I am wondering about the usefulness of getting an automated translation with the transliteration as a starting point, because getting to the transliteration implies a huge work of basically understanding the text, deciding between competing hypothesis about the value of all cuneiform sign (basically, is it used as a logogram, as a phonogram, if yes with which phonetic value, is it a determinative, etc.). Or may be I am wrong, but I would like to get some explanation about that.
", "messages": [ { "id": "3378335", "postDate": "12/17/2025 23:00:26", "content": "I am wondering about the usefulness of getting an automated translation with the transliteration as a starting point, because getting to the transliteration implies a huge work of basically understanding the text, deciding between competing hypothesis about the value of all cuneiform sign (basically, is it used as a logogram, as a phonogram, if yes with which phonetic value, is it a determinative, etc.). Or may be I am wrong, but I would like to get some explanation about that.
", "rawMarkdown": "I am wondering about the usefulness of getting an automated translation with the transliteration as a starting point, because getting to the transliteration implies a huge work of basically understanding the text, deciding between competing hypothesis about the value of all cuneiform sign (basically, is it used as a logogram, as a phonogram, if yes with which phonetic value, is it a determinative, etc.). Or may be I am wrong, but I would like to get some explanation about that.", "votes": null }, { "id": "3378338", "postDate": "12/17/2025 23:13:47", "content": "In this study from 2023 the researchers tested this very question: https://pmc.ncbi.nlm.nih.gov/articles/PMC10153418/\nYou can see that after aligning both the transliterated text with English (T2E) and the cuneiform signs with English (C2E) that the T2E performed much better. As you can imagine there is currently very little in terms of tools and models for tokenizing / embedding cuneiform signs.
", "rawMarkdown": "In this study from 2023 the researchers tested this very question: https://pmc.ncbi.nlm.nih.gov/articles/PMC10153418/\nYou can see that after aligning both the transliterated text with English (T2E) and the cuneiform signs with English (C2E) that the T2E performed much better. As you can imagine there is currently very little in terms of tools and models for tokenizing / embedding cuneiform signs.", "votes": null }, { "id": "3378933", "postDate": "12/18/2025 21:12:12", "content": "OK, I have also seen your article: https://pmc.ncbi.nlm.nih.gov/articles/PMC7592802/, which is also very interesting. The complete solution would be to combine the cuneiform -> transliteration process(C2T) + transliteration -> translation process (T2E), which would be the most useful for an assyriologist. After, I don't know which is better C2T + T2E or directly C2E. But in the first article you mentioned, in the conclusion I read: \"The marginal difference between T2E and C2E tasks exhibits a potential advantage when producing translations from the original text, without the need of transliteration. An option for further research is evaluating a multi-source task, using as source both the transliteration and the Cuneiform, as it may outperform each task separately.\"
", "rawMarkdown": "OK, I have also seen your article: https://pmc.ncbi.nlm.nih.gov/articles/PMC7592802/, which is also very interesting. The complete solution would be to combine the cuneiform -> transliteration process(C2T) + transliteration -> translation process (T2E), which would be the most useful for an assyriologist. After, I don't know which is better C2T + T2E or directly C2E. But in the first article you mentioned, in the conclusion I read: \"*The marginal difference between T2E and C2E tasks exhibits a potential advantage when producing translations from the original text, without the need of transliteration. An option for further research is evaluating a multi-source task, using as source both the transliteration and the Cuneiform, as it may outperform each task separately.*\"", "votes": null }, { "id": "3378938", "postDate": "12/18/2025 21:32:43", "content": "And I have got an additional question (may be irrelevant) on the tokenization issue. I always wondered why the current tokenization methods ignore the morphology of languages relying only on statistical algorithms. So there is very little in terms of models for tokenizing cuneiform signs, and the transliteration of cuneiforms into latin would also not be too well adapted to the language. In this context, is it not possible to use as the transliteration simply the cuneiform sign's id (in upper case, for instance KU3-UD for KUG-BABBAR, silver, but also for all syllabograms and determinatives), it would approximate a C2E.
", "rawMarkdown": "And I have got an additional question (may be irrelevant) on the tokenization issue. I always wondered why the current tokenization methods ignore the morphology of languages relying only on statistical algorithms. So there is very little in terms of models for tokenizing cuneiform signs, and the transliteration of cuneiforms into latin would also not be too well adapted to the language. In this context, is it not possible to use as the transliteration simply the cuneiform sign's id (in upper case, for instance KU3-UD for KUG-BABBAR, silver, but also for all syllabograms and determinatives), it would approximate a C2E.", "votes": null }, { "id": "3379347", "postDate": "12/19/2025 16:54:29", "content": "I appreciate your thoughtful questions. The current tokenization methods were primarily developed for Indo-European languages, which in general are stem-based and exhibit a low morphological inflection. In contrast, Semitic languages, among others, are root-based (i.e. employ radicals with affixes as morphological particles) and are highly morphologically inflected. \nAs to your question of the possibility to simplify the Latin transcription method, the different sign values or \"readings\" carry different meanings as well. So the name of the sign \"UD\" 𒌓 means: \"day; time; light; weather; storm;\" whereas BABBAR means: \"white (zalag): pure; bright\". The work of transliteration has in many ways reduced a significant amount of semantic ambiguity already. Considering this challenge is really about getting the best English translation as an outcome, it may be wise to keep as much of the specifics of the language as possible. I will leave that up to you to test.
", "rawMarkdown": "I appreciate your thoughtful questions. The current tokenization methods were primarily developed for Indo-European languages, which in general are stem-based and exhibit a low morphological inflection. In contrast, Semitic languages, among others, are root-based (i.e. employ radicals with affixes as morphological particles) and are highly morphologically inflected. \nAs to your question of the possibility to simplify the Latin transcription method, the different sign values or \"readings\" carry different meanings as well. So the name of the sign \"UD\"  𒌓 means: \"day; time; light; weather; storm;\" whereas BABBAR means: \"white (zalag): pure; bright\". The work of transliteration has in many ways reduced a significant amount of semantic ambiguity already. Considering this challenge is really about getting the best English translation as an outcome, it may be wise to keep as much of the specifics of the language as possible. I will leave that up to you to test.", "votes": null } ], "comments": [ { "id": 3378338, "author_name": "deeppast", "author_url": "", "post_date": "12/17/2025 23:13:47", "content": "In this study from 2023 the researchers tested this very question: https://pmc.ncbi.nlm.nih.gov/articles/PMC10153418/\nYou can see that after aligning both the transliterated text with English (T2E) and the cuneiform signs with English (C2E) that the T2E performed much better. As you can imagine there is currently very little in terms of tools and models for tokenizing / embedding cuneiform signs.
", "votes": null, "replies": [ { "id": 3378933, "author_name": "guokefei", "author_url": "", "post_date": "12/18/2025 21:12:12", "content": "OK, I have also seen your article: https://pmc.ncbi.nlm.nih.gov/articles/PMC7592802/, which is also very interesting. The complete solution would be to combine the cuneiform -> transliteration process(C2T) + transliteration -> translation process (T2E), which would be the most useful for an assyriologist. After, I don't know which is better C2T + T2E or directly C2E. But in the first article you mentioned, in the conclusion I read: \"The marginal difference between T2E and C2E tasks exhibits a potential advantage when producing translations from the original text, without the need of transliteration. An option for further research is evaluating a multi-source task, using as source both the transliteration and the Cuneiform, as it may outperform each task separately.\"
", "votes": null, "replies": [] }, { "id": 3378938, "author_name": "guokefei", "author_url": "", "post_date": "12/18/2025 21:32:43", "content": "And I have got an additional question (may be irrelevant) on the tokenization issue. I always wondered why the current tokenization methods ignore the morphology of languages relying only on statistical algorithms. So there is very little in terms of models for tokenizing cuneiform signs, and the transliteration of cuneiforms into latin would also not be too well adapted to the language. In this context, is it not possible to use as the transliteration simply the cuneiform sign's id (in upper case, for instance KU3-UD for KUG-BABBAR, silver, but also for all syllabograms and determinatives), it would approximate a C2E.
", "votes": null, "replies": [ { "id": 3379347, "author_name": "deeppast", "author_url": "", "post_date": "12/19/2025 16:54:29", "content": "I appreciate your thoughtful questions. The current tokenization methods were primarily developed for Indo-European languages, which in general are stem-based and exhibit a low morphological inflection. In contrast, Semitic languages, among others, are root-based (i.e. employ radicals with affixes as morphological particles) and are highly morphologically inflected. \nAs to your question of the possibility to simplify the Latin transcription method, the different sign values or \"readings\" carry different meanings as well. So the name of the sign \"UD\" 𒌓 means: \"day; time; light; weather; storm;\" whereas BABBAR means: \"white (zalag): pure; bright\". The work of transliteration has in many ways reduced a significant amount of semantic ambiguity already. Considering this challenge is really about getting the best English translation as an outcome, it may be wise to keep as much of the specifics of the language as possible. I will leave that up to you to test.
", "votes": null, "replies": [] } ] } ] } ], "raw_markdown_by_id": { "3378335": "I am wondering about the usefulness of getting an automated translation with the transliteration as a starting point, because getting to the transliteration implies a huge work of basically understanding the text, deciding between competing hypothesis about the value of all cuneiform sign (basically, is it used as a logogram, as a phonogram, if yes with which phonetic value, is it a determinative, etc.). Or may be I am wrong, but I would like to get some explanation about that.", "3378338": "In this study from 2023 the researchers tested this very question: https://pmc.ncbi.nlm.nih.gov/articles/PMC10153418/\nYou can see that after aligning both the transliterated text with English (T2E) and the cuneiform signs with English (C2E) that the T2E performed much better. As you can imagine there is currently very little in terms of tools and models for tokenizing / embedding cuneiform signs.", "3378933": "OK, I have also seen your article: https://pmc.ncbi.nlm.nih.gov/articles/PMC7592802/, which is also very interesting. The complete solution would be to combine the cuneiform -> transliteration process(C2T) + transliteration -> translation process (T2E), which would be the most useful for an assyriologist. After, I don't know which is better C2T + T2E or directly C2E. But in the first article you mentioned, in the conclusion I read: \"*The marginal difference between T2E and C2E tasks exhibits a potential advantage when producing translations from the original text, without the need of transliteration. An option for further research is evaluating a multi-source task, using as source both the transliteration and the Cuneiform, as it may outperform each task separately.*\"", "3378938": "And I have got an additional question (may be irrelevant) on the tokenization issue. I always wondered why the current tokenization methods ignore the morphology of languages relying only on statistical algorithms. So there is very little in terms of models for tokenizing cuneiform signs, and the transliteration of cuneiforms into latin would also not be too well adapted to the language. In this context, is it not possible to use as the transliteration simply the cuneiform sign's id (in upper case, for instance KU3-UD for KUG-BABBAR, silver, but also for all syllabograms and determinatives), it would approximate a C2E.", "3379347": "I appreciate your thoughtful questions. The current tokenization methods were primarily developed for Indo-European languages, which in general are stem-based and exhibit a low morphological inflection. In contrast, Semitic languages, among others, are root-based (i.e. employ radicals with affixes as morphological particles) and are highly morphologically inflected. \nAs to your question of the possibility to simplify the Latin transcription method, the different sign values or \"readings\" carry different meanings as well. So the name of the sign \"UD\"  𒌓 means: \"day; time; light; weather; storm;\" whereas BABBAR means: \"white (zalag): pure; bright\". The work of transliteration has in many ways reduced a significant amount of semantic ambiguity already. Considering this challenge is really about getting the best English translation as an outcome, it may be wise to keep as much of the specifics of the language as possible. I will leave that up to you to test." }, "source": "meta" }