Technology has long been the force behind modern advancements today – the phones we use, to the television sets we binge our favourite shows on to the appliances in the kitchen to whip up meals or the car we drive… there is very little that remains untouched by technology today.

India is a hugely exciting market for driving technology to every corner of the country – the country has the second highest number of smartphone users at 375 million and this number is expected to cross 600 million this year – and encourages blue-sky thinking in scientific minds and innovators. Another untapped, yet highly lucrative market is hinging on the crucial element of vernacular languages. Less than 10% of India speaks and transacts in English, and this leaves a wide section of the Indian population, comfortable in conversing in native languages, open to digital integration and economic inclusion.

At Microsoft India Development Center and Microsoft Research India, a team of highly passionate and committed engineers, social scientists, researchers, NLP and ML experts, algorithm analysts are working on a suite of technologies that will help integrate this largely untapped section of the population with the Internet and all the benefits it has to offer. Specifically, NLP for Indian languages has become a focal point for Microsoft, and the teams are keen on expanding Microsoft language technologies to as many Indian languages as possible.

Kalika Bali, principal researcher at MSR India & Niranjan Nayak, principal engineering manager at Microsoft India Development Center spent some time highlighting the immense opportunities that lie ahead of them in the Indian language domain, and touched upon their team’s journey developing products and solutions for the Indian native language user.

(Kalika Bali, Principal Researcher, MSR India)

Nayak says, “Currently, NLP R&D in India is focused on expanding Microsoft’s language technology to as many Indian languages as we can. To achieve this, our focus is on machine translation and speech technology such as Text-to-Speech synthesis and Automatic Speech Recognition.”

Broadly, within NLP technology, the team works on providing Indian language models that are exposed as Azure Cognitive Service APIs to developers, building on top of the Azure training infrastructure. Using these APIs further strengthens their own portfolio like MS Teams, MS Word, Bing etc. “In India, these are already available in some local languages like Hindi, while some others are in the process of getting this requisite support,” added Nayak.

With the Indian language landscape changing rapidly, Bali admits that technologies within this space run the risk of becoming obsolete quite rapidly. But a redeeming feature is that the local language ecosystem is more robust now, with increasing involvement from the government, startups, and research institutions; providing a much-needed impetus for local language technologies to scale. While Hindi is seeing a steady rise in adoption, other regional languages are catching up – and this allows innovation hubs like Microsoft Research to continue working on their solution suite to bring more languages into the mainstream. “We still need to do more in terms of qualitative and quantitative assessment of data, and scaling the availability of data for all languages. Efforts need to be aligned across organisations to create usable Indian language datasets,” added Bali.

On promising NLP applications like voice chatbots and sentiment analysis, Nayak says, “Chatbots are definitely becoming more human-like. They will become an important part of multiple enterprise functions.”

(Niranjan Nayak, principal engineering manager, Microsoft India Development Center)

Sentiment analysis, on the other hand, requires specific kind of labelled data to witness the scale of use like chatbots. Regardless, it is a much sought-after functionality for enterprises to better understand their customers and help them extract maximum value from transactional conversations and exchanges. “For sentiment analysis to be truly effective, a lot of data has to be collected in a certain manner. Also, existing technology models for identifying sentiment are mainly available for English and certain other European languages. These models have to be trained afresh with datasets in local languages, to be able to analyse Indian language specific terms and usage,” explains Bali.

Today, Microsoft’s Turing Multilingual Language Model (T-ULRv2) is the latest cross-lingual innovation from Microsoft that can represent 94 languages in the same vector space. T-ULRv2 is a transformer architecture with 24 layers and 1,024 hidden states. The architecture also includes a total of 550 million parameters. This kind of approach would allow for the trained model to be fine-tuned in one language and applied to a different one and helps overcome the challenge of requiring huge amount of training data for every language. The Turing models converge all language innovation across Microsoft and are then trained at large scale to support products like Bing, Office, Dynamics, and Azure Cognitive Services, powering a wide range of interactions via natural language such as chatbot, recommendation, question answering, search, personal assist, customer support automation, content generation, and others.

These capabilities can be used extensively to augment the nature of human-to-human, and human-to-machine interactions in sectors like healthcare, education, financial services and overall customer service, adds Bali. “In Microsoft Research, we’re doing a survey on language technologies available in the healthcare domain across the world. This will help us understand the reach and utility of the existing range of solutions and how to expand them.”

The team is also working on an interactive neural machine translation tool for social impact – this is a human-aided translation tool that allows people who are translating from one language to another to do so with a machine translation at the back end providing help through suggestions. This helps the human translators to come up with improved translations efficiently and develop parallel data for training. This INMT tool is now available in open source; both as a web version with support for training models and a mobile version for resource constrained environments.

Nayak says, “From a technology standpoint, we’re fully capable – we have a highly skilled and capable team rolling out some of the finest products in the market today. NLP presents a sea of endless opportunities; and the real impact of these innovations can be seen when implementation is just as creative, and can challenge the status quo.”