Wednesday, February 21, 2024

121 Voices, 1 Nation: India’s AI Mission for Linguistic Diversity


Earlier this year, people in villages in the southwest Indian state of Karnataka spoke sentences in their local Kannada language into a special app. This was part of a project to create India’s first AI-based chatbot for Tuberculosis.

This year, people in villages in a place called Karnataka in India helped make a smart computer program called Bhashini. They used a special phone app to say many sentences in their language called Kannada. The goal was to create a computer friend (chatbot) that can help with Tuberculosis, a sickness.

India has more than 40 million people who speak Kannada, and it’s an important language in the country. But when it comes to computer smarts that understand language, Kannada and many other languages are left out. This means lots of people, hundreds of millions in India, can’t get helpful information or good chances for jobs and money.

So, making a computer friend that speaks Kannada for Tuberculosis is a big step to fix this problem. It means more people can understand important stuff, and everyone gets a fair shot at opportunities.

“For AI tools to work for everyone, they need to also cater to people who don’t speak English or French or Spanish,” said Kalika Bali, principal researcher at Microsoft Research India.

“But if we had to collect as much data in Indian languages as went into a large language model like GPT, we’d be waiting another 10 years. So what we can do is create layers on top of generative AI models such as ChatGPT or Llama,” Bali told the Thomson Reuters Foundation.

In Karnataka, villagers are part of a big group of people from different parts of India who are talking into a device for a tech company called Karya. This company is making sets of information for big companies like Microsoft and Google. They use this information in smart computer programs (AI models) for things like teaching, healthcare, and other services.

The Indian government wants to provide more services using computers, so they’re also making sets of information with a system called Bhashini. This is a smart system that helps translate languages, and it’s creating open source sets of information in local languages for making smart tools.

Bhashini involves people helping out by sharing sentences in different languages, checking if what others said is right, translating texts, and labeling images. A lot of Indians, tens of thousands, have joined in and contributed to Bhashini.

According to Pushpak Bhattacharyya, who leads a lab in Mumbai, the government is really working hard to make sets of information to teach big computer systems Indian languages. These sets are already being used in tools that translate languages for education, tourism, and in the courts. However, there are challenges, like many Indian languages being mostly spoken, not written, and there isn’t a lot of electronic information. Also, collecting information in less common languages is tough and needs special effort.

Out of the many languages spoken around the world, only a small bunch, less than 100, get the spotlight in big language computer programs. The super-smart talking computers we have, like ChatGPT, are mainly taught to understand and generate text in English. Other popular ones, like Google’s Bard and Amazon’s Alexa, are mostly focused on English too. They can understand a bit of other languages like Arabic, Hindi, and Japanese, but it’s quite limited.

People are trying to fix this gap. There’s a group called Masakhane that’s working to make language research better for African languages. In the United Arab Emirates, they’ve made a big language computer program called Jais that can do cool things using Arabic.

In places like India, where there are lots of different languages, a smart way to get information is by asking regular people to help. This is called crowdsourcing. Bali, who got noticed as one of the big names in AI, says that in a country like India, getting people to share how they talk and what they say is a good way to make computers understand more languages.

“Crowdsourcing also helps to capture linguistic, cultural and socio-economic nuances”

said Bali.

“But there has to be awareness of gender, ethnic and socio-economic bias, and it has to be done ethically, by educating the workers, paying them, and making a specific effort to collect smaller languages,” she said. “Otherwise it doesn’t scale.”

As artificial intelligence (AI) continues to grow rapidly, there’s a need for understanding languages that are not widely known, according to Safiya Husain, co-founder of Karya. This demand comes not only from technological advancements but also from academics who want to preserve less common languages.

Karya collaborates with non-profit organizations to find workers living below the poverty line, with an annual income of less than $325. The company pays these workers around $5 per hour to generate data, which is well above the minimum wage in India. Importantly, the workers also own a share of the data they create, giving them the chance to earn royalties. Karya envisions using this data to develop AI products that benefit the community, particularly in areas like healthcare and farming.

Husain points out the significant economic value in speech data, noting that the cost of one hour of Odia speech data, a language spoken in the eastern Odisha state, has increased from about $3-$4 to $40. This shift highlights the growing recognition of the importance of diverse language data in the field of AI.

In India, where less than 11% of the 1.4 billion people speak English, many folks are not very comfortable with reading and writing. This has led to the development of several artificial intelligence (AI) models that focus on speech and speech recognition.

One such project is called Vaani, which means “voice” and is supported by Google. Vaani is gathering speech data from around 1 million Indians and sharing it freely for use in automatic speech recognition and speech-to-speech translation.

The EkStep Foundation, based in Bengaluru, has created AI-based translation tools. These tools are being used at the Supreme Court in India and Bangladesh, helping with language translation.

The AI4Bharat center, backed by the government, has introduced Jugalbandi, an AI-based chatbot. This chatbot is designed to answer questions about welfare schemes in various Indian languages, making information more accessible to people.

The chatbot, named Jugalbandi after a musical duet, combines language models from AI4Bharat and reasoning models from Microsoft. It’s accessible on WhatsApp, a widely used platform in India with around 500 million users. This bot, created to work like a musical duet, where two musicians play off each other, is helping break language barriers and connect with people at the grassroots level.

Gram Vaani, a social enterprise focused on working with farmers, is also utilizing AI-based chatbots to answer questions related to welfare benefits. Shubhmoy Kumar Garg, a product lead at Gram Vaani, emphasizes how automatic speech recognition technologies are making a positive impact by overcoming language barriers and reaching out to communities that need assistance the most.

For individuals like Swarnalata Nayak in Raghurajpur district, Odisha, the increasing demand for speech data in her native language, Odia, has provided a valuable additional income through her work with Karya. She shares that she does the work during her free time at night and appreciates the opportunity to contribute to her family’s well-being by talking on the phone.

Read more

Recommended For You