📄 Arabic Cross-Document Coreference Resolution

Summary

The paper addresses cross-document coreference for Arabic: deciding when entity mentions in different documents (e.g., two people both called "George Bush") denote the same real-world individual. The setting is resource-poor, with no annotated Arabic ground truth for the task. Within-document entities come from BBN's Serif resolver. Each system emits pairwise links between entities from different documents, and a union-find procedure groups these links into cross-document entities. Three systems are compared:

• an exact-match baseline on each entity's longest mention strings;
• a translingual projection system that machine-translates Arabic names into English and matches them there, using exact matches or Wikipedia-derived alias pairs;
• a context-similarity system that compares 50-token windows around mentions, computed at scale with MapReduce.

Both experimental systems are unioned with the baseline. On the NIST ACE 2008 Arabic evaluation, translingual projection beats the baseline on every measure. Context similarity gives no real gain. The work suggests that pivoting through a better-resourced language is a cheap, effective route for low-resource coreference.

Key points

• Data. The ACE 2008 evaluation covered about 10,000 Arabic documents across six genres, and NIST scored 412 of them. ACE 2005 Arabic (1,245 documents) served for development, and 12,000 LDC Arabic Gigaword documents were processed through Serif.
• Results.
  • Baseline: B-Cubed P/R/F of 37.5/44.1/40.6 and an ACE value of 19.2.
  • Translingual projection (TrnsProj): 38.4/44.8/41.3 and an ACE value of 21.2.
• Context-similarity threshold sweep. A threshold of 0.2 drops performance to an F of 36.4 and an ACE value of 15.9. Thresholds of 0.4956 and above essentially reproduce the baseline (ACE value 19.3/19.2).
• Threshold choice. The 0.4956 threshold was picked by a native Arabic speaker inspecting link scores on Gigaword, since no ground truth existed.
• Similarity computation. It follows Elsayed et al.'s MapReduce method: an inverted index, then summing per-term contributions to each pair. A 99.9% df-cut drops the most frequent terms.

Interesting ideas

• I1. Instead of fixing name variants (misspellings, titles, aliases) in the source language, the authors map names into a richer canonical space where variants collapse and alias resources already exist.
• I2. Coreference decisions are kept as simple pairwise links, and entity clusters are built afterwards with union-find, so each merge is traceable to one specific link.
• I3. Every experimental technique is evaluated as "cheap baseline ∪ new links," which isolates exactly what the added method contributes on top of a method that always runs.
• I4. Using only the longest mention in each within-document chain as the entity's representative name is a cheap heuristic that avoids matching on pronouns or short fragments.
• I5. The context-window similarity failed to help and even hurt at loose thresholds, a useful negative result about topical bag-of-words evidence for identity decisions.

For Mova

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

• I1 (frames): Coreference could map mentions to canonical forms such as WordNet/Wikidata identifiers or normalized aliases before matching, rather than relying on raw surface strings.
• I2 (world): Entity identity in the world of a text could be stored as explicit link evidence plus a union-find closure, so any merged entity can explain which links joined it.
• I3 (learn): New repair rules or witness parsers could be scored as additions to a fixed baseline, making their marginal gain or harm directly visible.
• I4 (frames): Cross-paragraph or cross-story entity linking could compare chains via their longest named mention rather than every mention.
• I5 (learn): A threshold sweep of this kind could serve as a negative control showing that structural evidence, not lexical overlap, drives coreference gains.