📄 QurAna: Corpus of the Quran annotated with Pronominal Anaphora

Summary

The paper presents QurAna, a resource in which pronouns in the original Arabic Quran are linked to the entities they refer to. The authors argue that pronouns in this Classical Arabic text are dense and often hard to resolve. Arabic verbs carry obligatory pronominal affixes, and referents may sit far back in the text or never be stated at all. They manually tagged over 24,500 pronouns. Where a reference was ambiguous, they settled it using the classical commentary of Ibn Kathir. Rather than storing antecedents only as text spans, they also collected them into a list of ontological concepts, so recurring referents such as "Children of Israel" or "mankind" become reusable nodes. The corpus can be browsed online by verse or by concept. As a demonstration, the authors measure verse relatedness with a root-based vector space model. Enriching verses with concept information roughly doubled the number of known related pairs that cross a similarity threshold. QurAna is offered as the first anaphora corpus for Classical Arabic, useful for building and evaluating resolution systems.

Key points

• Over 24,500 pronouns are annotated with antecedents across the Quran: 114 chapters, 6,243 verses, more than 77,000 words.
• Antecedents are normalized into an ontological concept list. A web tool shows, for any verse, its pronouns, their referents and its concepts, plus concordance lines for each concept.
• In one example verse (24:31), an English translation keeps only 16 of the 27 Arabic pronouns. Many second-person pronouns refer to Prophet Muhammad without any earlier mention of him.
• Verse similarity setup:
  • Verses are represented as tf-idf vectors over 758 roots, filtered from 1,226 unique roots, and compared by cosine.
  • The test set is 7,679 scholar-derived related verse pairs (from QurSim).
  • With plain lexical vectors, only 428 pairs (6%) scored above 0.5.
  • With concept-augmented vectors, 869 pairs did.
• Two challenges stood out:
  • Some antecedents lie very far from their pronouns, as much as 33 verses away in Chapter 2.
  • The scheme forbids split or multiple antecedents, so some antecedents became compound concepts.

Interesting ideas

• I1. Resolving each pronoun to a shared concept rather than only to a text span turns coreference output into an index of entities that recur across a whole document collection.
• I2. The authors expand a text unit's features with material from all other units that mention the same resolved referent. This trick can make it easier to find related passages when they share few words.
• I3. Ambiguous annotation decisions are anchored to one named authoritative external source, which keeps adjudication consistent and traceable.
• I4. Comparing pronoun counts between a source text and its translation shows how much referential structure translation silently drops or restores. Aligned translations can then inherit gold antecedents from the source.
• I5. The authors explicitly record that their scheme cannot represent split or plural antecedents, and that this forced compound concepts. This is a useful warning when designing a coreference format.

For Mova

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

• I1 (world): Coreference results could be lifted into persistent entity nodes in the world of the text. Then queries like "every place this character acts" become simple lookups.
• I2 (answer): Fable/tale QA could gather sentences linked by a resolved entity before matching a question, instead of relying on shared words alone.
• I3 (learn): When silver data from witness parsers disagree on a referent, a fixed, cited tiebreaker source keeps repairs explainable rather than ad hoc.
• I4 (back): Translation through the meta-world could be checked by verifying that each target pronoun maps back to the same entity as in the source graph.
• I5 (frames): The coref layer should allow set-valued antecedents (e.g., "they" = A + B) from the start, so groups do not collapse into artificial merged entities.