Updated: December 17, 2020 6:50:21 pm
Google Search will now show results in the relevant Indian language, even if the user types in their language using English or Roman characters. The company says this is aimed at improving search experiences for the bilingual Indian users, who might rely on transliteration for typing search queries, even when they are not the most comfortable with English. The company also announced a new language processing model called Multilingual Representations for Indian Languages or ‘MuRIL’, which is geared towards Indian languages. This is a text-to-text based model for language processing.
“One scenario unique to India is that devices are actually set up in English, and it is so much easier to interact with devices in English than in regional languages. But many Indian users who are very comfortable in their native language, may also be using the device in English,” Anand Rangarajan, Engineering Director – Search and Bangalore Site Lead explained in a call with the media.
Adding that it was difficult to type in Indian languages, which means that users often end up relying on Roman or Latin characters even for queries meant in their local language. “This results in our systems detecting a query in English and showing an English language result. They would have been better served with results in their local language. Over the next months, Google Search will now show results in the relevant Indian languages even when it is typed in English,” he added.
The feature will roll out in five Indian languages first: Hindi, Bangla, Tamil, Telugu and Marathi. Further, Google is extending the ability to toggle for Search results between English and regional languages to four more Indian languages. These are Tamil, Telugu, Bangla and Marathi. The feature was originally launched four years back.
Further, the Google Maps app will now support Indian languages. Nine Indian languages will be supported in the Maps app. Users will be able to to the language settings in the app and search for places, get directions in their preferred language.
Google is also adding the ability to get help with homework help in Hindi, which is a feature available with the Google Lens. According to the company, India is the number one market for the usage of Google Lens.
Currently, students can point to quadratic equations, math problems in Google Lens and get a solution in English. This feature will be extended to Hindi as well. When a maths problem is pointed at the Google Lens, it turns an image of a homework question into a query. Based on the query, it shows step-by-step guides and videos to help explain the problem.
Google’s MuRIL model for Indian languages
Google also showcased a new approach for natural language processing, which was developed in India called the Multilingual Representations for Indian Languages or ‘MuRIL’. The company said this is a powerful new multilingual model that can scale across multiple regional Indian languages, and it also provides support for transliterated text such as when writing Hindi using Roman script. Such features were missing from previous models of its kind. Google says MuRIL is also good at determining the sentiment within a sentence.
“The traditional approach has been to build one model for each language, specifically, so say one model for Hindi and a separate one for Tamil and then separate ones for any other language that we want to cover. So as you can imagine, this is a very labour intensive process, because most of the times we don’t have language-specific data to build one model for each language,” Dr Partha Talukdar, Research Scientist, Google Research India explained in the call.
He added that MuRIL is designed specifically for Indian languages, for the nuanced and specific uses that we see in the Indian context. MuRIL is a single model capable of handling multiple languages, and can transfer training and learnings from one language to another. It currently supports 16 Indian languages, and English as well, so a total of 17.
According to Talukdar, the new model will act as a bridge or a common foundation to map all of the different languages into one common platform or a framework, allowing researchers to transfer knowledge and learnings from one language to another.
Giving an example of how MuRIL is more accurate, Talukdar pointed out that the new model would accurately interpret the text such as “Achha hua account bandh nahi hua.” An earlier model would have interpreted this negatively, but MuRIL correctly identifies this as a positive statement.
Another example given was how MuRIL correctly interprets the phrase ‘Shirdi ke sai baba’ as being about a person, rather than a place. MuRIL is free & Open Source, and available on TensorFlow Hub to download for researchers.
📣 The Indian Express is now on Telegram. Click here to join our channel (@indianexpress) and stay updated with the latest headlines
- The Indian Express website has been rated GREEN for its credibility and trustworthiness by Newsguard, a global service that rates news sources for their journalistic standards.