📄 CODACT: Towards Identifying Orthographic Variants in Dialectal Arabic

Summary

Dialectal Arabic (DA) is now widely written online, but it has no spelling standard. The same word can appear in many forms, and this makes data sparse for NLP tools built for Modern Standard Arabic (MSA). The authors present CODACT, an initial system that detects spelling variants in Arabic-script DA. They treat the task as unsupervised clustering. Focal word types are placed in a vector space, and cell values come from three similarity measures:

• plain Levenshtein distance (LEDM),
• a phonologically biased edit distance (BEDM), which counts letters with similar sounds as matches,
• contextual string similarity (CSS).

Clusters are built with CLUTO repeated bisection using cosine similarity. Results are scored against gold-corrected Egyptian and Levantine data. The best setup, LEDM combined with BEDM over a large feature space, reaches an entropy of about 0.19, with 68% (EGY) and 64% (LEV) cluster precision. The baseline scores 0.47/0.51 entropy and 29% precision. The authors present this as the first attempt at this problem for DA and as a first step toward normalizing DA to the CODA convention.

Key points

• Applying MSA tokenizers to DA gives about 88% accuracy, compared with 99.2% on MSA. The authors attribute the gap mainly to differences between the language varieties, not only to genre.
• Data: the COMMENTDA set plus an augmented collection, RCorpora, with about 6M strings for each dialect. There are 729 focal types for EGY and 717 for LEV.
• Two matrix designs were compared:
  • NxN: focal words against the same focal words.
  • NxM: focal words against all corpus types, with roughly 205K–434K dimensions.
  • NxM was better for every metric in both dialects.
• BEDM consistently beat LEDM, and LEDM+BEDM beat both on precision, recall and entropy. CSS was the weakest measure and generally added noise, which the authors attribute to sparse surface forms in a morphologically rich language.
• Doubling the data helped only slightly, and mostly for CSS. Planned future work includes lemma-based co-occurrence and normalization to CODA.

Interesting ideas

• I1. Sound-class edit distance. Letters whose pronunciations can coincide form equivalence classes, and swapping within a class costs nothing. This captures variants that plain edit distance pushes too far apart.
• I2. Profile-based representation. Each candidate word is described by its similarity to every type in the corpus, not just to the other candidates. Comparing these profiles with cosine gives a more robust signal than a direct pairwise distance.
• I3. Ablating every similarity signal and every combination. The authors tested each measure alone and in all mixes. A weak signal (contextual similarity on small data) turned out to hurt a strong combination, so more evidence is not automatically better.
• I4. Scoring clusters against gold groupings with a random baseline. Clusters are evaluated with entropy and cluster precision/recall, and the random clustering serves as a floor.
• I5. Identify first, normalize later. Variant detection is separated from choosing a canonical form, and spelling errors are treated as a subset of legitimate variation.

For Mova

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

• I1 (tokens): A similar bias for English informal or phonetic spellings could help map unknown words to known lemmas before tagging.
• I2 (tokens): Clustering unknown or variant word forms by their similarity profiles could propose lemma merges that remain explainable.
• I3 (learn): When combining witness parsers for silver data, ablate each witness and drop any that lowers agreement quality.
• I4 (learn): This gives a ready negative-control setup for any induced grouping, such as verb classes or repair-rule families.
• I5 (tokens): Keeping a variant-cluster step separate from normalization would let each decision be inspected and explained on its own.