In the Mouth of the Wolf: A Taste of How AI Consumes Eggs and Idioms

Machine translation is tricky. Honestly, computers in general are tricky because they are quite literal while we humans are quite the opposite. Humans have a proclivity (an inclination or preference towards a particular thing) to blend actions, sounds, and words; essentially, to cut corners. In English, we regularly use “can’t”, “don’t”, “I’ll” and more, to save mere seconds of speech as opposed to saying, “cannot”, “do not”, “I will”. In English the word “bear” is pronounced as a single syllable rather than “bee-ahr”. Thanks to our habit of blending everything into their smallest parts, we get a single simple sound.

This extends even further to our use of idioms. If you hear the phrase “cutting corners” you probably correctly interpret the meaning as, “to undertake something in what appears to be the easiest, quickest, or cheapest way, especially by omitting to do something important or ignoring rules.” This is exactly why we use idioms, to explain something more succinctly. But in order to appropriately use a “stand-in” phrase, you must have the linguistic and cultural context that accompanies the use of the idiom. n machine translation, cutting corners is not an option, but in fact, it's the opposite.

Not Egg-actly Easy

When a computer is given a command, it performs the required action exactly as it was meant to do, in exactly the way it was told, unlike you, a human. If you were asked to make a cake, the recipe may instruct you to mix two large eggs with the other ingredients. What is not explicitly stated in those directions is that you have to crack the eggs, pour the contents into the bowl, discard the white shell, and mix the gooey inside into the rest of the ingredients. In our logic, we don’t see the need to explain this process. There is an assumption that you have the cultural context for understanding what “mix two large eggs” means.  We see the action of mixing eggs synonymous with the multi-step process of cracking the eggs, extracting the contents, and mixing them as one thing. In other words, we have concatenated (to link (things) together in a chain or series)these together into a seamless action.  

What happens if you ask a computer to perform the same action of “mixing two large eggs”? Say we had a robotic arm capable of doing all the work we would need. Then the computer gets to the same step, adding the eggs to the cake mix. To your surprise, if it was a simple program, you may see the robotic arm add two whole eggs, shell and all, to the mix. Why? Because the step said “Add two large eggs”. Even that would be impressive to me. You may disagree, as this was a failure to properly execute the directions, but you’d be missing out on the intricacy of what just happened.

What the computer still managed to do were the following steps as the computer would see it:

1. Locate 2 eggs.

2. Move arm to first egg.

3. Grab first egg.

4. Move arm to cake mix.

5. Release egg into cake mix

6. Move arm to second egg.

7. Grab second egg.

8. Move arm to cake mix.

9. Release egg into cake mix.

This 9-step process is what we humans do without thinking, on top of cracking the eggs. The incredible work it would take to get the computer to do this is a laborious (difficult and long) process. If that seems unlikely, this display of “dexterous manipulation” was recently achieved by Nayang Technological University in Singapore where they successfully programmed a robotic arm to assemble an IKEA chair. It is possible, just difficult.

And the connection to translation is….?

Now take that same step-by-step logic, and apply it to translation. Imagine having to use a computer’s linear style of logic to break a sentence apart. For example, take the English sentence “Bill is in the lead.” How would a computer translate that?

1. Break the sentence into smaller parts: “Bill, is, in, (the) lead.”

2. Analyze each part:
   “Bill”- A person’s name short for William, or a note for payment of money.
   “Is”- A simple verb to mean the condition, third person singular.
   “In”- Preposition, locates the attached word in a position in space or time.
   “Lead”- A toxic metal, or position of leadership, to lead. Partnered with “the”, signifying a noun.  

3. Analyze the pieces grammatically:
   “In the lead”- Meaning to  having the leading position or higher score in a contest. Functions as an adjective.
   “Bill is”- Lacks “the” or “a” in front of “Bill”, confirming that “Bill” is a name. “Bill is” signifies that “Bill” has the properties of “in the lead”

4. Generate meaning- Bill, a person, is in first place.

This is an oversimplification, but the logic is the same. The linear process must start at the base level of the sentence. If you are fluent in English you automatically assume that “Bill” is a person, not a note for payment of money. If you are a native speaker, or have native level proficiency in English, you also automatically assume “lead”, in this context, means the position of being first, rather than the toxic metal. Remember  computers or robotic arms speak different languages (i.e. Python, C/C++, Java, R) depending on the type of task it has been programmed to accomplish. This example is a glimpse into the effort that machine translation involves.

“A Tough Nut to Crack”

This is why you might have seen idioms break down to embarrassing mistranslations in programs such as Google Translate. The Italian “In bocca al lupo” translates to “in the mouth of the wolf” but is used as a means of bidding someone, “Good luck”.  Like our example with eggs, a machine can miss the context, and be too literal. Although you have the fascinating result of the computer translating, “In the mouth of the wolf”, you are missing the true meaning. We want to wish someone good luck, just as we want a cake that does not have any eggshells in it.

We cannot stress enough that, although machine translation is becoming a useful tool around the world, it is reliant on the human component. Have you ever noticed that when you use Google Translate, there is a button on the side that says, “Suggest an edit”? About a month or two ago, typing “Good luck” into Google Translate yielded the result, “Buona fortuna”, in Italian which is not a mistranslation, but certainly not the common phrase. Recently, if you try to translate “good luck” into Italian, it will now say, “In bocca al lupo.” Machine translation needs human guidance in order to flourish, and that’s where Tone comes in.

We use machine translation tools all the time, but they are far from perfect, and they are simply tools. Human translators have the cultural experience and understanding to make accurate translations. Tone’s human translators are nothing to shake a stick at. Our team is in the know, they are on top of their game, and they are without a doubt, the coolest. Drop the last couple sentences of this paragraph in a translation program, and you will understand exactly what we mean, but we’re willing to bet that the translation program won’t.