Humans can't stop Apple Translate

The much-anticipated iOS14 brings a new built-in translation function.

Opening it up, it looks unremarkable at first glance.

The translation application ontology is very simple. After selecting the language for mutual translation at the top, the user can enter the content that needs to be translated through text or voice, and the translated text will be displayed in blue font below.

Apple Translator currently supports 11 common languages including Chinese, English, Japanese, German, and Spanish.

However, when we carefully "played" with Apple's translation software, we found that things are not simple. At the beginning, we found that in addition to daily expressions, it can also easily solve some less common expressions:

For example, after typing "Happy Fat Home Water", the result given by Apple Translate is "Coke".

This aroused our interest, and we dug out some old memes to try, and they were faithfully "restored"——

The answer given by Apple to the question "Blue Skinny Shiitake Mushroom" is "Sad.gif"(“Blue Thin Mushroom” is a meme that went viral on the Internet in 2016. It originated from a self-portrait video taken by a lovelorn boy from Nanning. He expressed “uncomfortable and wanted to cry” in dialect, which was “literally translated” as “Blue Thin Mushroom”).

This is so amazing. You know, the answer to Google Translate is still:

We then tried more Internet buzzwords and found that Apple Translator has its own ideas.

After typing "oliji", Apple Translate said that this word means "cool".

It’s become a sperm! Moreover, Apple Translate has not lost when faced with delicate situations of cross-cultural communication.

The Japanese "Moonlight is So Beautiful" passed through its hands and became "I love you". Japanese netizens expressed their favorable impression of Apple.

Natsume Soseki taught for nothing

However, as the first batch of people who used iOS14 continued to "test" Apple Translate like us, everyone soon discovered that something was wrong with this application.

For example, if you enter "五五开", the English word that appears is "Lu Benwei"...(Lu Benwei is a former Douyu anchor. In a game competition, he was asked what it was like to play against opponents who were much stronger than him. He forcibly answered "50-50", and has since become synonymous with this word in the gaming circle. , but outside the gaming circle, not everyone may know this joke)

It's not a matter of accuracy or inaccuracy, it's really that kind of... that rare kind.

When entering "Top-type washing machine", "Kudo Shinichi" is directly displayed in Japanese, which makes people sweat for those who really need to buy a washing machine in Japan.(Because the Japanese pronunciation of "Kudo Shinichi" sounds very much like a drum washing machine. So many Chinese animation fans will call him that. But if it is placed in a "serious" translation software, is it too casual?).

Although to some extent there is no mistranslation

At this point, the perception of Apple Translator has changed from "it can accurately translate the essence of obscure Chinese memes" to "Is this a bit too casual and will delay the business of people who really want to translate?"

And, at other times, Apple Translator behaves like a kid obsessed with making bad jokes.

It's obviously just an ordinary sentence "how many floors does a bag of rice have to carry", but it is still translated into "feel the pain"(Because the Japanese pronunciation of the latter sounds like how many floors does a bag of rice have to carry in Chinese? You can open Apple Translate and give it a try).

You can also try "Spicy Tiansai"

There are more and more similar overturning moments, and everyone begins to feel that Apple’s concentration of “two dimensions as you wish” is too high.

At this point of "playing", we are really curious, who "taught" Apple bad? In fact, the corpus is "contaminated".

Although Apple has always been tight-lipped about its own technology, and this time it did not explain what model is used in iOS14, but we can refer to Google Translate, Apple’s predecessor.

Google Translate uses Seq2Seq(Sequence to Sequence)Model,Seq2Seq consists of two recurrent neural network models,one for encoding the input sequence and one for decoding the,output sequence.

When inputting the Chinese phrase "knowledge is power", the encoding model labels each word with a vector, where each vector represents the meaning of all the words read so far. After the entire sentence is encoded, the decoder starts generating the corresponding English sentence.

By analyzing a large amount of corpus data, the model can automatically learn the corresponding grammatical rules. That is to say,What the engineer teaches the model, the model learns.Therefore, Apple's engineers may have fed Apple Translate too much parallel online corpus, causing Apple's translation to be "polluted" by Internet terms and unable to recognize the original meaning of the text.

Another possibility for Apple Translate’s mistakes is that Apple Translate introduced the knowledge graph.

Knowledge graph is a concept proposed by Google in 2012. It is essentially a graph-based data structure. In the knowledge graph, each noun(also called entity)They are all nodes, and each node is connected by a logical relationship line. Through this knowledge graph, neural networks can better understand the correlation between contexts.

A schematic diagram of a knowledge graph

Perhaps in the knowledge graph built by Apple Translation, "五五开" is linked to the entity "Lu Benwei", and this entity can be translated as "Lu Benwei". In the same way, "Tumble Washing Machine" may also be linked to the entity "Kudo Shinichi".

Because of the existence of network parallel corpus and knowledge graph, the translation model can easily overturn when faced with independent nouns.For example, "guapi", Apple directly understands it according to the dialect and translates it into "idiot".

"Get started with dialect books"

However, based on our judgment of its principle, one of the solutions for more accurate translation is that when Apple Translate makes an error, we can try to add context to the text to help the model understand it better.

For example, change "melon rind" to "I don't eat melon rind", and change "drum-loading washing machine" to "how much does a drum-type washing machine cost".

These translations from Apple do bring a lot of fun, but when people really need to use them to complete cross-language communication, they can't help but sweat.

Now the question is, do you like this Apple Translator?

Author: Xing Yifan