{ "id": 666425, "title": "Academic Landscape of Old Akkadian: Chicago vs Munich vs French schools", "url": "/competitions/deep-past-initiative-machine-translation/discussion/666425", "author_name": "manwithacat", "post_date": "2026-01-07T04:20:25.899000", "votes": 9, "comment_count": 0, "views": 0, "content": "

Note - I'm a digital plumber, not a philologist, so this is my analysis trying to to puzzle out how to pull different data sources together for model training.

\n

Understanding Regional Scholarly Traditions in Akkadian Transliteration & Translation

\n

When working with Akkadian datasets for machine translation, you'll encounter texts prepared by scholars from different academic traditions. These differences matter for model training and evaluation. Here's a practical overview of the three major \"schools.\"

\n

The Three Major Traditions

\n

1. Chicago School (American)

\n

Key Resources:

\n\n

Characteristics:

\n\n

2. Munich/German School

\n

Key Resources:

\n\n

Characteristics:

\n\n

3. French School

\n

Key Resources:

\n\n

Characteristics:

\n\n

Practical Differences You'll See in Data

\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
FeatureChicagoMunichFrench
Semivowelyjvaries
Sign indicesCAD conventionsMZL numbersvaries
Translation styleInterpretive/contextualLiteral/philologicalContextual/archival
Vowel lengthā, ē, ī, ū (macron)samesame
SumerogramsUPPERCASEUPPERCASEOften lemmatized to Akkadian
\n

Key Technical Terms

\n\n

Why This Matters for ML

\n
    \n
  1. Training data consistency: Mixing sources may introduce inconsistent transliteration conventions
    2. \n
  2. Evaluation fairness: A model trained on Chicago-style data may score poorly against Munich-style test data on minor orthographic differences
    3. \n
  3. Translation style: Interpretive vs literal translations affect what the model learns as \"correct\"
    4. \n
\n

Recommendation

\n

When preparing training data, consider:

\n\n
\n

Sources:

\n", "messages": [ { "id": 3387463, "postDate": "2026-01-07T04:20:25.900Z", "content": "

Note - I'm a digital plumber, not a philologist, so this is my analysis trying to to puzzle out how to pull different data sources together for model training.

\n

Understanding Regional Scholarly Traditions in Akkadian Transliteration & Translation

\n

When working with Akkadian datasets for machine translation, you'll encounter texts prepared by scholars from different academic traditions. These differences matter for model training and evaluation. Here's a practical overview of the three major \"schools.\"

\n

The Three Major Traditions

\n

1. Chicago School (American)

\n

Key Resources:

\n\n

Characteristics:

\n\n

2. Munich/German School

\n

Key Resources:

\n\n

Characteristics:

\n\n

3. French School

\n

Key Resources:

\n\n

Characteristics:

\n\n

Practical Differences You'll See in Data

\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
FeatureChicagoMunichFrench
Semivowelyjvaries
Sign indicesCAD conventionsMZL numbersvaries
Translation styleInterpretive/contextualLiteral/philologicalContextual/archival
Vowel lengthā, ē, ī, ū (macron)samesame
SumerogramsUPPERCASEUPPERCASEOften lemmatized to Akkadian
\n

Key Technical Terms

\n\n

Why This Matters for ML

\n
    \n
  1. Training data consistency: Mixing sources may introduce inconsistent transliteration conventions
    2. \n
  2. Evaluation fairness: A model trained on Chicago-style data may score poorly against Munich-style test data on minor orthographic differences
    3. \n
  3. Translation style: Interpretive vs literal translations affect what the model learns as \"correct\"
    4. \n
\n

Recommendation

\n

When preparing training data, consider:

\n\n
\n

Sources:

\n", "rawMarkdown": "*Note - I'm a digital plumber, not a philologist, so this is my analysis trying to to puzzle out how to pull different data sources together for model training.*\n\n## Understanding Regional Scholarly Traditions in Akkadian Transliteration & Translation\n\nWhen working with Akkadian datasets for machine translation, you'll encounter texts prepared by scholars from different academic traditions. These differences matter for model training and evaluation. Here's a practical overview of the three major \"schools.\"\n\n### The Three Major Traditions\n\n#### 1. Chicago School (American)\n\nKey Resources:\n- https://isac.uchicago.edu/research/publications/chicago-assyrian-dictionary - 26 volumes, completed 2010\n- https://isac.uchicago.edu/research/publications/mad/mad-2-old-akkadian-writing-and-grammar (Materials for the Assyrian Dictionary) by I.J. Gelb\n\nCharacteristics:\n- Encyclopedic approach: CAD provides extensive context and idiomatic translations, not just word equivalents\n- Uses y for the semivowel (e.g., ayābu \"enemy\")\n- Tends toward interpretive translations that reconstruct cultural context\n- Strong emphasis on presenting words in \"meaningful context with full and idiomatic translation\"\n\n#### 2. Munich/German School\n\nKey Resources:\n- Wolfram von Soden's Akkadisches Handwörterbuch (AHw)\n- Rykle Borger's Mesopotamisches Zeichenlexikon (MZL) - the standard sign list\n- https://www.ebl.lmu.de/ at LMU Munich\n\nCharacteristics:\n- Uses j instead of y (e.g., ajābu in AHw vs ayābu in CAD)\n- Follows DMG (Deutsche Morgenländische Gesellschaft) conventions\n- More philologically conservative - closer to literal readings\n- Borger's MZL numbering system is now the de facto standard for sign identification\n- eBL project has ~25,000 tablets in transliteration with standardized ATF format\n\n#### 3. French School\n\nKey Resources:\n- https://www.archibab.fr/ - Archives babyloniennes project (CNRS/Collège de France)\n- Dominique Charpin's work on Old Babylonian texts\n\nCharacteristics:\n- Strong focus on archival/documentary texts (letters, contracts, administrative records)\n- Treats Sumerograms as Akkadian written logographically (e.g., transliterating ŠU.BA.AN.TI and tagging it as leqûm \"to take\")\n- Emphasizes historical and social context in translation\n- Extensive lemmatization connecting variant spellings to standard dictionary forms\n- Over 32,000 Old Babylonian documents in their corpus\n\nPractical Differences You'll See in Data\n\n | Feature | Chicago | Munich | French |\n |-------------------|-------------------------|----------------------|------------------------------|\n | Semivowel | y | j | varies |\n | Sign indices | CAD conventions | MZL numbers | varies |\n | Translation style | Interpretive/contextual | Literal/philological | Contextual/archival |\n | Vowel length | ā, ē, ī, ū (macron) | same | same |\n | Sumerograms | UPPERCASE | UPPERCASE | Often lemmatized to Akkadian |\n\n\n### Key Technical Terms\n\n- Transliteration: Sign-by-sign representation with hyphens: ṭu-up-šar-ru\n- Normalization/Transcription: Reconstructed pronunciation: ṭupšarru \"scribe\"\n- Lemmatization: Linking to dictionary headword (infinitive for verbs, nominative for nouns)\n\n### Why This Matters for ML\n\n1. Training data consistency: Mixing sources may introduce inconsistent transliteration conventions\n2. Evaluation fairness: A model trained on Chicago-style data may score poorly against Munich-style test data on minor orthographic differences\n3. Translation style: Interpretive vs literal translations affect what the model learns as \"correct\"\n\n### Recommendation\n\nWhen preparing training data, consider:\n- Normalizing j/y variations\n- Being consistent with sign index notation\n- Documenting which tradition your source texts follow\n- Accepting multiple valid translations in evaluation (literal AND interpretive)\n\n---\nSources:\n- https://isac.uchicago.edu/research/publications/chicago-assyrian-dictionary\n- https://www.ebl.lmu.de/\n- https://hal.science/hal-03752442/\n- https://academic.oup.com/edited-volume/27992/chapter/211695199\n- https://oracc.museum.upenn.edu/doc/help/languages/akkadian/akkadianstylesheet/index.html\n\n", "votes": 9 } ], "comments": [], "raw_markdown_by_id": { "3387463": "*Note - I'm a digital plumber, not a philologist, so this is my analysis trying to to puzzle out how to pull different data sources together for model training.*\n\n## Understanding Regional Scholarly Traditions in Akkadian Transliteration & Translation\n\nWhen working with Akkadian datasets for machine translation, you'll encounter texts prepared by scholars from different academic traditions. These differences matter for model training and evaluation. Here's a practical overview of the three major \"schools.\"\n\n### The Three Major Traditions\n\n#### 1. Chicago School (American)\n\nKey Resources:\n- https://isac.uchicago.edu/research/publications/chicago-assyrian-dictionary - 26 volumes, completed 2010\n- https://isac.uchicago.edu/research/publications/mad/mad-2-old-akkadian-writing-and-grammar (Materials for the Assyrian Dictionary) by I.J. Gelb\n\nCharacteristics:\n- Encyclopedic approach: CAD provides extensive context and idiomatic translations, not just word equivalents\n- Uses y for the semivowel (e.g., ayābu \"enemy\")\n- Tends toward interpretive translations that reconstruct cultural context\n- Strong emphasis on presenting words in \"meaningful context with full and idiomatic translation\"\n\n#### 2. Munich/German School\n\nKey Resources:\n- Wolfram von Soden's Akkadisches Handwörterbuch (AHw)\n- Rykle Borger's Mesopotamisches Zeichenlexikon (MZL) - the standard sign list\n- https://www.ebl.lmu.de/ at LMU Munich\n\nCharacteristics:\n- Uses j instead of y (e.g., ajābu in AHw vs ayābu in CAD)\n- Follows DMG (Deutsche Morgenländische Gesellschaft) conventions\n- More philologically conservative - closer to literal readings\n- Borger's MZL numbering system is now the de facto standard for sign identification\n- eBL project has ~25,000 tablets in transliteration with standardized ATF format\n\n#### 3. French School\n\nKey Resources:\n- https://www.archibab.fr/ - Archives babyloniennes project (CNRS/Collège de France)\n- Dominique Charpin's work on Old Babylonian texts\n\nCharacteristics:\n- Strong focus on archival/documentary texts (letters, contracts, administrative records)\n- Treats Sumerograms as Akkadian written logographically (e.g., transliterating ŠU.BA.AN.TI and tagging it as leqûm \"to take\")\n- Emphasizes historical and social context in translation\n- Extensive lemmatization connecting variant spellings to standard dictionary forms\n- Over 32,000 Old Babylonian documents in their corpus\n\nPractical Differences You'll See in Data\n\n | Feature | Chicago | Munich | French |\n |-------------------|-------------------------|----------------------|------------------------------|\n | Semivowel | y | j | varies |\n | Sign indices | CAD conventions | MZL numbers | varies |\n | Translation style | Interpretive/contextual | Literal/philological | Contextual/archival |\n | Vowel length | ā, ē, ī, ū (macron) | same | same |\n | Sumerograms | UPPERCASE | UPPERCASE | Often lemmatized to Akkadian |\n\n\n### Key Technical Terms\n\n- Transliteration: Sign-by-sign representation with hyphens: ṭu-up-šar-ru\n- Normalization/Transcription: Reconstructed pronunciation: ṭupšarru \"scribe\"\n- Lemmatization: Linking to dictionary headword (infinitive for verbs, nominative for nouns)\n\n### Why This Matters for ML\n\n1. Training data consistency: Mixing sources may introduce inconsistent transliteration conventions\n2. Evaluation fairness: A model trained on Chicago-style data may score poorly against Munich-style test data on minor orthographic differences\n3. Translation style: Interpretive vs literal translations affect what the model learns as \"correct\"\n\n### Recommendation\n\nWhen preparing training data, consider:\n- Normalizing j/y variations\n- Being consistent with sign index notation\n- Documenting which tradition your source texts follow\n- Accepting multiple valid translations in evaluation (literal AND interpretive)\n\n---\nSources:\n- https://isac.uchicago.edu/research/publications/chicago-assyrian-dictionary\n- https://www.ebl.lmu.de/\n- https://hal.science/hal-03752442/\n- https://academic.oup.com/edited-volume/27992/chapter/211695199\n- https://oracc.museum.upenn.edu/doc/help/languages/akkadian/akkadianstylesheet/index.html\n\n" } }