line comparison, and it took me all day.

I’m used to this kind of editing after years of working professionally in both languages.

Here are some observations I made using Deepl as a base to compare the translations of Google and Bard.

• Google Translate differed from Deepl’s translation into Spanish more than 60 times in a word or expression used. (For example, Deepl translated “pay” as la paga vs. el salario by Google and Bard).

• Bard differed from Deepl’s translation into Spanish 25 times. (It translated “he combined” in the past imperfect verb tense combinaba vs. combinó, the simple past or preterite used by Google and Bard.) One native speaker I consulted argued strongly for “combinaba” while another accepted “combinó”.

• Bard and Google agreed on a translation that was different from Deepl 12 times. (For example, Deepl translated “I did this” correctly as “hice esto” vs. the less precise translation rendered by Google and Bard, “lo hice” or “I did it.” )

If the goal of the translation is to give readers the basic idea of a text, without misleading them, I think all three of these do a good basic job. That was a big surprise for me.

All three of the translations were accurate in their use of Spanish and in conveying the meaning of my English newsletter — with three exceptions. Google made three minor errors having to do with agreement in gender or number of nouns or pronouns. Any experienced editor, or even automated grammar check, would have caught them.

A big translation boo-boo. Recently I was using Google to translate a ballet review from German into English. It translated the word “Unbekümmertheit” as “careless” in describing the choreographer’s adaptation of the traditional “Nutcracker” choreography. But another acceptable translation of the word is “carefree”, which was clearly the intent of the reviewer’s very positive article.

Flavor of Spain

At first, I found myself preferring Bard’s translation of my article. This was totally subjective. It could have been based on years of study of literature. The Spanish felt somehow smoother and more elegant than either Deepl or Google. My preference might have been influenced by the seven years I spent living and working in academia in Spain.

An example. In one place my newsletter talks about how to “deal with” problems. Google chose the Spanish word “afrontar”, or “confront” problems. Deepl used “abordar”, roughly “tackle” or “deal with”. And Bard used “lidiar”, which means to fight or combat. Lidiar is often used literally or figuratively to describe a torero confronting a bull.

Another flavor of Spain in Bard — verbs in the subjunctive mood. Spaniards, in my experience, use the subjunctive more than native speakers from other countries. (Linguists, please correct me if I’m wrong.)

The subjunctive is used in Spanish to express opinions, desires, or emotions. When Deepl could avoid using the subjunctive, it did.

Deepl’s translation of “No newsroom boss should let that happen” was “Ningún jefe de redacción debería permitirlo.” Deepl used an infinitive there (in boldface) rather than the subjunctive. By contrast, Bard and Google used the subjunctive: “Ningún jefe de redacción debería permitir que suceda” (in boldface).

I changed my mind and now prefer Deepl. It seemed less likely to use the subjunctive. The overall feel of the translation was less formal, more conversational, like a blog post, which it is. Deepl captured the tone and style, not just the words.

Conclusion: Two giants

Translation is, in fact, an art. And one of the best translators of Spanish and Portuguese authors, Gregory Rabassa, explained how he often collaborated with them in bringing their works into English (“If This Be Treason: Translation and its Dyscontents, a Memoir”.

Rabassa’s best known translation is of Gabriel García Márquez’s “One Hundred Years of Solitude”.

Another great translator, Edith Grossman, who died in September, was best known for her translation of Cervantes’ Don Quixote. She also translated many contemporary works, including García Márquez’s “Love in the Time of Cholera,” another of my favorites.

She described her craft in the book “Why Translation Matters.” Grossman once said, “there are as many ways to translate a text as there are translators.” Indeed.

Choosing the right word or phrase whose sound and rhythm capture that of the original requires a fantastic ear for the subtleties of language, as well as a knowledge of the historical and cultural context of the work.

AI is helping make routine translation better, but they are a long way from replacing Rabassa or Grossman.