My background is delivery driving, not NLP. My dad has 40+ years in seismic interpretation and when I first looked at this competition I kept thinking — sparse data, formulaic patterns, unknown tokens you have to infer from context — this is just a well log problem with different labels on it.\nSo that's basically the direction we went. We started with a Self-Organizing Map on the Akkadian token space. The idea is that if you cluster the vocabulary spatially before you do any translation, unknown tokens can inherit meaning from their neighbors in the map. A token you've never seen that lands in the commerce cluster is probably a trade word, even if it never appeared in training.\nFrom there we layered in rule-based translation — CAD lexicon, onomasticon, logogram anchors, Old Assyrian epistolary formulas. These tablets are repetitive enough that pattern matching gets you pretty far on its own.\nThen ByT5 came in last as a post-editor. We trained it on the SOM's outputs rather than raw Akkadian, so it was learning to fix our specific broken English instead of trying to do translation from scratch on 1,500 pairs.\nIt's two of us — my dad on the domain knowledge, me on the code. .I have no CS background and I vibe-code everything through AI tools, so there are probably decisions in here that would make an NLP person's eye twitch. But it was a genuinely interesting problem.\nCurious what other people did with this one.\n— Alvaro & Alvaro Sr
", "messages": [ { "id": 3423654, "postDate": "2026-03-18T20:32:32.890Z", "content": "My background is delivery driving, not NLP. My dad has 40+ years in seismic interpretation and when I first looked at this competition I kept thinking — sparse data, formulaic patterns, unknown tokens you have to infer from context — this is just a well log problem with different labels on it.\nSo that's basically the direction we went. We started with a Self-Organizing Map on the Akkadian token space. The idea is that if you cluster the vocabulary spatially before you do any translation, unknown tokens can inherit meaning from their neighbors in the map. A token you've never seen that lands in the commerce cluster is probably a trade word, even if it never appeared in training.\nFrom there we layered in rule-based translation — CAD lexicon, onomasticon, logogram anchors, Old Assyrian epistolary formulas. These tablets are repetitive enough that pattern matching gets you pretty far on its own.\nThen ByT5 came in last as a post-editor. We trained it on the SOM's outputs rather than raw Akkadian, so it was learning to fix our specific broken English instead of trying to do translation from scratch on 1,500 pairs.\nIt's two of us — my dad on the domain knowledge, me on the code. .I have no CS background and I vibe-code everything through AI tools, so there are probably decisions in here that would make an NLP person's eye twitch. But it was a genuinely interesting problem.\nCurious what other people did with this one.\n— Alvaro & Alvaro Sr
", "rawMarkdown": "\nMy background is delivery driving, not NLP. My dad has 40+ years in seismic interpretation and when I first looked at this competition I kept thinking — sparse data, formulaic patterns, unknown tokens you have to infer from context — this is just a well log problem with different labels on it.\nSo that's basically the direction we went. We started with a Self-Organizing Map on the Akkadian token space. The idea is that if you cluster the vocabulary spatially before you do any translation, unknown tokens can inherit meaning from their neighbors in the map. A token you've never seen that lands in the commerce cluster is probably a trade word, even if it never appeared in training.\nFrom there we layered in rule-based translation — CAD lexicon, onomasticon, logogram anchors, Old Assyrian epistolary formulas. These tablets are repetitive enough that pattern matching gets you pretty far on its own.\nThen ByT5 came in last as a post-editor. We trained it on the SOM's outputs rather than raw Akkadian, so it was learning to fix our specific broken English instead of trying to do translation from scratch on 1,500 pairs.\nIt's two of us — my dad on the domain knowledge, me on the code. .I have no CS background and I vibe-code everything through AI tools, so there are probably decisions in here that would make an NLP person's eye twitch. But it was a genuinely interesting problem.\n\nCurious what other people did with this one.\n— Alvaro & Alvaro Sr", "votes": 3 } ], "comments": [], "raw_markdown_by_id": { "3423654": "\nMy background is delivery driving, not NLP. My dad has 40+ years in seismic interpretation and when I first looked at this competition I kept thinking — sparse data, formulaic patterns, unknown tokens you have to infer from context — this is just a well log problem with different labels on it.\nSo that's basically the direction we went. We started with a Self-Organizing Map on the Akkadian token space. The idea is that if you cluster the vocabulary spatially before you do any translation, unknown tokens can inherit meaning from their neighbors in the map. A token you've never seen that lands in the commerce cluster is probably a trade word, even if it never appeared in training.\nFrom there we layered in rule-based translation — CAD lexicon, onomasticon, logogram anchors, Old Assyrian epistolary formulas. These tablets are repetitive enough that pattern matching gets you pretty far on its own.\nThen ByT5 came in last as a post-editor. We trained it on the SOM's outputs rather than raw Akkadian, so it was learning to fix our specific broken English instead of trying to do translation from scratch on 1,500 pairs.\nIt's two of us — my dad on the domain knowledge, me on the code. .I have no CS background and I vibe-code everything through AI tools, so there are probably decisions in here that would make an NLP person's eye twitch. But it was a genuinely interesting problem.\n\nCurious what other people did with this one.\n— Alvaro & Alvaro Sr" } }