![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F9193114%2F7b1bd7110aa16ecb725aac18749f71b8%2FScreenshot%202026-01-16%20135852.png?generation=1768552700074005&alt=media\")
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F9193114%2F71425da193757ea99f82f924d23b7d55%2FScreenshot%202026-01-16%20140325.png?generation=1768552718898704&alt=media\")
I am using OA_Lexicon_eBL.csv to post-process and standardize Proper Names (PN) and Geographical Names (GN) in my translations. However, I have noticed a systemic conflict between the dictionary forms and the Ground Truth (GT) in the training set.
\n[1]
\nOARE-ID : ab1021ff-2aff-4a13-86fa-6b32c4cf239e\n---\nDEBUG: Replacing 'Ilī-tūram' with 'Ilī-tūra' (Score: 94.11764705882352)\nDEBUG: Replacing 'Aluwa' with 'Alua' (Score: 88.88888888888889)\n---\nNormalized Transliteration : ma-na KÙ.BABBAR ṣa-ru-pá-am ú ša 3 GÍN sà-hi-ir-tám ku-nu-ki-a ili5-tù-ra-am ú-šé-bi4-lá-kum-ma 17 GÍN KÙ.BABBAR ṣa-ru-pá-am ku-nu-ki-a ší-im ṣú-ba-tí-kà id-na-a ú a-lu-wa na-áš-ú-ni-kum\nTranslation (GT) : I sent Ilī-tūram to you with ½ mina of refined silver and small goods worth 3 shekels under my seal, and Idnaya and Aluwa are bringing you 17 shekels of silver under my seal, the proceeds from your textiles.\nPrediction : I sent Ilī-tūram to you with Ilī-tūram's seal, and Idnaya and Aluwa bring you 17 shekels of refined silver under my seal, the proceeds from your textiles.\nPN-GN Fixed :I sent Ilī-tūra to you with Ilī-tūra's seal, and Idnaya and Alua bring you 17 shekels of refined silver under my seal, the proceeds from your textiles.\n\n
[2]
\nOARE-ID : ab1021ff-2aff-4a13-86fa-6b32c4cf239e\n---\nDEBUG: Replacing 'Enišārum' with 'Enišaru' (Score: 80.0)\n---\nAkkadian :a-ṣé-er e-ni-ša-ri-im a-dí AN.NA-ki-kà ú-ṣú-ur-ša-IŠTAR áš-ta-pá-ar\nClean : I have sent Uṣur-ša-Ištar to Enišārum concerning your tin.\nPrediction : I have sent Uṣur-ša-Ištar to Enišārum concerning your tin.\nPN-GN Fixed :I have sent Uṣur-ša-Ištar to Enišaru concerning your tin.\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F9193114%2Fb82c904da9d82858ad877d466ebd2ba7%2FScreenshot%202026-01-16%20140726.png?generation=1768552841513247&alt=media\")
Even If I used norm as replacement instead of lexeme , it still doesn't match ground truth
\n[3]
\nDEBUG: Replacing 'Ikūn-pīya' with 'Ikūn-pīa' (Score: 94.11764705882352)\nDEBUG: Replacing 'Šu-Anum' with 'Šu-Anu' (Score: 92.3076923076923)\n---\nAkkadian :ma-na KÙ.BABBAR ṣa-ru-pá-am i-ku-pí-a DUMU šu-a-nim i-na ší-im SÍG.HI.A ša ú-ší-na-lam il5-qé\nClean : Ikūn-piya son of Šu-Anum has received 8 minas of refined silver from the proceeds of Ušinalam's wool.\nPrediction : Ikūn-pīya son of Šu-Anum has received 1 mina of refined silver from the proceeds of Ušinalam's wool.\nPN-GN Fixed :Ikūn-pīa son of Šu-Anu has received 1 mina of refined silver from the proceeds of Ušinalam's wool.\n[4]
\nDEBUG: Replacing 'Dan-Aššur' with 'Dān-Aššur' (Score: 88.88888888888889)\nDEBUG: Replacing 'Enna-Suen' with 'Ennam-Suen' (Score: 94.73684210526315)\nDEBUG: Replacing 'Iddin-abum' with 'Iddin-abu' (Score: 94.73684210526315)\nAkkadian :dan-a-šùr sí-ki e-na-sú-in DUMU i-dí-a-bi4-im ú-kà-al-ma a-na ha-nu DUMU i-tur4-DINGIR a-na qá-ta-tim i-dí-šu\nClean : Dan-Aššur detains Enna-Suen son of Iddin-abum and he gave him to Hanu son of Itūr-ilī as guarantor.\nPrediction : Dan-Aššur detains Enna-Suen son of Iddin-abum and he gave him to Ḫanu son of Itūr-ilī as guarantor.\nPN-GN Fixed :Dān-Aššur detains Ennam-Suen son of Iddin-abu and he gave him to Ḫanu son of Itūr-ilī as guarantor.\nI am using OA_Lexicon_eBL.csv to post-process and standardize Proper Names (PN) and Geographical Names (GN) in my translations. However, I have noticed a systemic conflict between the dictionary forms and the Ground Truth (GT) in the training set.
\n[1]
\nOARE-ID : ab1021ff-2aff-4a13-86fa-6b32c4cf239e\n---\nDEBUG: Replacing 'Ilī-tūram' with 'Ilī-tūra' (Score: 94.11764705882352)\nDEBUG: Replacing 'Aluwa' with 'Alua' (Score: 88.88888888888889)\n---\nNormalized Transliteration : ma-na KÙ.BABBAR ṣa-ru-pá-am ú ša 3 GÍN sà-hi-ir-tám ku-nu-ki-a ili5-tù-ra-am ú-šé-bi4-lá-kum-ma 17 GÍN KÙ.BABBAR ṣa-ru-pá-am ku-nu-ki-a ší-im ṣú-ba-tí-kà id-na-a ú a-lu-wa na-áš-ú-ni-kum\nTranslation (GT) : I sent Ilī-tūram to you with ½ mina of refined silver and small goods worth 3 shekels under my seal, and Idnaya and Aluwa are bringing you 17 shekels of silver under my seal, the proceeds from your textiles.\nPrediction : I sent Ilī-tūram to you with Ilī-tūram's seal, and Idnaya and Aluwa bring you 17 shekels of refined silver under my seal, the proceeds from your textiles.\nPN-GN Fixed :I sent Ilī-tūra to you with Ilī-tūra's seal, and Idnaya and Alua bring you 17 shekels of refined silver under my seal, the proceeds from your textiles.\n\n
[2]
\nOARE-ID : ab1021ff-2aff-4a13-86fa-6b32c4cf239e\n---\nDEBUG: Replacing 'Enišārum' with 'Enišaru' (Score: 80.0)\n---\nAkkadian :a-ṣé-er e-ni-ša-ri-im a-dí AN.NA-ki-kà ú-ṣú-ur-ša-IŠTAR áš-ta-pá-ar\nClean : I have sent Uṣur-ša-Ištar to Enišārum concerning your tin.\nPrediction : I have sent Uṣur-ša-Ištar to Enišārum concerning your tin.\nPN-GN Fixed :I have sent Uṣur-ša-Ištar to Enišaru concerning your tin.\n
Even If I used norm as replacement instead of lexeme , it still doesn't match ground truth
\n[3]
\nDEBUG: Replacing 'Ikūn-pīya' with 'Ikūn-pīa' (Score: 94.11764705882352)\nDEBUG: Replacing 'Šu-Anum' with 'Šu-Anu' (Score: 92.3076923076923)\n---\nAkkadian :ma-na KÙ.BABBAR ṣa-ru-pá-am i-ku-pí-a DUMU šu-a-nim i-na ší-im SÍG.HI.A ša ú-ší-na-lam il5-qé\nClean : Ikūn-piya son of Šu-Anum has received 8 minas of refined silver from the proceeds of Ušinalam's wool.\nPrediction : Ikūn-pīya son of Šu-Anum has received 1 mina of refined silver from the proceeds of Ušinalam's wool.\nPN-GN Fixed :Ikūn-pīa son of Šu-Anu has received 1 mina of refined silver from the proceeds of Ušinalam's wool.\n[4]
\nDEBUG: Replacing 'Dan-Aššur' with 'Dān-Aššur' (Score: 88.88888888888889)\nDEBUG: Replacing 'Enna-Suen' with 'Ennam-Suen' (Score: 94.73684210526315)\nDEBUG: Replacing 'Iddin-abum' with 'Iddin-abu' (Score: 94.73684210526315)\nAkkadian :dan-a-šùr sí-ki e-na-sú-in DUMU i-dí-a-bi4-im ú-kà-al-ma a-na ha-nu DUMU i-tur4-DINGIR a-na qá-ta-tim i-dí-šu\nClean : Dan-Aššur detains Enna-Suen son of Iddin-abum and he gave him to Hanu son of Itūr-ilī as guarantor.\nPrediction : Dan-Aššur detains Enna-Suen son of Iddin-abum and he gave him to Ḫanu son of Itūr-ilī as guarantor.\nPN-GN Fixed :Dān-Aššur detains Ennam-Suen son of Iddin-abu and he gave him to Ḫanu son of Itūr-ilī as guarantor.\nIf I understand correctly, for PN/GN the lexemes are exclusively in the Nominative, whereas for norm they retain their case (N, G or D). So technically, for English translations, lexeme should be ok.
\nHave you tried compiling a list of PN/GN from Ground Truth and comparing it to the OA_lexicone_eBL entries? I wonder how wide-spread the problem is.
", "rawMarkdown": "If I understand correctly, for PN/GN the lexemes are exclusively in the Nominative, whereas for norm they retain their case (N, G or D). So technically, for English translations, lexeme should be ok.\n\nHave you tried compiling a list of PN/GN from Ground Truth and comparing it to the OA_lexicone_eBL entries? I wonder how wide-spread the problem is.", "votes": null }, { "id": "3392141", "postDate": "01/16/2026 10:50:22", "content": "The technique I’ve been using is using the soundex algorithm to mine the training set for variations of the same proper name from the lexicon. There is some variation maybe this approach could be used to converge on consensus spellings. To be honest though this type of approach is kind of hacking the evaluation metric you’ll get a better score but you won’t get a better translation model considering the quality of the leaderboard scores and makes me think there’s much lower hanging fruit than optimizing naming conventions
", "rawMarkdown": "The technique I’ve been using is using the soundex algorithm to mine the training set for variations of the same proper name from the lexicon. There is some variation maybe this approach could be used to converge on consensus spellings. To be honest though this type of approach is kind of hacking the evaluation metric you’ll get a better score but you won’t get a better translation model considering the quality of the leaderboard scores and makes me think there’s much lower hanging fruit than optimizing naming conventions", "votes": null }, { "id": "3392148", "postDate": "01/16/2026 10:59:21", "content": "We should get a well-established norm for the names (maybe the onomasticon that @deeppast is working on will shed more light on the topic) but from what see the field of Old Assyrian is relatively new/fragmented into many groups, each using a slightly different norm.
", "rawMarkdown": "We should get a well-established norm for the names (maybe the onomasticon that @deeppast is working on will shed more light on the topic) but from what see the field of Old Assyrian is relatively new/fragmented into many groups, each using a slightly different norm.", "votes": null }, { "id": "3392501", "postDate": "01/17/2026 04:16:44", "content": "Thank you all for your efforts here, these are great observations. I just posted the onomasticon.csv dataset so I'll look forward to hearing if it is useful:\nhttps://www.kaggle.com/datasets/deeppast/old-assyrian-grammars-and-other-resources/data?select=onomasticon.csv\nI was not able to list all the different aliases, but my hope is that by providing the spellings, and cross-referencing the OA_lexicone_eBL entries, that a lot of the many-to-many issues with the names in translations vs. transliteration can be resolvable.
", "rawMarkdown": "Thank you all for your efforts here, these are great observations. I just posted the onomasticon.csv dataset so I'll look forward to hearing if it is useful:\nhttps://www.kaggle.com/datasets/deeppast/old-assyrian-grammars-and-other-resources/data?select=onomasticon.csv\nI was not able to list all the different aliases, but my hope is that by providing the spellings, and cross-referencing the OA_lexicone_eBL entries, that a lot of the many-to-many issues with the names in translations vs. transliteration can be resolvable.", "votes": null } ], "comments": [ { "id": 3392118, "author_name": "angantyr", "author_url": "", "post_date": "01/16/2026 10:05:48", "content": "If I understand correctly, for PN/GN the lexemes are exclusively in the Nominative, whereas for norm they retain their case (N, G or D). So technically, for English translations, lexeme should be ok.
\nHave you tried compiling a list of PN/GN from Ground Truth and comparing it to the OA_lexicone_eBL entries? I wonder how wide-spread the problem is.
", "votes": null, "replies": [ { "id": 3392141, "author_name": "aaronbornstein", "author_url": "", "post_date": "01/16/2026 10:50:22", "content": "The technique I’ve been using is using the soundex algorithm to mine the training set for variations of the same proper name from the lexicon. There is some variation maybe this approach could be used to converge on consensus spellings. To be honest though this type of approach is kind of hacking the evaluation metric you’ll get a better score but you won’t get a better translation model considering the quality of the leaderboard scores and makes me think there’s much lower hanging fruit than optimizing naming conventions
", "votes": null, "replies": [ { "id": 3392148, "author_name": "angantyr", "author_url": "", "post_date": "01/16/2026 10:59:21", "content": "We should get a well-established norm for the names (maybe the onomasticon that @deeppast is working on will shed more light on the topic) but from what see the field of Old Assyrian is relatively new/fragmented into many groups, each using a slightly different norm.
", "votes": null, "replies": [ { "id": 3392501, "author_name": "deeppast", "author_url": "", "post_date": "01/17/2026 04:16:44", "content": "Thank you all for your efforts here, these are great observations. I just posted the onomasticon.csv dataset so I'll look forward to hearing if it is useful:\nhttps://www.kaggle.com/datasets/deeppast/old-assyrian-grammars-and-other-resources/data?select=onomasticon.csv\nI was not able to list all the different aliases, but my hope is that by providing the spellings, and cross-referencing the OA_lexicone_eBL entries, that a lot of the many-to-many issues with the names in translations vs. transliteration can be resolvable.
", "votes": null, "replies": [] } ] } ] } ] } ], "raw_markdown_by_id": { "3392072": "I am using OA_Lexicon_eBL.csv to post-process and standardize Proper Names (PN) and Geographical Names (GN) in my translations. However, I have noticed a systemic conflict between the dictionary forms and the Ground Truth (GT) in the training set.\n\n[1]\n```\nOARE-ID : ab1021ff-2aff-4a13-86fa-6b32c4cf239e\n---\nDEBUG: Replacing 'Ilī-tūram' with 'Ilī-tūra' (Score: 94.11764705882352)\nDEBUG: Replacing 'Aluwa' with 'Alua' (Score: 88.88888888888889)\n---\nNormalized Transliteration : ma-na KÙ.BABBAR ṣa-ru-pá-am ú ša 3 GÍN sà-hi-ir-tám ku-nu-ki-a ili5-tù-ra-am ú-šé-bi4-lá-kum-ma 17 GÍN KÙ.BABBAR ṣa-ru-pá-am ku-nu-ki-a ší-im ṣú-ba-tí-kà id-na-a ú a-lu-wa na-áš-ú-ni-kum\nTranslation (GT) : I sent Ilī-tūram to you with ½ mina of refined silver and small goods worth 3 shekels under my seal, and Idnaya and Aluwa are bringing you 17 shekels of silver under my seal, the proceeds from your textiles.\nPrediction : I sent Ilī-tūram to you with Ilī-tūram's seal, and Idnaya and Aluwa bring you 17 shekels of refined silver under my seal, the proceeds from your textiles.\nPN-GN Fixed :I sent Ilī-tūra to you with Ilī-tūra's seal, and Idnaya and Alua bring you 17 shekels of refined silver under my seal, the proceeds from your textiles.\n```\n\n\n\n\n[2]\n\n```\nOARE-ID : ab1021ff-2aff-4a13-86fa-6b32c4cf239e\n---\nDEBUG: Replacing 'Enišārum' with 'Enišaru' (Score: 80.0)\n---\nAkkadian :a-ṣé-er e-ni-ša-ri-im a-dí AN.NA-ki-kà ú-ṣú-ur-ša-IŠTAR áš-ta-pá-ar\nClean : I have sent Uṣur-ša-Ištar to Enišārum concerning your tin.\nPrediction : I have sent Uṣur-ša-Ištar to Enišārum concerning your tin.\nPN-GN Fixed :I have sent Uṣur-ša-Ištar to Enišaru concerning your tin.\n```\n \n\nEven If I used norm as replacement instead of lexeme , it still doesn't match ground truth \n\n\n[3]\n```\nDEBUG: Replacing 'Ikūn-pīya' with 'Ikūn-pīa' (Score: 94.11764705882352)\nDEBUG: Replacing 'Šu-Anum' with 'Šu-Anu' (Score: 92.3076923076923)\n---\nAkkadian :ma-na KÙ.BABBAR ṣa-ru-pá-am i-ku-pí-a DUMU šu-a-nim i-na ší-im SÍG.HI.A ša ú-ší-na-lam il5-qé\nClean : Ikūn-piya son of Šu-Anum has received 8 minas of refined silver from the proceeds of Ušinalam's wool.\nPrediction : Ikūn-pīya son of Šu-Anum has received 1 mina of refined silver from the proceeds of Ušinalam's wool.\nPN-GN Fixed :Ikūn-pīa son of Šu-Anu has received 1 mina of refined silver from the proceeds of Ušinalam's wool.\n```\n\n[4]\n```\nDEBUG: Replacing 'Dan-Aššur' with 'Dān-Aššur' (Score: 88.88888888888889)\nDEBUG: Replacing 'Enna-Suen' with 'Ennam-Suen' (Score: 94.73684210526315)\nDEBUG: Replacing 'Iddin-abum' with 'Iddin-abu' (Score: 94.73684210526315)\nAkkadian :dan-a-šùr sí-ki e-na-sú-in DUMU i-dí-a-bi4-im ú-kà-al-ma a-na ha-nu DUMU i-tur4-DINGIR a-na qá-ta-tim i-dí-šu\nClean : Dan-Aššur detains Enna-Suen son of Iddin-abum and he gave him to Hanu son of Itūr-ilī as guarantor.\nPrediction : Dan-Aššur detains Enna-Suen son of Iddin-abum and he gave him to Ḫanu son of Itūr-ilī as guarantor.\nPN-GN Fixed :Dān-Aššur detains Ennam-Suen son of Iddin-abu and he gave him to Ḫanu son of Itūr-ilī as guarantor.\n```", "3392118": "If I understand correctly, for PN/GN the lexemes are exclusively in the Nominative, whereas for norm they retain their case (N, G or D). So technically, for English translations, lexeme should be ok.\n\nHave you tried compiling a list of PN/GN from Ground Truth and comparing it to the OA_lexicone_eBL entries? I wonder how wide-spread the problem is.", "3392141": "The technique I’ve been using is using the soundex algorithm to mine the training set for variations of the same proper name from the lexicon. There is some variation maybe this approach could be used to converge on consensus spellings. To be honest though this type of approach is kind of hacking the evaluation metric you’ll get a better score but you won’t get a better translation model considering the quality of the leaderboard scores and makes me think there’s much lower hanging fruit than optimizing naming conventions", "3392148": "We should get a well-established norm for the names (maybe the onomasticon that @deeppast is working on will shed more light on the topic) but from what see the field of Old Assyrian is relatively new/fragmented into many groups, each using a slightly different norm.", "3392501": "Thank you all for your efforts here, these are great observations. I just posted the onomasticon.csv dataset so I'll look forward to hearing if it is useful:\nhttps://www.kaggle.com/datasets/deeppast/old-assyrian-grammars-and-other-resources/data?select=onomasticon.csv\nI was not able to list all the different aliases, but my hope is that by providing the spellings, and cross-referencing the OA_lexicone_eBL entries, that a lot of the many-to-many issues with the names in translations vs. transliteration can be resolvable." }, "source": "meta" }