Karnataka govt. writes to Meta on ‘incorrect auto-translations'

The Hindu

a day ago

Chief Minister Siddaramaiah's media advisor has written to tech giant Meta flagging faulty auto-translation of Kannada content on Meta platforms like Facebook and Instagram.
'We have noted with concern that the auto-translation from Kannada to English is frequently inaccurate, and, in some cases, grossly misleading. This poses a significant risk, especially when public communications, official statements, or important messages from the Chief Minister and the government are incorrectly translated,' read the letter, which further noted that it can lead to misinterpretation among users.
Terming flawed translation as 'unacceptable', the letter, written by K.V. Prabhakar, media advisor to the Chief Minister, urged Meta to temporarily suspend the auto-translation feature for Kannada content until the translation accuracy is reliably improved. It has also called on the platform to collaborate with qualified Kannada language experts and linguistic professionals to enhance the quality and contextual accuracy of translations between Kannada and English.
'Social media platforms must act responsibly. I caution citizens to be aware that translations shown are often inaccurate. Such negligence by tech giants can harm public understanding and trust,' Mr. Siddaramaiah wrote on X.
Training important
While the State government's letter comes closely on the heels of the mistranslation of a Facebook post by the Chief Minister on the demise of actress B. Saroja Devi, faulty auto-translation of regional Indian languages to English has been a persisting concern.
Srinath Srinivasa, who heads the Web Science lab and is the Dean (R&D) at IIIT Bangalore, speaking to The Hindu, said that the availability of good quality and large enough corpuses to train language models is vital for translation.
'Most of the corpuses are based on literary - not colloquial - versions of the language. This is called diglossia. In Natural Language Processing community, languages such as Kannada are known as low resource languages - not because there is less literature in the language, but there is less trainable literature. It is a major problem in most Indian languages,' he noted.
Nirlendu Saha, who works as an AI Engineer with a global MNC, said the accuracy of translation varies highly across languages, with Spanish and French faring better than Asian languages like Kannada, Tamil, Mandarin, and Korean.
'While artificial intelligence has made significant strides in simplifying auto-translations, it's crucial to acknowledge its limitations at this point. Until we have more comprehensive training data, we can't expect perfection. Models need to be trained in the digital literature of Kannada,' he said.
Unreliable platforms?
Meanwhile, experts also expressed concern about government officials using platforms, which are prone to error, for official communication.
Chandan Gowda, professor and dean of the School of Liberal Arts at Vidyashilp University, noted that while the government is well within its right to complain about being mistranslated, there remains the question of why it is relying on a translation instrument it is unsure of.
'The large official machinery of the country should take safeguards in communicating to the public, like using images of messages, which don't run the risk of being poorly translated, instead of text messages. The Meta platform ought to have been responsible and cautioned the government about the risks of distortion in its translation and offered the option of turning it off,' he said.
The Hindu reached out to Meta, but there was no response at the time of publication.