Hello there, new to Kaggle here (this would be my first competition to which I'm actually submitting)\nCould someone explain to me why a procedural approach scores so bad? Why can't I just translate using a dictionary of words/phrases? Why should I use a language model?
\nAnother question would be regarding ChatGPT (5.2) performance on translating Akkadian to English. Seemed really good when I tried it on a few pieces of text from train.csv and I really doubt that any model within the Kaggle Notebook memory range would perform better. Why don't we use it instead?
", "messages": [ { "id": "3382225", "postDate": "12/26/2025 23:41:20", "content": "Hello there, new to Kaggle here (this would be my first competition to which I'm actually submitting)\nCould someone explain to me why a procedural approach scores so bad? Why can't I just translate using a dictionary of words/phrases? Why should I use a language model?
\nAnother question would be regarding ChatGPT (5.2) performance on translating Akkadian to English. Seemed really good when I tried it on a few pieces of text from train.csv and I really doubt that any model within the Kaggle Notebook memory range would perform better. Why don't we use it instead?
", "rawMarkdown": "Hello there, new to Kaggle here (this would be my first competition to which I'm actually submitting)\nCould someone explain to me why a procedural approach scores so bad? Why can't I just translate using a dictionary of words/phrases? Why should I use a language model?\n\nAnother question would be regarding ChatGPT (5.2) performance on translating Akkadian to English. Seemed really good when I tried it on a few pieces of text from train.csv and I really doubt that any model within the Kaggle Notebook memory range would perform better. Why don't we use it instead?", "votes": null }, { "id": "3382358", "postDate": "12/27/2025 10:16:27", "content": "Hello and congrats on your first submissions!
\n\n\nWhy can't I just translate using a dictionary of words/phrases?
\n
My guess would be that this is due to the difference in how Akkadian and English build words -- Akkadian relies heavily on word roots, prefixes, suffixes and infixes, much more than English does. What in English would be 3 words, e.g., \"in the house\" is a single word in Akkadian: \"bētānu\". However, \"bētānu\" can also mean \"inside/interior\" (of a palace, mountain/valley), or describing movement as \"inwardly/inwards\". (Here's the entry on the assyrianlanguages.com). So I'm guessing that you would have to account for a lot of N-grams to capture the context properly and there aren't that many texts to begin with.
\n\n\nWhy should I use a language model?
\n
It captures the context much easier. Additionally you can build on existing similar languages, e.g., Aramaic, Arabic or even Akkadian. I saw similar projects start from multilanguage models, e.g., NLLB-200 from Meta or Google's UMT5-small. It enables one to start from some information about the language, even if it's not perfect.
\n\n\nAnother question would be regarding ChatGPT (5.2) performance on translating Akkadian to English. Seemed really good when I tried it on a few pieces of text from train.csv and I really doubt that any model within the Kaggle Notebook memory range would perform better. Why don't we use it instead?
\n
The Akkademia project I cited here (they created the AKK-300m language model for Akkadian) is super small compared to ChatGPT 5.2. As far as I know people don't have access to the model itself (it's only available via API), so it cannot be used as is (Internet OFF, model not publicly available). What ou could try is using a distilled GPT5.2 model. I found a list of Qwen models on HuggingFace, though haven't vetted them myself for safety, efficiency, etc.
\nGood luck with the Competition!
", "rawMarkdown": "Hello and congrats on your first submissions!\n\n> Why can't I just translate using a dictionary of words/phrases?\n\nMy guess would be that this is due to the difference in how Akkadian and English build words -- Akkadian relies heavily on word roots, prefixes, suffixes and infixes, much more than English does. What in English would be 3 words, e.g., \"in the house\" is a single word in Akkadian: \"bētānu\". However, \"bētānu\" can also mean \"inside/interior\" (of a palace, mountain/valley), or describing movement as \"inwardly/inwards\". ([Here's the entry on the assyrianlanguages.com](https://www.assyrianlanguages.org/akkadian/dosearch.php?searchkey=6519&language=id)). So I'm guessing that you would have to account for a lot of N-grams to capture the context properly and there aren't that many texts to begin with.\n\n\n> Why should I use a language model?\n\nIt captures the context much easier. Additionally you can build on existing similar languages, e.g., Aramaic, Arabic or even [Akkadian](https://huggingface.co/Thalesian/AKK_300m). I saw similar projects start from multilanguage models, e.g., [NLLB-200 from Meta](https://cointegrated.medium.com/how-to-fine-tune-a-nllb-200-model-for-translating-a-new-language-a37fc706b865) or [Google's UMT5-small](https://huggingface.co/Thalesian/AKK_300m). It enables one to start from some information about the language, even if it's not perfect.\n\n\n> Another question would be regarding ChatGPT (5.2) performance on translating Akkadian to English. Seemed really good when I tried it on a few pieces of text from train.csv and I really doubt that any model within the Kaggle Notebook memory range would perform better. Why don't we use it instead?\n\n[The Akkademia project I cited here](https://github.com/gaigutherz/Akkademia) (they created the [AKK-300m language model for Akkadian](https://huggingface.co/Thalesian/AKK_300m)) is super small compared to ChatGPT 5.2. As far as I know people don't have access to the model itself (it's only available via API), so it cannot be used as is (Internet OFF, model not publicly available). What ou could try is using a distilled GPT5.2 model. I found [a list of Qwen models on HuggingFace](https://huggingface.co/collections/TeichAI/gpt-52), though haven't vetted them myself for safety, efficiency, etc.\n\n\nGood luck with the Competition!", "votes": null } ], "comments": [ { "id": 3382358, "author_name": "angantyr", "author_url": "", "post_date": "12/27/2025 10:16:27", "content": "Hello and congrats on your first submissions!
\n\n\nWhy can't I just translate using a dictionary of words/phrases?
\n
My guess would be that this is due to the difference in how Akkadian and English build words -- Akkadian relies heavily on word roots, prefixes, suffixes and infixes, much more than English does. What in English would be 3 words, e.g., \"in the house\" is a single word in Akkadian: \"bētānu\". However, \"bētānu\" can also mean \"inside/interior\" (of a palace, mountain/valley), or describing movement as \"inwardly/inwards\". (Here's the entry on the assyrianlanguages.com). So I'm guessing that you would have to account for a lot of N-grams to capture the context properly and there aren't that many texts to begin with.
\n\n\nWhy should I use a language model?
\n
It captures the context much easier. Additionally you can build on existing similar languages, e.g., Aramaic, Arabic or even Akkadian. I saw similar projects start from multilanguage models, e.g., NLLB-200 from Meta or Google's UMT5-small. It enables one to start from some information about the language, even if it's not perfect.
\n\n\nAnother question would be regarding ChatGPT (5.2) performance on translating Akkadian to English. Seemed really good when I tried it on a few pieces of text from train.csv and I really doubt that any model within the Kaggle Notebook memory range would perform better. Why don't we use it instead?
\n
The Akkademia project I cited here (they created the AKK-300m language model for Akkadian) is super small compared to ChatGPT 5.2. As far as I know people don't have access to the model itself (it's only available via API), so it cannot be used as is (Internet OFF, model not publicly available). What ou could try is using a distilled GPT5.2 model. I found a list of Qwen models on HuggingFace, though haven't vetted them myself for safety, efficiency, etc.
\nGood luck with the Competition!
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "3382225": "Hello there, new to Kaggle here (this would be my first competition to which I'm actually submitting)\nCould someone explain to me why a procedural approach scores so bad? Why can't I just translate using a dictionary of words/phrases? Why should I use a language model?\n\nAnother question would be regarding ChatGPT (5.2) performance on translating Akkadian to English. Seemed really good when I tried it on a few pieces of text from train.csv and I really doubt that any model within the Kaggle Notebook memory range would perform better. Why don't we use it instead?", "3382358": "Hello and congrats on your first submissions!\n\n> Why can't I just translate using a dictionary of words/phrases?\n\nMy guess would be that this is due to the difference in how Akkadian and English build words -- Akkadian relies heavily on word roots, prefixes, suffixes and infixes, much more than English does. What in English would be 3 words, e.g., \"in the house\" is a single word in Akkadian: \"bētānu\". However, \"bētānu\" can also mean \"inside/interior\" (of a palace, mountain/valley), or describing movement as \"inwardly/inwards\". ([Here's the entry on the assyrianlanguages.com](https://www.assyrianlanguages.org/akkadian/dosearch.php?searchkey=6519&language=id)). So I'm guessing that you would have to account for a lot of N-grams to capture the context properly and there aren't that many texts to begin with.\n\n\n> Why should I use a language model?\n\nIt captures the context much easier. Additionally you can build on existing similar languages, e.g., Aramaic, Arabic or even [Akkadian](https://huggingface.co/Thalesian/AKK_300m). I saw similar projects start from multilanguage models, e.g., [NLLB-200 from Meta](https://cointegrated.medium.com/how-to-fine-tune-a-nllb-200-model-for-translating-a-new-language-a37fc706b865) or [Google's UMT5-small](https://huggingface.co/Thalesian/AKK_300m). It enables one to start from some information about the language, even if it's not perfect.\n\n\n> Another question would be regarding ChatGPT (5.2) performance on translating Akkadian to English. Seemed really good when I tried it on a few pieces of text from train.csv and I really doubt that any model within the Kaggle Notebook memory range would perform better. Why don't we use it instead?\n\n[The Akkademia project I cited here](https://github.com/gaigutherz/Akkademia) (they created the [AKK-300m language model for Akkadian](https://huggingface.co/Thalesian/AKK_300m)) is super small compared to ChatGPT 5.2. As far as I know people don't have access to the model itself (it's only available via API), so it cannot be used as is (Internet OFF, model not publicly available). What ou could try is using a distilled GPT5.2 model. I found [a list of Qwen models on HuggingFace](https://huggingface.co/collections/TeichAI/gpt-52), though haven't vetted them myself for safety, efficiency, etc.\n\n\nGood luck with the Competition!" }, "source": "meta" }