In the previous topic, we discussed the currently highly regarded
neural network machine translation.
Even free translation services like Google Translate and Microsoft Translator have adopted the mechanisms of neural networks, making them a hot topic.

By the way, everyone, you must be curious about how much the accuracy of translations improves by adopting the mechanism of neural networks, right??

In this topic, we will introduce the results of comparing the output accuracy of traditional statistical-based machine translation engines and
neural network machine translation engines.

 

●Accuracy Comparison Results

This is the Japanese text for machine translation.

——————————————–
This command applies to the display content after the command is executed,
If an error message is displayed, it will apply to that specific error message.
——————————————–

Let's compare the results of translating this Japanese text into English using a company's "statistical-based MT engine" and
"neural network MT engine".
(Both use the same company's MT engine.)

The sentence is long and at first glance seems unsuitable for machine translation,
but what kind of results did it produce?

・Statistical-based MT engine machine translation results

——————————————–
This command applies to the contents after running the command,
an error message is displayed if it applies to the error message.
——————————————–

・Neural Network MT Engine Machine Translation Results

——————————————–
This command applies to the displayed content
after the command execution, so if you receive an error message,
it will be applied to the error message.
——————————————–

How about comparing the accuracy?

 

●Discussion of Accuracy Comparison Results

In the machine translation results of the statistics-based MT engine,
“the contents of the after running the command”,
“if the applies to the error message” have
grammatical issues, resulting in sentences that do not make sense.

On the other hand, the machine translation results from the neural network MT engine may have some unnatural parts,
but compared to statistical MT engines, the grammar is correct,
and you can see that the meaning of the Japanese is conveyed.

In this way, the MT engine that adopts the mechanism of neural networks has,
compared to statistical-based MT engines,
significantly improved output accuracy!!

 

●Why are neural networks amazing?

Many traditional MT engines segmented text into individual words,
translated each word, and combined them
using a mechanism called "phrase-based".

On the other hand, the new engine that adopts the mechanism of neural networks does not simply replace words with their translations,
but rather,
takes into account the context based on the connections with other words in the same sentence,
and can determine how each word should be translated.

With this system, it has become possible to find translation candidates with higher accuracy than the conventional version,
and,
resulting in a quality that is closer to human translation.

If such high quality can be achieved,
it seems that the introduction of machine translation will increasingly progress even in Japan.

We encourage everyone to consider introducing machine translation for items that require shortened delivery times and cost reductions.
At Human Science, we also introduce machine translation engines that can be used in a secure environment
with neural networks.

Please feel free to contact us!!

 

>>Contact Form

If the form is not available, please send your inquiry via email to hsweb_inquiry@science.co.jp.
Thank you.

Alternatively, please feel free to contact us by phone at TEL: 03-5321-3111.

 

Related Articles

Market Trends: Utilization of Neural Networks in Machine Translation

http://www.science.co.jp/mt/market/blog42.html