📄 Quotations, Relevance and Time Depth: Medieval Arabic Literature in Grids and Networks

Summary

The authors map how medieval Arabic writers reused one another's texts. They treat reused passages, not explicit quotation markers, as evidence that one author read and drew on another. Their data is CLAUDIA, a diachronic historical corpus of Arabic. They built a matcher for repeated passages that does not depend on word order and accepts some lexical and morphological variation, because Arabic is inflectional, its word order is free, and copyists rarely reproduced text exactly. Each pair of works gets a length-normalized link weight. The works are then placed on a grid with a time axis and joined into a weighted network that the authors call semi-directed. Two case studies show what the network can do. One measures centrality and finds that az-Zayla'i, a writer little known today, was a major hub. The other traces how a single work's influence grows and fades over the centuries. The value of the approach is that it measures how much a work was actually used, which can differ sharply from how highly modern scholars rate it.

Key points

• Corpus: CLAUDIA holds about 2,000 complete works and roughly 420 million words of raw, unannotated text from the 7th to the mid-20th century. Author, date and genre were assigned by hand from reference sources, and dates are based on each author's year of death.
• Matching method:
  • Passages of 7 to 200 tokens are considered, with up to 10% of word types allowed to differ.
  • For each candidate, the round(0.1n)+1 rarest tokens are looked up in a reverse index, and the text around each hit is checked.
  • Overlapping matches are merged so that longer quotations are also found.
• Link weight and visualization: The index Ξ is a log2 score built from the number of quoted tokens, scaled by the lengths of both documents. A parameter h controls how many edges appear in the graph, and the graphs are drawn with Graphviz.
• Centrality results:
  • Vol. 2 of az-Zayla'i's Nasab ar-Raya has the highest degree centrality (0.0958).
  • For Ibn Xaldun, vol. 3 of his History is the hub, while the Muqaddima, much admired today, is isolated in the data.
• Parameters and caveats: Settings were chosen by subjective judgment, with the minimum length set above common stock phrases and eulogies, and were not formally evaluated. The authors say small changes barely affect the case studies. They also stress that a missing intermediate text in the corpus can distort the lineage of a quotation.

Interesting ideas

• I1. Looking up only the rarest few tokens of a window, one more than the number of allowed substitutions, guarantees that at least one exact anchor survives, which makes fuzzy, order-free matching across a huge corpus cheap.
• I2. The minimum match length is deliberately set longer than stock formulas, so that conventional phrases do not create false links.
• I3. Links between nearly contemporary works are left without a direction, and only a clear time gap licenses an arrow, giving a semi-directed graph.
• I4. Multi-volume works are split into separate nodes, which showed that a single volume often carries all the influence.
• I5. Edge weights are normalized by the lengths of both documents and log-scaled, so long texts do not dominate the network simply because they are big.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (learn): Use this to find near-duplicate sentences between silver training data and evaluation sets, so that test contamination is caught even after paraphrase-like edits.
• I2 (answer): When matching a question against a fable or tale, discount evidence that comes only from formulaic openings or set phrases and require overlap beyond that threshold.
• I3 (world): In event ordering, keep "unordered" as a valid state when temporal evidence is weak instead of forcing a before/after relation that cannot be explained.
• I4 (learn): Report evaluation per document section or genre slice rather than one aggregate score, so that localized failures or strengths are visible.
• I5 (world): Apply similar size normalization when weighting co-occurrence or linking edges in the knowledge graph, so that frequent but long sources do not drown out the rest.