📄 Semi-Automatic Data Annotation, POS Tagging and Mildly Context-Sensitive Disambiguation: the eXtended Revised AraMorph (XRAM)

Summary

Buckwalter's AraMorph (AM) is a widely used Arabic morphological analyzer, but it was built mostly around contemporary newspaper prose. Applied to other text types it fails in two ways: it misses many words, and it adds readings that are implausible for the genre. Lancioni and colleagues argue that an earlier incremental revision (RAM) cannot get past this limit, so they restructure the analyzer as XRAM. It adds three things:

• genre and register flags on lexicon entries, which the user toggles per corpus;
• a light layer of context sensitivity: regular-expression filters drop impossible candidate sequences, and a language model ranks the remaining analyses;
• lexical growth drawn semi-automatically from existing resources, namely a TEI-encoded classical dictionary and Arabic Wikipedia.

The output is TEI P5 XML. The best-ranked reading becomes the default, and the alternatives stay in the file so a human annotator can promote one of them. Unknown-word rates drop sharply for Classical, colloquial and specialized texts. The paper gives few details about the test data.

Key points

• Genre flags: XRAM_CA, XRAM_MSA, XRAM_ICA, XRAM_SPEC_MED/ALCH/GRAM, XRAM_NE, XRAM_FNE and XRAM_CAP. They are implemented as bit flags, so filtering by text type is fast.

• Ranking: a hybrid model combines an order-3 model trained on a manually corrected sample with POS–lemma frequencies from Buckwalter & Parkinson's (2011) frequency dictionary.

• Unknown-word rate, AM → XRAM:

  Genre	AM	XRAM
  Classical Arabic	12.4%	3.4%
  Informal colloquial	18.5%	7.6%
  Medical sublanguage	7.5%	1.3%
  Colloquial aspectual preverbs	23.4%	6.7%
  MSA	2.5%	1.7%

• Accuracy:

  Measure	Precision	Recall	F1	Error rate
  Top-ranked analysis	86.44	95.25	90.63	19.07
  Any analysis in the output	96.43	97.25	96.84	7.57

• Lexicon sources: Salmoné's 1889 Arabic-English dictionary (TEI XML) for Classical vocabulary, Arabic Wikipedia for aligning spellings of transliterated foreign words, and a manually built list for dialect words.

Interesting ideas

• I1. Tag each lexicon entry with register or genre markers and switch whole groups off per corpus, so analyses that are out of place for the text never enter the ambiguity set.
• I2. Before any statistical ranking, run a short list of hard local constraints (e.g., "this preposition is never followed by a verb") that delete impossible candidate sequences; this gives some context sensitivity without a full model.
• I3. Keep every surviving analysis in the output, mark the top-ranked one as the default, and let a reviewer flip the choice with one attribute that an automatic transform then applies.
• I4. Report scores both for the top-ranked answer and for "correct answer anywhere among the candidates". The gap between the two separates ranking errors from coverage errors.
• I5. Measure unknown-word rate separately for each genre, which exposes coverage holes that an aggregate score hides.

For Mova

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

• I1 (tokens): mark archaic, fable-style, technical or slang senses in crate en's lexicon and enable them per input, which cuts spurious readings before tagging.
• I2 (tokens): a small, readable set of pattern-based vetoes over tagger candidates would prune options in an explainable way, with each veto named in the trace.
• I3 (learn): store ranked alternatives alongside silver annotations so that a human or LLM teacher can correct an item by choosing an option rather than re-annotating it.
• I4 (learn): add an oracle-candidate metric to tagger and parser evaluation so that repair work targets either the lexicon or the ranker, depending on which one is failing.
• I5 (answer): track out-of-lexicon rates for fables, school tests and maths word problems separately to see which reader corpora need lexicon growth.