Custom Search
 



 
 
Google leverages neural networks to sharpen instant translation between 9 Indian languages

After latest additions last week, to its Translate app,  phone  users can carry out  a bilingual conversation  with on-the fly translation
Bangalore, September 20 2017: Users  who have been trying out Google's translation tool on its website for some years, might have noticed a sharp improvement in recent months. The classic technique where  translation was being done phrase by phrase -- with some weird syntax resulting -- has given way to  more mature systems driven by neural networks  -- computers that mimic the  neurons of the human brain. 
Neural machines translate  full sentences at a time  and the  computer system uses the  broader context to help it figure out the most relevant translation. It then rearranges and adjusts the translation so that it is  conversational, like a human speaking with proper grammar. The use of neural machines,  which Google described in a paper last year,  has drastically improved translation accuracy   and  the technology is being applied progressively to all the 103 language   that the Google Translate app supports.
Last week,  the  impact was felt on the 9 Indian languages that the app embraces: Hindi, Bengali, Marathi, Tamil, Telugu, Gujarati, Punjabi, Malayalam and Kannada.  Google researchers in India harnessed the similarity between groups of Indian languages  -- Marathi and Hindi or Malayalam and Tamil  for example.  The neural machine  "learns"  multiple Indian languages  simultaneously,  and when  it trains on them  together, the translations for all  are better,  than if the machine was trained on  each individually.  Strength lies in numbers!
Perhaps the most useful aspect of the Google Translate updates is  Conversation Mode.  Users can have a bilingual conversation with someone,  by talking to the Google Translate app. To activate this,  the user needs to  tap the microphone to start speaking in a selected language, then tap the mike again: Google Translate  will automatically recognize which of the two languages are being spoken, letting you have a  smooth conversation.  
Recognizing that many  phone owners may not have always-on Internet, Google has also  introduced an offline mode in some of the Indian languages  --  Hindi, Bengali, Gujarati, Kannada, Marathi, Tamil, Telugu, and Urdu --   where  they can obtain a translation of a word or phrase even when there is no Internet connection. For this one has to download the specific language pack.  To translate between two languages, one has to download both packs.
A feature previously available in Google Translate  was  the ability to use  the phone-camera mode to click a  photo of English text and get a translation.   Now this feature too, has been  fuelled by machine learning. Using something called Word Lens,    one can obtain instant translation in real time -- very useful if you want to read  a street sign or a restaurant menu that you can't understand.
India is an important test bed for Google's translation technology:  it ranks among the world's top four countries using the Translate tool and the number of users has doubled in six months since the Indian languages  were introduced.   Interestingly, more Indians like to speak  their translation requirement rather than entering text. In fact voice based search  in Hindi quadruples  year on year.
The Google Translate app is available both for Android and iOS.




    


Post Your Comments Now
       
  Name  
  Email  
  Comment