Graph convolutional networks for exploring authorship hypotheses

What it is about

The paper casts the Documentary Hypothesis, the claim that the Torah combines several named sources, as a span classification task. The data is the Westminster Leningrad Codex (vowel points removed), with 378 labelled spans across eight source labels. To avoid topic overfitting, the bag-of-morph features keep only the morphs that occur in every source, which cuts 6,625 morph types to 70 (about half function morphs, a fifth verbs, a fifth common nouns). Graph convolutional networks, which see each span's neighbours (a sibling relation between spans that share a parent), beat logistic regression and MLPs on macro F-score; the best row of its Table 2 is about 48.6. Granularity indicators help the models with hidden layers but hurt when the full vocabulary is used. The error analysis notes that a shared word such as "wife" can push a span to the wrong source.

Ideas

• I1. A vocabulary filter for stylistic tests: keep only the morphs that appear in every class before any Torah style measure. This blocks a topic from identifying a source.
• I2. Granularity (book, chapter, verse, word) as a feature has to be tested with and without the topic features, twice, because the paper shows it can help in one setting and do harm in another.
• I3. The sibling relation between spans can be an explicit, deterministic edge in the world store: are neighbouring spans from the same source? This is a graph query, not a neural model.

For Mova

I1 is a filter for stylistic tests in tools/macula and a guide for the function-word tables in Pronouns, particles and prepositions. I2 is a measurement rule for any feature test. I3 belongs to world (relations between units) and to Discourse: narrative, speech and the shape of a text for structure across a text.

Cite

The paper itself, not this card. BibTeX from the library's record of the work. DOI 10.18653/v1/w19-2510

@inproceedings{lippincott2019graph,
  title = {{Graph convolutional networks for exploring authorship hypotheses}},
  author = {Lippincott, Tom},
  booktitle = {Proceedings of the 3rd Joint SIGHUM Workshop on Computational Linguistics for Cultural Heritage, Social Sciences, Humanities and Literature},
  year = {2019},
  doi = {10.18653/v1/w19-2510},
  url = {https://doi.org/10.18653/v1/w19-2510}
}

References

10 references, 1 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Adam Amram, Anat Ben-David, and Reut Tsarfaty. 2018. Representations and Architectures in Neural Sentiment Analysis for Morphologically Rich Languages: A Case Study from Modern Hebrew. In Proceedings of the 27th International Conference on Computational Linguistics, pages 2242–2252. find our card
2. Association for Literary and Linguistic Computing. 1977. Oxford Archive of Electronic Literature. find
3. David M Blei, Andrew Y Ng, and Michael I Jordan. 2003. Latent Dirichlet allocation. Journal of machine Learning research, 3(Jan):993–1022. find
4. TEI Consortium editors. 2019. TEI P5: Guidelines for Electronic Text Encoding and Interchange. find
5. Richard Elliott Friedman. 1987. Who Wrote the Bible? Simon and Schuster. find
6. Richard Elliott Friedman. 2003. The Bible with Sources Revealed. HarperCollins. find
7. Thomas N Kipf and Max Welling. 2016. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907. ↗
8. John Boaz Lee, Ryan A Rossi, Xiangnan Kong, Sungchul Kim, Eunyee Koh, and Anup Rao. 2018. Higher-order graph convolutional networks. arXiv preprint arXiv:1809.07697. ↗
9. Thomas Lippincott. 2009. A Framework for Multilayered Boundary Detection. find
10. Digital Humanities 2009. find

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/papers/lingbulk-2019-lippincott-graph-convolutional-networks-exploring-authorship/