📄 Towards A Sign Language Gloss Representation Of Modern Standard Arabic

Summary

The paper tackles the first half of a text-to-sign-language pipeline for Modern Standard Arabic (MSAr). It converts a written sentence into a gloss, a textual intermediate form that a later stage can render as signs with an animated avatar. The authors note that this task is hard for Arabic because the script has no capital letters, the language is morphologically and grammatically complex, and there is little data. Their approach builds on the glossing rules of Aouiti & Jemni (2018), which were modelled on ASL gloss, and adds more Arabic grammar rules. These hand-written rules work over lemmas and part-of-speech tags. They extract the features the animation stage needs: sentence tense, sentence form (negative, interrogative or conditional) and emphasis words. Verbs are treated as past by default, and verb prefixes or particles switch them to future or present. The paper shows its results only as two example glosses, one for a question and one for a negated sentence, and gives no quantitative evaluation. The authors call the outcome promising but incomplete. The work matters as an early step toward accessibility tools for Arabic-speaking deaf people.

Key points

• The paper motivates the work with WHO-style figures: about 466 million people (over 5% of the world) have disabling hearing loss, and 4 million of them are children.
• Two text representations of signs are mentioned: textual gloss and the symbolic HamNoSys (Hanke, 2004). The paper targets gloss.
• Tense rules work as follows:
  • Verbs are past by default.
  • The prefix sin (sa-) marks future.
  • Certain verb-initial letters mark present.
  • The negation particles la and laysa signal present, lan signals future, and lam signals past.
• Questions are detected by the question mark and labelled by their question word (is, who, where, when, what, how). Conditionals are detected by two Arabic words for "if".
• Known gaps are that pronoun-based gender and number are not handled and named entities (places, people, times) are not included. The work was funded by the AI4D programme.

Interesting ideas

• I1. The rules set a default value (past tense) that is overridden only when a specific morphological or lexical cue appears, which keeps the logic small and every decision traceable to one cue. — For Mova (world): Event-time anchoring could start from a narrative-past default and record which token (auxiliary, modal, time adverb) overrode it. This would make each timeline placement explainable.
• I2. One negation particle can encode polarity and tense at the same time, so both features are read from the same token instead of being computed separately. — For Mova (frames): English auxiliaries such as "didn't" or "won't" could fill a frame's negation and tense slots jointly from one aux node in the UD tree. This avoids conflicts between two independent detectors.
• I3. The intermediate representation is defined by what the downstream consumer (the avatar) strictly needs, not by a full linguistic analysis. — For Mova (back): A gloss-like minimal feature set (predicate lemma, tense, polarity, sentence form, intensifiers) could serve as a compact target for the text → graph → text compressor. It would also give a baseline for checking which features generation really requires.
• I4. Questions are not just flagged but sub-typed by their wh-word, which tells the next stage what kind of answer or gesture is expected. — For Mova (frames): Speech-act detection could attach a question subtype (who/where/when/what/how/yes-no) taken from the wh-word's UD attachment. The answer readers could then use that subtype to pick the entity type to look for.
• I5. Intensity words like "very" and "often" are treated as a separate required feature rather than dropped as modifiers. — For Mova (world): Degree and frequency adverbs could be stored as explicit attributes on states and events in the world model. Answers about "how much" or "how often" would then have a recorded source.