📄 The Impact of Morphological Stemming on Arabic Mention Detection and Coreference Resolution

Summary

Arabic attaches articles, prepositions, conjunctions and pronouns directly to words, so one whitespace-delimited token may hold several mentions, such as a noun and a possessive pronoun that point to different entities. The authors build an Entity Detection and Recognition (EDR) system for the ACE 2004 Arabic data that first splits words into prefixes, stems and suffixes.

The segmenter chains three finite-state machines:

• affix-expansion rules that produce a lattice of candidate splits,
• a stem dictionary,
• a Kneser-Ney trigram language model, plus a character-level model for unknown words.

A weaker character n-gram segmenter auto-segments part of Arabic Gigaword, which enlarges the stem vocabulary. The segments then feed two maximum-entropy components: a cascaded mention detector that uses stem, prefix and suffix n-gram features, and a coreference resolver with stem-matching features.

Stem features mostly raise recall in mention detection and add about 2 F-points to coreference. The data partition is clearly documented, and the system ranked among the top ACE 2004 participants. The work shows that subword segmentation is a basic requirement for entity tracking in agglutinative languages.

Key points

• The dictionary-based segmenter was trained on about 0.5M words from Arabic Treebanks 1V2, 2V2 and 3V1 and reaches 97.8% exact word-match segmentation. The character n-gram model alone gets 94.5%, but it handles unseen words better.

• Bootstrapped vocabulary: a stem or affix is kept if it has a frequency above two in the supervised data, or a frequency above ten in the auto-segmented Gigaword portion. This gives roughly 53K entries, about six times the supervised vocabulary.

• Mention detection: stem n-gram features improve first-level classification mainly through recall, with precision nearly unchanged. Full-cascade F-measure moves only from 59.4 to 59.7. Dropping the prefix/suffix n-gram features barely matters.

• Coreference with stem features:

  Mentions	Metric	Base	Base+Stem
  True mentions	ECM-F	77.7	80.0
  True mentions	ACE-Value	86.9	88.2
  System mentions	ECM-F	62.3	64.2
  System mentions	ACE-Value	61.9	63.1

  ACE-Value gains are smaller because that metric weights names heavily and discounts pronouns.

• Affix information acts as a hard constraint: mentions that disagree in gender or number are never chained.

Interesting ideas

• I1. Let a less accurate model that generalizes well to unseen words label a large raw corpus, then admit only frequent items into the precise dictionary model, with a stricter threshold for silver data than for gold.
• I2. Build segmentation as a composition of transducers (affix rules → candidate lattice, dictionary, n-gram scorer) with an explicit, penalized unknown-word path.
• I3. In coreference, match mentions on stems rather than surface strings, and veto links between mentions whose morphological agreement features conflict.
• I4. Score coreference twice, on gold mentions and on system-detected mentions, and with two metrics, explaining any gap by how each metric weights mention types.
• I5. Treat a pronoun buried inside a word as its own mention that must not corefer with its host word.

For Mova

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

• I1 (learn): Silver output agreed by witness parsers could grow lexicons or repair-rule inventories only past a frequency bar set higher than the one for gold treebank material.
• I2 (tokens): Splitting English clitics and contractions ('s, n't, 'll) could keep every candidate in a lattice, so the chosen split stays explainable instead of hard-coded.
• I3 (frames): Coreference could compare lemmas from crate en (company/companies) and use number/gender from morphology as hard filters before any scoring.
• I4 (learn): Evaluations could separate upstream detection errors from linking errors and check whether a gain holds under pronoun-sensitive and name-weighted metrics.
• I5 (world): When building entities, possessive determiners in phrases like "their company" should spawn a separate entity reference that is never merged with the noun it attaches to.