Universal Translator For All Languages In The World - Meta Is Already On!

By Author: John
Total Articles: 1
Comment this article

The social media group Meta (Facebook) is working on its Metaversum , which is to include virtual work and leisure worlds in the future. According to Meta, this also includes a universal translator who masters all languages in the world, no matter how rare, and translates them into other languages in real time.
Our ability to communicate is the basis of our humanity, writes Meta in this report: "Teaching AI to translate 100s of spoken and written languages in real time". Technologies have changed our way of communicating and exchanging ideas time and time again. And the power of technologies would be even greater if they could work the same way for billions of people around the world - providing similar access to information and making it possible to communicate with much larger audiences, regardless , in which language. As Meta strives for a more inclusive and connected world, it will be even more important to break down the existing barriers to information and opportunity by supporting people in their respective languages like polish translation services:
With a universal translator that includes the languages ...
... of all people in the world, the digital and physical world can be opened up in ways that have not been possible before, Meta continues.
Accordingly, removing language barriers is crucial and would enable billions of people to access information online in their mother tongue or their preferred language. Advances would not only help those who don't speak any of the languages that dominate the Internet today, but would also fundamentally change the way people around the world connect and exchange ideas.
There is still no universal translator. But our professional translators and interpreters are proficient in all the languages and dialects of the world.
Meta is already working on this and making strides in enabling translations for low-resource languages, which would pose a significant barrier to universal translation for most of the world's population. By getting work in areas like
• corpus creation,
• multilingual modeling
• and evaluation
advance and disclose, it is hoped that other researchers could build on this work and bring the real world uses of translation systems closer to reality.
Meta explains that for people who understand world languages like English, Mandarin or Spanish, it may seem like today's applications and web tools already provide the translation technology we need. But billions of people are still left out - they can't easily access most information on the web or connect to most of the online world in their native language, according to Meta.
While current machine translation (MT) systems are improving rapidly, they still rely heavily on learning from large amounts of text data . Therefore, in general, they would not be well suited to low-resource languages, i.e., languages
• which either lack training data
• or who did not have a standardized writing system.
As a result, today's AI translation systems are not designed to serve the thousands of languages used worldwide or to provide real-time language-to-language translations. In order to truly serve everyone, research must address three key challenges.
1. Overcome data scarcity by collecting more training data in more languages and finding new ways to use the data already available.
2. Overcome modeling by overcoming the challenges that would arise if the models were used for more and more languages.
3. Evaluate and improve results by finding new ways to do it.
Data scarcity is one of the biggest hurdles to expanding translation of documents tools to more languages. Machine translation systems for text translations usually rely on learning from millions of sentences of annotated data. Because of this, MT systems capable of delivering high-quality translations have so far only been developed for the few languages that are dominant on the Internet. Expanding to other languages means finding ways to acquire and use training samples from languages with little web presence.
When translating directly from language to language, data acquisition is even more challenging. Most speech MT systems would use text as an intermediate step: speech in one language is first converted to text, then translated into text in the target language, and finally fed into a text-to-speech system to produce audio. Thus, the speech-to-speech translations are dependent on the text in a way that limits their efficiency and makes it difficult to scale to languages that are mainly spoken orally.
Direct language-to-language translation models could allow translations for languages that would not have standardized writing systems. This language-based approach could also lead to much faster and more efficient translation systems by eliminating the extra steps of converting speech to text, translating, and then generating speech in the target language.
Meta goes on to explain that today's machine translation systems are not designed to meet the needs of everyone around the world - not to mention that they require proper training data in thousands of languages. Many MT systems are bilingual, so each language pair has its own model, such as English-Russian or Japanese-Spanish. This approach is extraordinarily difficult to scale to dozens of language pairs, let alone all the languages used around the world.
According to Meta, many experts would have suggested that multilingual systems could be helpful here. However, it is extremely difficult to integrate many languages into a single efficient, powerful multilingual model that can map all languages.
Real-time language processing faces many of the same challenges as text-based models, according to the meta-report. In addition, the latency, i.e. the delay that occurs when translating from one language to another, must be overcome before they can be used effectively for real-time translations. The greatest challenge therefore arises from the fact that a sentence can be spoken in different languages in different word orders.
According to Meta, even professional simultaneous interpreters lag behind the original language by about three seconds. The report gives this as an example:
Suppose the German sentence "I want to translate all languages" and its Spanish equivalent "Quisiera traducir todos los idiomas" are to be translated into English. Real-time translation from German to English would be more challenging, according to the meta-report, because the verb "translate" is at the end of the sentence, while the word order in Spanish and English is similar.
Finally, as it expands to more and more languages, new methods for evaluating the translations produced by MT models would also need to be developed. According to Meta, there are already resources for assessing the quality of translations, for example from English into Russian, but the question is: what about translations from Amharic into Kazakh?
In order to expand the number of languages that current MT models can translate, new approaches to training data and measurements are also needed to cover more languages.
It's not just about accuracy, but also about carrying out the translations responsibly. Ways would have to be found to ensure that MT systems would preserve cultural sensibilities and not create or reinforce prejudice.
And that is exactly what Meta KI is currently working on. It remains exciting.
Do you need translators or interpreters for rarely spoken languages? No problem! Our language experts are on hand to help you with all the languages and dialects of the world, we promise!