🔍Human Expertise Meets AI: The Quality Assurance Process
Next, we move to the Human Labeling phase. If the sentences generated by the AI models were perfect, this human quality assurance step might be unnecessary. However, for now, it's essential to meticulously review the content for semantic suitability and accurate reflection of the Korean context. Therefore, we commissioned an external specialized institution to conduct reviews by experts in both Korean and English.
For testing purposes, we also ran an interesting experiment: we had ChatGPT review Gemini’s results, and Gemini review ChatGPT’s results.
First, we presented the five sentences generated by ChatGPT and asked Gemini to review them. Our specific request was to "correct any sentences that do not fit the image or contain inappropriate expressions."
Conversely, we asked ChatGPT to review the sentences generated by Gemini, making the exact same request for an accurate comparison.
Here is how Gemini corrected some of the sentences originally generated by ChatGPT:
<Sentences Generated by ChatGPT, Corrected by Gemini (English)>
The corrections mainly focused on some awkward phrasing, and the resulting expressions certainly feel more natural.
Similarly, ChatGPT also reviewed and corrected the sentences generated by Gemini.
<Sentences Generated by Gemini, Corrected by ChatGPT (English)>
In this case, ChatGPT tended to make the sentences more concise. It also revised parts that felt overly emotional or verbose, suggesting that the model perceived the original phrasing as somewhat elaborate and sentimental.
In our actual process, the review and correction work was handled by human experts, not AI. However, given this level of performance, it seems the day is not far off when AI can also take over quality assurance tasks.
Following this, we finalized the quality verification through a Data Validity Evaluation. We were able to achieve excellent validity scores: 83% (English) and 62% (Korean).
🌐 Conclusion: Our Journey Toward Sovereign AI
As we've emphasized, Sovereign AI can only hold true meaning if it accurately reflects the unique context and culture of the nation and society it serves. The Korean Video Understanding Data Project we've introduced clearly illustrates this direction.
The 41,000 Korean-centric images and 205,000 captioning sentences we constructed are more than just a dataset; they are a foundation that captures cultural nuances and regional context. This work is a core basis for realizing what Sovereign AI fundamentally aims for: securing data sovereignty, achieving technological self-reliance, and respecting cultural diversity.
Moving forward, we are committed to ensuring that various global cultures are naturally integrated into AI, contributing to the establishment of unbiased and well-balanced AI models.
<Until the day AI models can seamlessly generate images of characteristic parks from around the world>
Thank you for reading today.
We'll be back soon with more insightful stories!
