📄 Building the Emirati Arabic FrameNet

Summary

The authors aim to start a FrameNet for Emirati Arabic, a spoken Gulf dialect with very few semantic resources. Their target is something like the early Berkeley FrameNet. The base is the Emirati Arabic Corpus (EAC), a spoken-language collection of about three million words. It is transcribed in narrow IPA and carries morpheme, gloss, part-of-speech and translation tiers.

Work runs in two parallel tracks:

• Manual track: native-speaker annotators, trained on Berkeley protocols, assign frames and frame elements to sampled corpus sentences.
• Automatic track: frames are linked to Arabic lexical units by connecting the English FrameNet to an Arabic resource.

The manual track gave the project an immediate start while software for the automatic track was still being built. It also supplies gold data for judging the automatic output.

The paper reports early results. Human raters scored automatically assigned frames on a 5-point scale, and agreement was measured with weighted kappa. Agreement rose clearly when the Arabic script included vowel marks. This suggests vowel information matters for automatic collection.

The resource matters because it would be the first frame-semantic layer over annotated spoken Arabic.

Key points

• EAC: about 3 million words of Emirati speech from radio and TV interviews, daily conversation and some scripted material. It is described as the first annotated corpus of spoken Arabic.
• Planned scope: the first two-year phase (begun 2015) targeted 1000 lexical-unit senses, up to 10,000 annotated sentences and over 100 frames. For comparison, the Berkeley FrameNet has more than 13,000 senses, 190,000 sentences and 1200 frames.
• Annotation format: each frame-element realization is a triple of FE name, grammatical function and phrase type.
• Evaluation of the automatic track: each annotator rated 198 randomly sampled automatic LUs on a 1–5 correctness scale. The sample was split by whether the Arabic script carried vowels.
• Agreement statistic: Cohen's kappa for the manual task, and squared-weight weighted kappa for the ordinal ratings, computed with the irr package in R. The JWKTL library is credited for early technical work.

Interesting ideas

• I1. Build a small hand-made gold core alongside a cheap automatic pipeline, so work begins at once and the gold later grades the automatic output. — For Mova (learn): keep a hand-checked frame set as a fixed referee for silver data from witness parsers and LLM teachers, reporting where each source drifts from it.
• I2. Judge automatic output on an ordered scale from fully right to fully wrong, scored with squared-weight kappa, so a near-miss costs less than a total failure. — For Mova (learn): grade frame or role outputs with partial credit instead of exact-match only, which separates nearly correct predictions from nonsense.
• I3. Split the evaluation sample by whether a disambiguating cue was present (here, vowel marks) to find out what drives the errors. — For Mova (frames): split frame-assignment accuracy by whether the verb lemma is ambiguous or has a clear tag and morphology, to see if mistakes come from earlier stages.
• I4. Record each role filler as frame element plus grammatical function plus phrase type, so semantics stays tied to its syntactic realization. — For Mova (frames): store each role binding together with the UD relation and phrase type that realized it, keeping every frame decision traceable to the tree.
• I5. Reuse English frames as a shared layer across languages, while keeping a manual channel to discover frames specific to one language. — For Mova (back): treat FrameNet frames as the pivot for translation through the meta-world, with explicit slots for frames that have no English counterpart.