{
  "id": 681543,
  "title": "RAG is Needed?",
  "url": "/competitions/deep-past-initiative-machine-translation/discussion/681543",
  "author_name": "WOOSUNG YOON",
  "post_date": "2026-03-15T18:32:49.484000",
  "votes": 1,
  "comment_count": 0,
  "views": 0,
  "content": "<p>For everyone in the Deep Past Challenge (Akkadian → English):</p>\n<h2>(1) RAG is Needed?</h2>\n<p>Just read a <strong>low-resource translation</strong> paper that feels custom-made for this competition.</p>\n<h3><strong>Title:</strong>  \"Enhancing Low-Resource Minority Language Translation with LLMs and Retrieval-Augmented\"</h3>\n<p><img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F4808143%2F38688511d3d592d494ce94e0adaa0991%2F111.png?generation=1773599229739907&amp;alt=media\" alt=\"rag\"></p>\n<p>They tested 7 models on <strong>Mandarin → Hakka (Sìxiàn dialect)</strong>.</p>\n<ul>\n<li><strong>Authors:</strong> Chen-Chi Chang et al. (National United University, Taiwan)</li>\n<li>Results:<ul>\n<li>Model 4 (Gemini 2.0 + RAG): 31% BLEU ← Best  </li>\n<li>Dictionary + Gemini refinement: 26%  </li>\n<li>Dictionary-only: 24%  </li>\n<li>Pure LLM baselines: 12~18%  </li></ul></li>\n</ul>\n<p>I’ve been quietly following the Deep Past Challenge.<br>\nIf a team has space for one more member,<br>\nI’d be grateful to join and do my best in the final week to add new layers and improve the score.  </p>\n<p>Thank you!</p>",
  "messages": [
    {
      "id": 3421523,
      "postDate": "2026-03-15T18:32:49.483Z",
      "content": "<p>For everyone in the Deep Past Challenge (Akkadian → English):</p>\n<h2>(1) RAG is Needed?</h2>\n<p>Just read a <strong>low-resource translation</strong> paper that feels custom-made for this competition.</p>\n<h3><strong>Title:</strong>  \"Enhancing Low-Resource Minority Language Translation with LLMs and Retrieval-Augmented\"</h3>\n<p><img src=\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F4808143%2F38688511d3d592d494ce94e0adaa0991%2F111.png?generation=1773599229739907&amp;alt=media\" alt=\"rag\"></p>\n<p>They tested 7 models on <strong>Mandarin → Hakka (Sìxiàn dialect)</strong>.</p>\n<ul>\n<li><strong>Authors:</strong> Chen-Chi Chang et al. (National United University, Taiwan)</li>\n<li>Results:<ul>\n<li>Model 4 (Gemini 2.0 + RAG): 31% BLEU ← Best  </li>\n<li>Dictionary + Gemini refinement: 26%  </li>\n<li>Dictionary-only: 24%  </li>\n<li>Pure LLM baselines: 12~18%  </li></ul></li>\n</ul>\n<p>I’ve been quietly following the Deep Past Challenge.<br>\nIf a team has space for one more member,<br>\nI’d be grateful to join and do my best in the final week to add new layers and improve the score.  </p>\n<p>Thank you!</p>",
      "rawMarkdown": "For everyone in the Deep Past Challenge (Akkadian → English):\n\n## (1) RAG is Needed?  \nJust read a **low-resource translation** paper that feels custom-made for this competition.\n\n\n### **Title:**  \"Enhancing Low-Resource Minority Language Translation with LLMs and Retrieval-Augmented\"\n\n\n![rag](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F4808143%2F38688511d3d592d494ce94e0adaa0991%2F111.png?generation=1773599229739907&alt=media)\n\nThey tested 7 models on **Mandarin → Hakka (Sìxiàn dialect)**.\n\n- **Authors:** Chen-Chi Chang et al. (National United University, Taiwan)\n- Results:\n - Model 4 (Gemini 2.0 + RAG): 31% BLEU ← Best  \n - Dictionary + Gemini refinement: 26%  \n - Dictionary-only: 24%  \n - Pure LLM baselines: 12~18%  \n\n\nI’ve been quietly following the Deep Past Challenge.  \nIf a team has space for one more member,   \nI’d be grateful to join and do my best in the final week to add new layers and improve the score.  \n\nThank you!",
      "votes": 1
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "3421523": "For everyone in the Deep Past Challenge (Akkadian → English):\n\n## (1) RAG is Needed?  \nJust read a **low-resource translation** paper that feels custom-made for this competition.\n\n\n### **Title:**  \"Enhancing Low-Resource Minority Language Translation with LLMs and Retrieval-Augmented\"\n\n\n![rag](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F4808143%2F38688511d3d592d494ce94e0adaa0991%2F111.png?generation=1773599229739907&alt=media)\n\nThey tested 7 models on **Mandarin → Hakka (Sìxiàn dialect)**.\n\n- **Authors:** Chen-Chi Chang et al. (National United University, Taiwan)\n- Results:\n - Model 4 (Gemini 2.0 + RAG): 31% BLEU ← Best  \n - Dictionary + Gemini refinement: 26%  \n - Dictionary-only: 24%  \n - Pure LLM baselines: 12~18%  \n\n\nI’ve been quietly following the Deep Past Challenge.  \nIf a team has space for one more member,   \nI’d be grateful to join and do my best in the final week to add new layers and improve the score.  \n\nThank you!"
  }
}