Restoration of Fragmentary Babylonian Texts Using Recurrent Neural Networks

What it is about

Clay tablets from Achaemenid Babylonia are often broken, and experts fill the gaps by hand. The authors collect about 2,250 late Babylonian archival texts (legal, economic and administrative) from an online corpus, tokenise them with a custom Akkadian tokenizer (names, gods, places and months become class tokens; broken parts become a break token; rare words become an unknown token), and train an LSTM language model. The vocabulary has 1,549 types over about 221,000 words. The LSTM reaches a test perplexity of 5.02 against 22.87 for a bigram model. On a cloze test that removes the middle word of test sentences of length 10 or more (166 sentences), the LSTM with full sentence context reaches a mean reciprocal rank of 0.89 and hit@10 of 95.5%, against 0.64 and 83.6% for the bigram. A 52-question multiple-choice test gives 88.5% accuracy, and most errors are semantic, not syntactic.

Ideas

• I1. A gap-filler for damaged Hebrew and Aramaic text in tools/hbo: remove a word, rank candidates from the lexicon and the grammar, and report MRR and hit@k on held-out verses, using the same cloze protocol as the paper.
• I2. A transparent baseline first: a bigram model read in both directions (previous and next word) already reaches hit@1 of 52% in the paper. Build it from counts over MACULA before any neural model, and keep it as the reference that any learned ranker must beat.
• I3. Class tokens and a break token in the Hebrew tokenizer for fragmentary inscriptions: a token for a damaged span, and for personal names and place names marked by a determinative, so that the model does not spend its vocabulary on them.

For Mova

I1 and I2 go into tools/hbo as a restoration test, with the bigram counts stored as a table; the neural ranker is not part of the SLM. I3 applies to fragmentary texts in Biblical Aramaic texts, annotated and Discourse: narrative, speech and the shape of a text.

Cite

The paper itself, not this card. BibTeX from the library's record of the work. DOI 10.1073/pnas.2003794117 · arXiv 2003.01912

@misc{fetaya2020restoration,
  title = {{Restoration of Fragmentary Babylonian Texts Using Recurrent Neural Networks}},
  author = {Fetaya, Ethan and Lifshitz, Yonatan and Aaron, Elad and Gordin, Shai},
  year = {2020},
  doi = {10.1073/pnas.2003794117},
  eprint = {2003.01912},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.1073/pnas.2003794117}
}

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/papers/langstream-2020-fetaya-restoration-fragmentary-babylonian-texts-using/