📄 Multi-Level Analysis and Annotation of Arabic Corpora for Text-to-Sign Language MT

Summary

Arabic has syntactic treebanks such as PATB and QADT, but semantic annotation of Arabic has received much less attention. The authors present a semi-automatic tool, still in development, that annotates Modern Standard Arabic text at three levels: morphology, syntax and frame semantics. The morphological level uses the BAMA analyzer and adds lexical semantics from Arabic WordNet. The syntactic level is a dependency analysis with fine-grained Arabic-specific relations, plus FrameNet-style constituent annotation. The semantic level reuses frames and frame elements from Berkeley FrameNet.

The first corpus comes from Algerian primary-school textbooks. These texts are the development data for the authors' system that translates Arabic into Algerian Sign Language for deaf children. In a pilot, the team annotated more than 1,000 sentences about Motion events. The XML output follows the FrameNet format, and the authors compile it into rules that link syntactic patterns to semantic roles.

The paper reports no quantitative evaluation. Its value is in showing that English frames can carry over to Arabic, and that annotations can be turned directly into form-to-meaning rules for knowledge-based MT.

Key points

• The external resources are the BAMA lexicon, Berkeley FrameNet and Arabic WordNet (AWN). AWN is mapped to Princeton WordNet 2.0 and SUMO and is used to widen Arabic FrameNet coverage.
• The annotation is stored in named layers ("FE", "GF", "Sumo", etc.), grouped by Lexical Unit. Layers can be exported selectively and filled in incrementally.
• Sentences for each target word are collected FrameNet-style and split into sub-corpora by syntactic pattern. Planned sources include Quranic texts and the CCA Corpus of Contemporary Arabic.
• In the pilot, over 1,000 Motion-event sentences were annotated. One sample rule comes from the Placing frame:
  • the pattern is a VP with a nominative NP, an accusative NP and a PP;
  • the roles map as Agent/Subj, Theme/Obj and Source/OBL.
• Arabic properties highlighted as relevant:
  • pro-drop subjects;
  • object pronouns attached to the verb;
  • case diacritics;
  • argument semantics needed to tell apart the two accusative objects of ditransitives.

Interesting ideas

• I1. Annotated sentences are compiled into valence rules: a phrase-type pattern plus voice determines each argument's frame element, phrase type and grammatical function. This turns a corpus into an inspectable lexicon.
• I2. Each lexical unit gets one file with separate, independently fillable layers. This allows partial, incremental annotation without redoing earlier work.
• I3. A rule-based morphological analyzer's output is enriched at the token level with WordNet senses. This closes the gap where morphology alone carries no meaning.
• I4. The tool adds label subtypes beyond standard dependency relations, such as cognate accusative and accusative of cause, because finer function labels make argument classification easier.
• I5. Argument semantics is used to decide which accusative object fills which role in double-object constructions, where syntax alone is ambiguous.

For Mova

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

• I1 (frames): induce per-verb rules from annotated examples that map UD argument configurations to VerbNet roles, so every role choice can be traced to a stated rule.
• I2 (learn): store silver data per verb lemma, with each witness parser or teacher tier kept as its own layer, so agreement and disagreement can be checked layer by layer.
• I3 (tokens): attach candidate WordNet synsets to lemmas in crate en, so later stages receive sense options instead of bare strings.
• I4 (tree): keep or derive UD subtypes (e.g., cause or manner obliques) during dialect conversion, giving the role mapper sharper input.
• I5 (frames): apply selectional preferences from the WordNet ontology to separate Recipient from Theme in English ditransitives, recording the semantic test as the explanation.