Conversation

Notices

1. Embed this notice
   Upol Ehsan (upol@hci.social)'s status on Monday, 05-Jun-2023 08:57:19 JST Upol Ehsan

   🧵
   1/ English is the predominant language of the Internet. So what's different about ChatGPT squishing non-English languages farther into the margins?

   Key points from this piece, my reflections, and one key recommendation for those in Responsible AI:

   #AI #chatgpt #responsibleAI #academia #llm #mastodon

   https://www.wired.com/story/chatgpt-non-english-languages-ai-revolution/

   • Embed this notice
     Upol Ehsan (upol@hci.social)'s status on Monday, 05-Jun-2023 08:57:17 JST Upol Ehsan

     in reply to

     4/
     🎯 "The chatbots’ utterances are “the average of what’s on the internet,”
     ▶ 🎙️ This is my favorite quote from the article. The "norming" effect of the chatbots often lead to very bland, albeit coherent, responses. Is this how we really want our writing to be? Our linguistic fingerprints chatbot-washed?

     Paul Cantrell repeated this.
   • Embed this notice
     Upol Ehsan (upol@hci.social)'s status on Monday, 05-Jun-2023 08:57:17 JST Upol Ehsan

     in reply to

     5/
     ⚡️For my Responsible AI colleagues: it is irresponsible if we don't think of under-resourced languages as we think of Responsible AI in LLMs. As we advocate for responsible use of technology, it's imperative that we advocate for under-resourced languages in an equitable manner. Otherwise, those not at the table will end up on the menu.

     / 🧵 END

   • Embed this notice
     Upol Ehsan (upol@hci.social)'s status on Monday, 05-Jun-2023 08:57:18 JST Upol Ehsan

     in reply to

     3/
     🎯"Fung’s group is curating Indonesian-language data for training models...But collecting more data is unlikely to be enough, because the reams of English text are so large—and still growing"

     ▶ 🎙️ Yes! You won't be able to "collect your data" out of this problem. The same approach that worked for resourceFUL languages will not work for under-resourced one. We need to adopt radically different paradigms based on an equitable outlook.

   • Embed this notice
     Upol Ehsan (upol@hci.social)'s status on Monday, 05-Jun-2023 08:57:19 JST Upol Ehsan

     in reply to

     2/
     🎯 "The AI systems are good at translating other languages into English, but they struggle with rewriting English into ...those with non-Latin scripts."

     ▶ 🎙️This is key-- often, we hear people touting how good these LLMs are on multilingual outputs; often implicit is the fact they are talking about non-English *Latin-based* languages. Non-Latin scripts are severely under-resourced on the web, and by extension in the training of these LLMs, which is why they do so poorly for languages Bangla.