📄 Expanding Arabic Treebank to Speech: Results from Broadcast News

Summary

The paper reports on the completed Arabic Treebank BN corpus. It is the first large Arabic treebank built from transcribed speech rather than written newswire. The source is manually transcribed broadcast news, which is mostly scripted Modern Standard Arabic. It still contains spoken-language phenomena such as restarts, repetitions, fillers and unfinished words, plus a small amount of dialect from interviews.

The authors adapted the Penn Treebank conventions for English speech to Arabic. They added tags for dialect tokens, transcription errors and partial words, and analyses for dialect-specific constructions. They also explain how each token links to the SAMA morphological analyzer. A surprising finding is that speech transcription removes some ambiguity found in written text, so annotation is more consistent in places. An initial Bikel-parser experiment shows that BN parses only slightly worse than newswire. The gap shrinks further when disfluent EDITED material is removed. The resource matters because it gives a baseline and a set of guidelines for moving treebanking toward informal and dialectal Arabic.

Key points

• The corpus has 120 files of broadcast news with 432,976 source tokens and 517,080 tree tokens. It was released through LDC in 2012.
• 4,941 tree tokens (under 1%) receive a single catch-all DIALECT POS tag regardless of function. Other special tags are TRANSERR (3,001 source tokens) and PARTIAL (4,765).
• SAMA status flags: 96.1% of BN source tokens exactly match a SAMA solution, compared with 84.6% for ATB3. The excluded class differs by genre: dialect, errors and partial words in BN, but punctuation and digits in ATB3.
• Initial hamza is audible in speech, so transcribers already resolve the <in~a / >an~a distinction. The orthographically neutralized form, about 1.5% of ATB3 tokens, essentially does not occur in BN.
• The Bikel parser was tested on sentences of up to 40 words. Scores with predicted tags / gold tags were:
  • ATB3: 78.2 / 79.6
  • BN: 76.1 / 77.8
  • BN with EDITED subtrees removed: 77.2 / 78.9

Interesting ideas

• I1. A misspelled or mistranscribed token keeps its surface form and gets an error tag, but the tree is built as if the word were correct, so syntax stays clean despite noisy input.
• I2. Disfluencies get dedicated structure: EDITED for material repaired by later speech, -UNF for unfinished constituents, INTJ for filled pauses. This lets the core sentence be recovered by deleting those nodes.
• I3. A frequent discourse filler meaning roughly "you know" is always annotated as a parenthetical, whatever the surrounding sentence looks like. Discourse particles are thus kept out of the main argument structure.
• I4. Every token carries a status flag recording how its annotation relates to the morphological analyzer: exact match, reduced solution, pending manual solution, or deliberately excluded. This makes lexicon coverage auditable.
• I5. Parsing is run both with predicted and with forced gold tags, and again with EDITED subtrees stripped. This separates tagging errors and speech-specific structure from core parsing difficulty.

For Mova

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

• I1 (tokens): The tagger could flag typo tokens with an explicit error marker while handing the parser the intended category, keeping the tree explainable instead of mis-parsed.
• I2 (tree): Parsing of dialogue in fables or transcripts could mark repaired fragments as EDITED, so later stages can drop them before building frames.
• I3 (frames): Fillers and hedges like "you know" or "I mean" could be treated as parentheticals, so they never become arguments in who-did-what frames but can still feed speech-act or stance cues.
• I4 (tokens): The morphology crate could record a per-token provenance status, showing which lemmas came from the lexicon, which are guesses awaiting review, and which were deliberately skipped.
• I5 (learn): Evaluations could include similar ablations, such as gold versus predicted tags and with versus without disfluent material, to pinpoint which stage causes an accuracy drop.