UPDATED 12:00 EDT / OCTOBER 23 2018

AI

Baidu creates the world’s first simultaneous translation system

Baidu Inc.’s advances in artificial intelligence often go unheralded next to the likes of Google LLC and Microsoft Corp. But the company is actually one of the leaders in the field, having developed its own specialized chip for AI workloads, an operating system for AI, cancer-detecting algorithms and even a no-code platform that lets unskilled workers create their own AI models.

Now, it’s looking for a bit more recognition of its efforts. Today it revealed that it has created what it said is the world’s first simultaneous machine translation system with “controllable latency,” which works by using machine learning to anticipate speakers’ next words.

Baidu said it built the Simultaneous Translation with Anticipation and Controllable Latency, or STACL, system in order to make simultaneous translation more widely available. The system could be useful because most translators can only do what’s known as “consecutive interpretation,” where they wait until a speaker has paused, usually at a sentence boundary. The problem with consecutive translation is that it takes quite a long time, and can sometimes disrupt the flow of conversation between the people speaking.

That’s why most government leaders and other high profile individuals prefer to use translators who can perform “simultaneous translation,” in which they begin translating just a few seconds into the speaker’s speech. The problem with simultaneous translation, however, is that very few translators are capable of doing it.

And those that can generally can only do so for 20 to 30 minutes, after which they begin to tire and make too many errors. As a result, speakers often require a team of two of three translators who can alternate in order to keep up with the conversation. “Therefore, there is a critical need to develop automated systems to expand the access to simultaneous translation,” Baidu said in a blog post announcing STACL.

Although translation tools such as Google Translate are becoming increasingly accurate, simultaneous translation is a whole different ballgame. The biggest challenge is the word order differences between languages, which makes it very difficult for such systems to keep up.

“For example, in a Chinese sentence Bùshí Zǒngtǒng zài Mòsīkē yǔ Pǔjīng huìwù, which means “President Bush meets with Putin in Moscow,” the Chinese verb huìwù (“meet”), appears at the very end, similar to a German or Japanese verb. In the English translation, however, the verb “meets” appears much earlier,” Baidu’s blog post explained.

In order to translate that sentence simultaneously and accurately, a translation system would need to predict that the speaker intends to use the word “meet,” otherwise he would be too far behind. “As a result, virtually all commercial “real-time” translation systems still today use conventional full-sentence (i.e., non-simultaneous) translation methods, causing the undesirable latency of at least one sentence, rendering the user out of sync with the speaker,” Baidu said.

But Baidu has managed to reduce that latency to less than a second with its STACL system, which uses machine learning to anticipate what the speaker’s next words will be, in much the same fashion the best human interpreters can do.

The system also allows users to set their desired latency – for example, they can specify a latency of just one word behind the speaker, or a longer latency of, say, five words behind. This is useful too, because translating between some languages, such as French and Spanish, is far easier since the grammar structure and sentence order is usually very similar. So in that instance, a low latency would work just fine. For vastly different languages such as Chinese and English, however, a longer latency of up to five words would be more accurate because it involves less guessing of what the next words will be.

Baidu’s STACL system is still a work in progress, and the company noted that it has its limitations and is not intended to replace human translators altogether. Baidu also admits it has no idea if the system is actually better, or worse, than human translators.

“We have not yet compared our results to professional simultaneous interpreters because there is no such data that we are aware of,” said Liang Huang, principal scientist of Baidu’s Silicon Valley AI Lab. “It would be very interesting to collect such data and conduct a comparison between our system and human interpreters. Simultaneous interpretation is an extremely challenging and mentally exhaustive task, and it is well-known that the best simultaneous interpreters can only cover about 60 percent of the source material.”

The idea, then, is not really to try and compete with humans, but to make simultaneous translation more accessible in situations when a human translator isn’t available.

“The technology opens doors to many applications such as simultaneous translations for international conferences, closed captioning, business meetings, press conferences, and legal proceedings,” Liang said. “This technology has already been integrated into Baidu’s internal speech-to-text translation product. It will be used at Baidu World Tech Conference in November, where all speeches will be simultaneously translated by this technology.”

Analyst Holger Mueller of Constellation Research Inc. told SiliconANGLE that Baidu’s system is welcome, because even though language barriers have been slowly eroded by modern technology, there are still limits. “This is especially true when it comes to Asian languages, so it’s good to a see a local infrastructure-as-a-service provider getting into the game,” he added.

Image: Baidu

A message from John Furrier, co-founder of SiliconANGLE:

Your vote of support is important to us and it helps us keep the content FREE.

One click below supports our mission to provide free, deep, and relevant content.  

Join our community on YouTube

Join the community that includes more than 15,000 #CubeAlumni experts, including Amazon.com CEO Andy Jassy, Dell Technologies founder and CEO Michael Dell, Intel CEO Pat Gelsinger, and many more luminaries and experts.

“TheCUBE is an important partner to the industry. You guys really are a part of our events and we really appreciate you coming and I know people appreciate the content you create as well” – Andy Jassy

THANK YOU