OpenAI and Government of Iceland Partnership for Icelandic Language Preservation

OpenAI has partnered with the Government of Iceland, the non-profit language technology center Almannarómur, and the language technology company Miðeind ehf to integrate GPT-4 into efforts to preserve the Icelandic language. This collaboration aims to prevent the digital extinction of Icelandic by ensuring the language remains functional and accurate within the AI applications and software that citizens use daily.

Using RLHF to Solve Low-Resource Language Challenges

Reinforcement Learning from Human Feedback (RLHF) is being used to improve GPT-4's proficiency in Icelandic, overcoming the limitations of traditional fine-tuning for low-resource languages. While GPT-4 is inherently more capable than previous versions, it still initially produced grammatical errors and incorrect cultural knowledge in Icelandic.

To address this, a team of 40 volunteers led by Miðeind ehf implemented an RLHF process where human testers:

  1. Provide a prompt to GPT-4.
  2. Review four generated completions.
  3. Select the best response and edit it to create an ideal completion.

This feedback loop is then used to further train the model. This method is significantly more efficient than previous attempts to fine-tune GPT-3, which failed to produce grammatical Icelandic even after 300,000 examples. In contrast, RLHF can produce meaningful results with as few as 100 examples, making it a viable path for other "long-tail" or low-resource languages to avoid an "AI divide."

Improvements in Linguistic and Cultural Accuracy

GPT-4 demonstrates a marked improvement over GPT-3 and GPT-3.5 in handling Icelandic, specifically regarding cultural nuances and correct translations.

Translation Accuracy

One example of this improvement is the translation of "Donald Duck." Where GPT-3 and ChatGPT (GPT-3.5) provided incorrect answers, GPT-4 correctly identified the Icelandic name as "Andrés Önd."

Cultural and Contextual Awareness

GPT-4 exhibits different responses based on the language of the prompt, allowing it to provide localized context:

  • Icelandic Prompt: When asked about the number of parliamentarians, GPT-4 correctly identified the Alþingi (Icelandic parliament) as having 63 members.
  • English Prompt: When asked the same question in English, the model provided a general list of parliamentarians in various countries (UK, Canada, India, Australia).
  • Presidential Context: When asked in Icelandic who the president is, the model correctly identified Guðni Th. Jóhannesson (as of 2021).

Practical Applications and Future Goals

The ultimate goal of the partnership is to enable Icelandic companies to deploy GPT-4 in their interactive applications without defaulting to English. This ensures that the native tongue remains the default language for digital interaction.

Key planned applications include:

  • Voice Assistants: Embla, a voice assistant app by Miðeind, will use GPT-4 as a backend to provide fluent Icelandic conversations and translations.
  • Business Integration: Icelandic companies can implement native Icelandic chatbots on their websites rather than relying on English-speaking alternatives.

Despite these advancements, the team continues to work on refining the model to eliminate remaining grammatical and spelling errors, ensuring the model can power complex and creative applications in Icelandic.

Sources