More than 19,000 dialects are spoken in India, creating difficulties for an AI bot to explain somewhat taboo sexual health topics. The creators of the Myna Bolo project have found a solution.
Category
🗞
NewsTranscript
00:00 The main Abolo bot can answer tricky questions about sex and health in India
00:04 But it has to understand cultural nuances and different dialects
00:08 Not only is India one of the most linguistically diverse areas with you know hundreds of languages
00:13 But within those languages there are so many sub languages sub dialects slangs
00:18 Not only slangs within those languages, but slangs that are different within those families
00:23 So how do we then produce answers in a chat bot that is meant to target women that are maybe not literate to the standardised
00:29 version of our languages
00:31 Advanced large language models enable the chat bot to comprehend and engage in multiple languages
00:37 The LLMs used by the main Abolo app are trained to connect with users in their native languages like Marathi and even
00:44 Hinglish Hindi mixed with English words. While people don't think sex they do say sambandh karna
00:50 But maybe a mom says sambandh karna or a daughter says a relation rakhna
00:55 So there are different ways of saying the same thing so creating that nuance then using slang so things that are replaced
01:03 Little words that could be replaced by a metaphor that is actually used instead so having those metaphors compiled
01:09 Lots of back-end database is needed so the bot can understand a prompt correctly and doesn't get lost in translation
01:17 or even worse create confusion with incorrect information or
01:22 hallucinations
01:24 This requires intensive groundwork
01:27 We check the answers that the chatbot provides and if it is wrong then we have a correct version that we create
01:35 So then we take all the correct questions and the correct answers and we feed it back into the model and that is actually
01:41 20 to 25 thousand questions that we are fact-checking and vetting not only with our staff
01:48 But also with medical professionals and doctors and health professionals who've been in the field for 10 years
01:53 Hallucinations is the AI issue that almost across industries
01:56 We are trying to figure out how to decrease it and along with other things like, you know
02:01 Data bias, gender bias, medical bias that comes up in the data because it is in the world
02:06 They're the reason why there's gender bias in the data because there's gender bias in the world