Article Citation: Gutherz G, Gordin S, Sáenz L, Levy O, Berant J. Translating Akkadian to English with neural machine translation. PNAS Nexus. 2023 May 2;2(5):pgad096. doi: 10.1093/pnasnexus/pgad096. PMID: 37143863; PMCID: PMC10153418.
\n\"Cuneiform is one of the earliest writing systems in recorded human history (ca. 3,400 BCE–75 CE). Hundreds of thousands of such texts were found over the last two centuries, most of which are written in Sumerian and Akkadian. The authors show the high potential in assisting scholars and interested laypeople alike, by using natural language processing (NLP) methods such as convolutional neural networks (CNN), to automatically translate Akkadian from cuneiform Unicode glyphs directly to English (C2E: Cuneiform to English Task) and from transliteration to English (T2E: Transliteration to English Task).\"
\n\"The authors showed that high-quality translations can be obtained when translating directly from cuneiform to English, as they got 36.52 and 37.47 Best Bilingual Evaluation Understudy 4 (BLEU4) scores for C2E (Cuneiform to English Task) and T2E (Transliteration to English Task), respectively. For C2E, their model was better than the translation memory baseline in 9.43, and for T2E, the difference was even higher and stands at 13.96. The model achieves best results in short- and medium-length sentences (c. 118 or less characters).\"
\n\"For the purposes of C2E, the signs of each text were encoded as strings of Unicode cuneiform glyphs generated by the Cuneify tool.\"
\n\"For the Cuneiform script (C2E), the authors used character-based tokenization with a small vocabulary of characters. For transliteration (T2E), they used BytePair Encoding (BPE) with the SentencePiece package, where the size of the vocabulary for the transliteration and English was set to 1,000 and 10,000, respectively.\"
\n\"The authors used Fairseq. Fairseq is a sequence modeling toolkit that allows researchers and developers to train custom models for translation, summarization, language modeling, and other text generation tasks.'
\n\"The best results overall were to be found in the short- and middle-length sentences. Longer sources produced more hallucinations or missing translations in the NMT results. This is promising for the usage in realistic scenarios, since all cuneiform texts are divided into manageable lines on the clay tablet.\"
\n\"The number of characters on an inscribed clay tablet can vary from period to period (signs in the **Old Babylonian period **are bigger than in the Neo-Babylonian period) and from genre to genre. Also, the number of columns by which a tablet is divided will determine the number of characters. Even in a single tablet, the number can vary to fill the space in the line.\"
\n\"The usage of standard sentence aligners such as the Gale–Church algorithm can be considered as a first step of the process to handle the alignment problem of the data set.\"
\n\"The input data for C2E could originate from already existing OCR tools in the Babylonian Engine (CuRe Demo and Decuneify), which would even allow a layperson to produce an NMT with the online notebook accompanying that publication, as well as part of the python package Akkademia (Akkadian).\"
\nhttps://pmc.ncbi.nlm.nih.gov/articles/PMC10153418/
\nhttps://github.com/cdli-gh/Machine-Translation
\nhttps://github.com/gaigutherz/Akkademia
", "messages": [ { "id": "3377812", "postDate": "12/17/2025 01:39:04", "content": "Article Citation: Gutherz G, Gordin S, Sáenz L, Levy O, Berant J. Translating Akkadian to English with neural machine translation. PNAS Nexus. 2023 May 2;2(5):pgad096. doi: 10.1093/pnasnexus/pgad096. PMID: 37143863; PMCID: PMC10153418.
\n\"Cuneiform is one of the earliest writing systems in recorded human history (ca. 3,400 BCE–75 CE). Hundreds of thousands of such texts were found over the last two centuries, most of which are written in Sumerian and Akkadian. The authors show the high potential in assisting scholars and interested laypeople alike, by using natural language processing (NLP) methods such as convolutional neural networks (CNN), to automatically translate Akkadian from cuneiform Unicode glyphs directly to English (C2E: Cuneiform to English Task) and from transliteration to English (T2E: Transliteration to English Task).\"
\n\"The authors showed that high-quality translations can be obtained when translating directly from cuneiform to English, as they got 36.52 and 37.47 Best Bilingual Evaluation Understudy 4 (BLEU4) scores for C2E (Cuneiform to English Task) and T2E (Transliteration to English Task), respectively. For C2E, their model was better than the translation memory baseline in 9.43, and for T2E, the difference was even higher and stands at 13.96. The model achieves best results in short- and medium-length sentences (c. 118 or less characters).\"
\n\"For the purposes of C2E, the signs of each text were encoded as strings of Unicode cuneiform glyphs generated by the Cuneify tool.\"
\n\"For the Cuneiform script (C2E), the authors used character-based tokenization with a small vocabulary of characters. For transliteration (T2E), they used BytePair Encoding (BPE) with the SentencePiece package, where the size of the vocabulary for the transliteration and English was set to 1,000 and 10,000, respectively.\"
\n\"The authors used Fairseq. Fairseq is a sequence modeling toolkit that allows researchers and developers to train custom models for translation, summarization, language modeling, and other text generation tasks.'
\n\"The best results overall were to be found in the short- and middle-length sentences. Longer sources produced more hallucinations or missing translations in the NMT results. This is promising for the usage in realistic scenarios, since all cuneiform texts are divided into manageable lines on the clay tablet.\"
\n\"The number of characters on an inscribed clay tablet can vary from period to period (signs in the **Old Babylonian period **are bigger than in the Neo-Babylonian period) and from genre to genre. Also, the number of columns by which a tablet is divided will determine the number of characters. Even in a single tablet, the number can vary to fill the space in the line.\"
\n\"The usage of standard sentence aligners such as the Gale–Church algorithm can be considered as a first step of the process to handle the alignment problem of the data set.\"
\n\"The input data for C2E could originate from already existing OCR tools in the Babylonian Engine (CuRe Demo and Decuneify), which would even allow a layperson to produce an NMT with the online notebook accompanying that publication, as well as part of the python package Akkademia (Akkadian).\"
\nhttps://pmc.ncbi.nlm.nih.gov/articles/PMC10153418/
\nhttps://github.com/cdli-gh/Machine-Translation
\nhttps://github.com/gaigutherz/Akkademia
", "rawMarkdown": "##Akkadian to English with neural machine translation\n\nArticle Citation: Gutherz G, Gordin S, Sáenz L, Levy O, Berant J. Translating Akkadian to English with neural machine translation. PNAS Nexus. 2023 May 2;2(5):pgad096. doi: 10.1093/pnasnexus/pgad096. PMID: 37143863; PMCID: PMC10153418.\n\n\"Cuneiform is one of the earliest writing systems in recorded human history (ca. 3,400 BCE–75 CE). Hundreds of thousands of such texts were found over the last two centuries, most of which are written in Sumerian and Akkadian. The authors show the high potential in assisting scholars and interested laypeople alike, by using natural language processing (NLP) methods such as convolutional neural networks (CNN), to automatically translate Akkadian from cuneiform Unicode glyphs directly to English (C2E: Cuneiform to English Task) and from transliteration to English (T2E: Transliteration to English Task).\"\n\n\"The authors showed that high-quality translations can be obtained when translating directly from cuneiform to English, as they got 36.52 and 37.47 Best Bilingual Evaluation Understudy 4 **(BLEU4)** scores for C2E (Cuneiform to English Task) and T2E (Transliteration to English Task), respectively. For C2E, their model was better than the translation memory baseline in 9.43, and for T2E, the difference was even higher and stands at 13.96. The model achieves best results in short- and medium-length sentences (c. 118 or less characters).\"\n\n\"For the purposes of C2E, the signs of each text were encoded as strings of Unicode cuneiform glyphs generated by the **Cuneify tool.**\"\n\n\"For the Cuneiform script (C2E), the authors used character-based tokenization with a small vocabulary of characters. For transliteration (T2E), they used BytePair Encoding (BPE) with the SentencePiece package, where the size of the vocabulary for the transliteration and English was set to 1,000 and 10,000, respectively.\"\n\n### Fairseq toolkit\n\n\"The authors used **Fairseq**. Fairseq is a sequence modeling toolkit that allows researchers and developers to train custom models for translation, summarization, language modeling, and other text generation tasks.'\n\n\"The best results overall were to be found in the short- and middle-length sentences. Longer sources produced more hallucinations or missing translations in the NMT results. This is promising for the usage in realistic scenarios, since all cuneiform texts are divided into manageable lines on the clay tablet.\"\n\n\"The number of characters on an inscribed clay tablet can vary from period to period (signs in the **Old Babylonian period **are bigger than in the Neo-Babylonian period) and **from genre to genre**. Also, the number of columns by which a tablet is divided will determine the number of characters. Even in a single tablet, the number can vary to fill the space in the line.\"\n\n\"The usage of standard sentence aligners such as the **Gale–Church algorithm** can be considered as a first step of the process to handle the alignment problem of the data set.\"\n\n### Python Package Akkademia (Akkadian)\n\n\"The input data for C2E could originate from already existing OCR tools in the Babylonian Engine (CuRe Demo and Decuneify), which would even allow a layperson to produce an NMT with the online notebook accompanying that publication, as well as part of the **python package Akkademia (Akkadian)**.\"\n\nhttps://pmc.ncbi.nlm.nih.gov/articles/PMC10153418/\n\nhttps://github.com/cdli-gh/Machine-Translation\n\nhttps://github.com/gaigutherz/Akkademia", "votes": null }, { "id": "3378241", "postDate": "12/17/2025 19:40:19", "content": "Yes, the Akkademia repo is a benchmark MT study for Neo-Assyrian texts (ca 911–612 BCE). Note that the time period here is about 1000 years later from the Old Assyrian period (ca. 1950-1750 BCE). It would be a mistake to assume these two different textual datasets contain the very same language usage, in terms of genre and register.
", "rawMarkdown": "Yes, the Akkademia repo is a benchmark MT study for Neo-Assyrian texts (ca 911–612 BCE). Note that the time period here is about 1000 years later from the Old Assyrian period (ca. 1950-1750 BCE). It would be a mistake to assume these two different textual datasets contain the very same language usage, in terms of genre and register.", "votes": null }, { "id": "3378255", "postDate": "12/17/2025 20:22:03", "content": "Indeed DeepPast,
\nI didn't assume that both periods are equal just because the Assyrian word. \nLanguages are dynamic and subject to changes. We see Portugal/portuguese and Brazilian(my country) \"only\" in 500 years we notice many modifications. Imagine these Assyrian (Old and Neo) during One thousand years how far could their language have evolved.
\nWhat I really assume is, that I'm not able to translate any of them Assyrian, Sumerian, Amorite, Babylonian.
\nAny way, I tried to use python Akkadian on my Kaggle Notebook, unfortunately it resulted in:
\n!pip install akkadian
\n\"No module named 'allennlp'\" after I ran
\n\n\nimport akkadian.transliterate as akk
\nprint(akk.transliterate(\"𒁹𒀭𒌍𒋀𒈨𒌍𒌷𒁀\"))
\n
Which was a pity cause it would be a nice experience.
\nThank you.
\nMarília Prata.
", "rawMarkdown": "Indeed DeepPast,\n\nI didn't assume that both periods are equal just because the Assyrian word. \nLanguages are dynamic and subject to changes. We see Portugal/portuguese and Brazilian(my country) \"only\" in 500 years we notice many modifications. Imagine these Assyrian (Old and Neo) during One thousand years how far could their language have evolved.\n\nWhat I really assume is, that I'm not able to translate any of them Assyrian, Sumerian, Amorite, Babylonian.\n\nAny way, I tried to use python Akkadian on my Kaggle Notebook, unfortunately it resulted in:\n\n!pip install akkadian\n\n\"No module named **'allennlp'**\" after I ran\n\n>import akkadian.transliterate as akk\n\n>print(akk.transliterate(\"𒁹𒀭𒌍𒋀𒈨𒌍𒌷𒁀\"))\n\nWhich was a pity cause it would be a nice experience.\n\nThank you.\n\nMarília Prata.", "votes": null }, { "id": "3378290", "postDate": "12/17/2025 21:05:17", "content": "Good to hear you're considering the time difference, just making sure. I worked on replicating this workflow in a jupyter notebook, perhaps that will be useful, but it still requires some tinkering… \nhttps://github.com/ancient-world-citation-analysis/EvaCun-Colab-Notebook
", "rawMarkdown": "Good to hear you're considering the time difference, just making sure. I worked on replicating this workflow in a jupyter notebook, perhaps that will be useful, but it still requires some tinkering... \nhttps://github.com/ancient-world-citation-analysis/EvaCun-Colab-Notebook", "votes": null }, { "id": "3378370", "postDate": "12/18/2025 00:57:55", "content": "I'll try to run them in a Private Notebook. Due to my poor knowledge, only when I read the outcomes I'd be able to get it.
\nThen, I'll try to understand the EvaCun: ORACC Akkadian Parallel Corpus
\nI've checked some of the FactGrid Cuneiform:\ne. g. akkadian_train.txt.
\nThank you.
", "rawMarkdown": "I'll try to run them in a Private Notebook. Due to my poor knowledge, only when I read the outcomes I'd be able to get it. \n\nThen, I'll try to understand the EvaCun: ORACC Akkadian Parallel Corpus\n\nI've checked some of the FactGrid Cuneiform:\ne. g. akkadian_train.txt. \n\nThank you.", "votes": null }, { "id": "3378476", "postDate": "12/18/2025 09:50:23", "content": "Thank you for the best Bilingual Evaluation Understudy 4 (BLEU4) scores @mpwolke
", "rawMarkdown": "Thank you for the best Bilingual Evaluation Understudy 4 (BLEU4) scores @mpwolke", "votes": null }, { "id": "3380697", "postDate": "12/22/2025 20:34:22", "content": "One thing to note is that Gale–Church algorithm relies on the sentence length ratio of the sequences for alignment this does not transfer well to morphologically rich Semitic languages like Akkadian take a look at the descriptive stats below from the train set if you use Gale–Church you are likely to get very noisy alignments.
\n| | |\n|Character-length statistics |Akkadian transliteration |English translation\n|Mean |~426 chars |~500 chars\n|Std dev |~270 |~466\n|Min | 21 |6\n|Max |932 |3,895\n| | |
", "rawMarkdown": "One thing to note is that Gale–Church algorithm relies on the sentence length ratio of the sequences for alignment this does not transfer well to morphologically rich Semitic languages like Akkadian take a look at the descriptive stats below from the train set if you use Gale–Church you are likely to get very noisy alignments.\n\n| | |\n|Character-length statistics |Akkadian transliteration\t|English translation\n|Mean\t |~426 chars\t |~500 chars\n|Std dev\t |~270\t |~466\n|Min |\t21\t |6\n|Max\t |932\t |3,895\n| | |", "votes": null }, { "id": "3380721", "postDate": "12/22/2025 22:13:46", "content": "Thank you for this additional information Bornstein. I've just reproduced what I read on that article:\nhttps://pmc.ncbi.nlm.nih.gov/articles/PMC10153418/
", "rawMarkdown": "Thank you for this additional information Bornstein. I've just reproduced what I read on that article:\nhttps://pmc.ncbi.nlm.nih.gov/articles/PMC10153418/", "votes": null } ], "comments": [ { "id": 3378241, "author_name": "deeppast", "author_url": "", "post_date": "12/17/2025 19:40:19", "content": "Yes, the Akkademia repo is a benchmark MT study for Neo-Assyrian texts (ca 911–612 BCE). Note that the time period here is about 1000 years later from the Old Assyrian period (ca. 1950-1750 BCE). It would be a mistake to assume these two different textual datasets contain the very same language usage, in terms of genre and register.
", "votes": null, "replies": [ { "id": 3378255, "author_name": "mpwolke", "author_url": "", "post_date": "12/17/2025 20:22:03", "content": "Indeed DeepPast,
\nI didn't assume that both periods are equal just because the Assyrian word. \nLanguages are dynamic and subject to changes. We see Portugal/portuguese and Brazilian(my country) \"only\" in 500 years we notice many modifications. Imagine these Assyrian (Old and Neo) during One thousand years how far could their language have evolved.
\nWhat I really assume is, that I'm not able to translate any of them Assyrian, Sumerian, Amorite, Babylonian.
\nAny way, I tried to use python Akkadian on my Kaggle Notebook, unfortunately it resulted in:
\n!pip install akkadian
\n\"No module named 'allennlp'\" after I ran
\n\n\nimport akkadian.transliterate as akk
\nprint(akk.transliterate(\"𒁹𒀭𒌍𒋀𒈨𒌍𒌷𒁀\"))
\n
Which was a pity cause it would be a nice experience.
\nThank you.
\nMarília Prata.
", "votes": null, "replies": [ { "id": 3378290, "author_name": "deeppast", "author_url": "", "post_date": "12/17/2025 21:05:17", "content": "Good to hear you're considering the time difference, just making sure. I worked on replicating this workflow in a jupyter notebook, perhaps that will be useful, but it still requires some tinkering… \nhttps://github.com/ancient-world-citation-analysis/EvaCun-Colab-Notebook
", "votes": null, "replies": [ { "id": 3378370, "author_name": "mpwolke", "author_url": "", "post_date": "12/18/2025 00:57:55", "content": "I'll try to run them in a Private Notebook. Due to my poor knowledge, only when I read the outcomes I'd be able to get it.
\nThen, I'll try to understand the EvaCun: ORACC Akkadian Parallel Corpus
\nI've checked some of the FactGrid Cuneiform:\ne. g. akkadian_train.txt.
\nThank you.
", "votes": null, "replies": [] } ] } ] } ] }, { "id": 3378476, "author_name": "navneetbende", "author_url": "", "post_date": "12/18/2025 09:50:23", "content": "Thank you for the best Bilingual Evaluation Understudy 4 (BLEU4) scores @mpwolke
", "votes": null, "replies": [] }, { "id": 3380697, "author_name": "aaronbornstein", "author_url": "", "post_date": "12/22/2025 20:34:22", "content": "One thing to note is that Gale–Church algorithm relies on the sentence length ratio of the sequences for alignment this does not transfer well to morphologically rich Semitic languages like Akkadian take a look at the descriptive stats below from the train set if you use Gale–Church you are likely to get very noisy alignments.
\n| | |\n|Character-length statistics |Akkadian transliteration |English translation\n|Mean |~426 chars |~500 chars\n|Std dev |~270 |~466\n|Min | 21 |6\n|Max |932 |3,895\n| | |
", "votes": null, "replies": [ { "id": 3380721, "author_name": "mpwolke", "author_url": "", "post_date": "12/22/2025 22:13:46", "content": "Thank you for this additional information Bornstein. I've just reproduced what I read on that article:\nhttps://pmc.ncbi.nlm.nih.gov/articles/PMC10153418/
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "3377812": "##Akkadian to English with neural machine translation\n\nArticle Citation: Gutherz G, Gordin S, Sáenz L, Levy O, Berant J. Translating Akkadian to English with neural machine translation. PNAS Nexus. 2023 May 2;2(5):pgad096. doi: 10.1093/pnasnexus/pgad096. PMID: 37143863; PMCID: PMC10153418.\n\n\"Cuneiform is one of the earliest writing systems in recorded human history (ca. 3,400 BCE–75 CE). Hundreds of thousands of such texts were found over the last two centuries, most of which are written in Sumerian and Akkadian. The authors show the high potential in assisting scholars and interested laypeople alike, by using natural language processing (NLP) methods such as convolutional neural networks (CNN), to automatically translate Akkadian from cuneiform Unicode glyphs directly to English (C2E: Cuneiform to English Task) and from transliteration to English (T2E: Transliteration to English Task).\"\n\n\"The authors showed that high-quality translations can be obtained when translating directly from cuneiform to English, as they got 36.52 and 37.47 Best Bilingual Evaluation Understudy 4 **(BLEU4)** scores for C2E (Cuneiform to English Task) and T2E (Transliteration to English Task), respectively. For C2E, their model was better than the translation memory baseline in 9.43, and for T2E, the difference was even higher and stands at 13.96. The model achieves best results in short- and medium-length sentences (c. 118 or less characters).\"\n\n\"For the purposes of C2E, the signs of each text were encoded as strings of Unicode cuneiform glyphs generated by the **Cuneify tool.**\"\n\n\"For the Cuneiform script (C2E), the authors used character-based tokenization with a small vocabulary of characters. For transliteration (T2E), they used BytePair Encoding (BPE) with the SentencePiece package, where the size of the vocabulary for the transliteration and English was set to 1,000 and 10,000, respectively.\"\n\n### Fairseq toolkit\n\n\"The authors used **Fairseq**. Fairseq is a sequence modeling toolkit that allows researchers and developers to train custom models for translation, summarization, language modeling, and other text generation tasks.'\n\n\"The best results overall were to be found in the short- and middle-length sentences. Longer sources produced more hallucinations or missing translations in the NMT results. This is promising for the usage in realistic scenarios, since all cuneiform texts are divided into manageable lines on the clay tablet.\"\n\n\"The number of characters on an inscribed clay tablet can vary from period to period (signs in the **Old Babylonian period **are bigger than in the Neo-Babylonian period) and **from genre to genre**. Also, the number of columns by which a tablet is divided will determine the number of characters. Even in a single tablet, the number can vary to fill the space in the line.\"\n\n\"The usage of standard sentence aligners such as the **Gale–Church algorithm** can be considered as a first step of the process to handle the alignment problem of the data set.\"\n\n### Python Package Akkademia (Akkadian)\n\n\"The input data for C2E could originate from already existing OCR tools in the Babylonian Engine (CuRe Demo and Decuneify), which would even allow a layperson to produce an NMT with the online notebook accompanying that publication, as well as part of the **python package Akkademia (Akkadian)**.\"\n\nhttps://pmc.ncbi.nlm.nih.gov/articles/PMC10153418/\n\nhttps://github.com/cdli-gh/Machine-Translation\n\nhttps://github.com/gaigutherz/Akkademia", "3378241": "Yes, the Akkademia repo is a benchmark MT study for Neo-Assyrian texts (ca 911–612 BCE). Note that the time period here is about 1000 years later from the Old Assyrian period (ca. 1950-1750 BCE). It would be a mistake to assume these two different textual datasets contain the very same language usage, in terms of genre and register.", "3378255": "Indeed DeepPast,\n\nI didn't assume that both periods are equal just because the Assyrian word. \nLanguages are dynamic and subject to changes. We see Portugal/portuguese and Brazilian(my country) \"only\" in 500 years we notice many modifications. Imagine these Assyrian (Old and Neo) during One thousand years how far could their language have evolved.\n\nWhat I really assume is, that I'm not able to translate any of them Assyrian, Sumerian, Amorite, Babylonian.\n\nAny way, I tried to use python Akkadian on my Kaggle Notebook, unfortunately it resulted in:\n\n!pip install akkadian\n\n\"No module named **'allennlp'**\" after I ran\n\n>import akkadian.transliterate as akk\n\n>print(akk.transliterate(\"𒁹𒀭𒌍𒋀𒈨𒌍𒌷𒁀\"))\n\nWhich was a pity cause it would be a nice experience.\n\nThank you.\n\nMarília Prata.", "3378290": "Good to hear you're considering the time difference, just making sure. I worked on replicating this workflow in a jupyter notebook, perhaps that will be useful, but it still requires some tinkering... \nhttps://github.com/ancient-world-citation-analysis/EvaCun-Colab-Notebook", "3378370": "I'll try to run them in a Private Notebook. Due to my poor knowledge, only when I read the outcomes I'd be able to get it. \n\nThen, I'll try to understand the EvaCun: ORACC Akkadian Parallel Corpus\n\nI've checked some of the FactGrid Cuneiform:\ne. g. akkadian_train.txt. \n\nThank you.", "3378476": "Thank you for the best Bilingual Evaluation Understudy 4 (BLEU4) scores @mpwolke", "3380697": "One thing to note is that Gale–Church algorithm relies on the sentence length ratio of the sequences for alignment this does not transfer well to morphologically rich Semitic languages like Akkadian take a look at the descriptive stats below from the train set if you use Gale–Church you are likely to get very noisy alignments.\n\n| | |\n|Character-length statistics |Akkadian transliteration\t|English translation\n|Mean\t |~426 chars\t |~500 chars\n|Std dev\t |~270\t |~466\n|Min |\t21\t |6\n|Max\t |932\t |3,895\n| | |", "3380721": "Thank you for this additional information Bornstein. I've just reproduced what I read on that article:\nhttps://pmc.ncbi.nlm.nih.gov/articles/PMC10153418/" }, "source": "meta" }