There is no such thing as “one-size-fits all” with Machine Translation

In my last article, I went over the growth of Machine Translation (MT) in today’s world. today’s world. Now that you have a retrospective on the origins of MT and the latest trends in the field, let’s see if using MT is a good choice for your needs.

In this article, you will soon realize that the phrase “one size fits all” doesn’t apply to Machine Translation:

  • Some content can only be translated by a human
  • Some can be processed by MT but require post-editing
  • Some can be processed through MT with no human intervention.

Let’s start with content for which, we “the humans”, are indispensable.

A good example is advertising.

Designers work hand in hand with copywriters to craft messages which only a human can create. Let’s take the example of this ad in which an airline company is positioning itself as a dynamic airline covering all continents.

 

Using a free translation MT tool based on neural MT, the French rendition of the tagline above reads:

Allez voyagez.

There are several issues in this MT output:

The machine is not clever enough to recognize that in French, two verbs in the imperative form cannot follow each other, and the result is a tagline with grammar issues, which also sounds dismissive to any French reader (it sounds as “Go take a trip!”)

MT runs into similar issues for marketing copy and newsletters, which require human creativity to draw our attention.

Another category which cannot be processed through MT, is legal text and contracts.

One mistranslated word is likely to lead to a lawsuit. Not worth trying to save time and money!

Readers, let’s not despair!

There are several categories of business content which can make a very good use of MT, for specific domains, but need our intervention.

When humans are involved in machine translated output, they are no longer reviewers or editors, they are post-editors.

User guides, manuals or help content usually require less post-editing than websites, due to the fact that they are more consistent and repetitive. Let’s take a look of some help content.

Let’s take a look of some help content.

Here’s the original help content posted on a website:

To delete a conversation:

Click the Messaging icon at the top of your homepage.

Select the message you want to delete from the list on the left rail.

Click the More icon on the top of your conversation thread and select Delete conversation.

Click Yes, delete to confirm.

Here’s the MT raw (not post-edited) output in French:

“Pour supprimer une conversation:

Cliquez sur l’icône Messagerie en haut de votre page d’accueil.

Sélectionnez le message que vous voulez supprimer de la liste sur le rail de gauche.

Cliquez sur l’icône Plus en haut de votre discussion et sélectionnez Supprimer la conversation.

Cliquez sur Oui, supprimez pour confirmer.”

As a post-editorI only had to change two errors:

In the raw output , the word “rail” was translated as a “track” designed for a train.  Within 3 minutes, I followed the steps on that site to visualize what “left track” meant and corrected the meaning. I also had to insert a space before the colon, to follow French punctuation rules.

“Pour supprimer une conversation :

Cliquez sur l’icône Messagerie en haut de votre page d’accueil.

Sélectionnez le message que vous voulez supprimer de la liste sur la barre de gauche.

Cliquez sur l’icône Plus en haut de votre discussion et sélectionnez Supprimer la conversation. Cliquez sur Oui, supprimez pour confirmer.”

That translation is ready to roll!

 

Aside from Help, user guides and manuals, MT can also be used in business specific domains.

Let’s take the example of the Finance Industry.

Interestingly enough, a system has been developed just for the translation of financial documents. This MT platform called VERTO was developed by Lingua Custodia.

It relies on a pre-trained engine which contains corpora and terminology for the financial domain. Financial firms often rely on it for translation which require a fast turn-around on highly specialized documents.

Giant financial-related corpora and terminology databases need to be fed in the machine before it can be used with a minimal margin of error. Very light post-editing is needed, but sometimes professionals using a highly customized and reliable version of VERTO can leverage raw output without any post-editing.

You may now wonder, can we use MT without any post-editing? Is it safe enough?

In my first article about MT, I did an experiment with a highly sophisticated product description.

Although it has three minor flaws, the translation processed through a Neural MT system was impressive. Today, many e-commerce sites rely strictly on NMT for the millions of products sold on their site.

Yet, e-commerce is not the only industry for which NMT has become indispensable.

User-generated content (UGC) such as blogs, chats, instant messaging can only rely on NMT for the benefit of millions of customers around the world.

What determines the raw output quality rating depends on what the user needs. In the field of UGC, the goal is to allow users to be able to understand the writer’s general intentbehind the message, blog or review.

Let’s take the example of a tourist visiting Paris for the first time.

Ben is American and has been warned about tourist traps in the French capital. He also happens to be a great cook and is looking for an authentic restaurant, recommended by French customers.

 

Ben is an avid traveler and mainly uses this particular review site because it can translate reviews left in several languages.

And here comes the magic of NMT:

Everything is perfect just perfect! The ambiance, service & food. Oh la! A true delight. French gastronomy at its best really Hat! It’s worth the detour. The gourmet cellar, my favorite Parisian 2017!

In this case, Ben gets the message, good food, good service and good vibe. He may wonder what “it’s really best Hat” means but with such an enthusiastic review, why should he care?

Ben’s choice is made and he only spent 5 minutes to make his decision. Enjoy your trip Ben!

I’ll use the same review site when I travel to Japan because I like Japanese food even if I don’t read the language.

At Spartan, we facilitate your MT content decisions and strategies!

Contact me (sabine@spartansoftwareinc.com) at Spartan Software, Inc. for a consultation.