📄 Disfluency and Out-of-vocabulary Word Processing in Arabic Speech Understanding

Summary

The paper addresses spoken Modern Standard Arabic in a narrow-domain dialogue system. Automatic transcripts of spontaneous requests often contain hesitations, repeated or self-corrected phrases, and broken words. They also contain out-of-vocabulary (OOV) tokens: words the recognizer got wrong, words missing from the lexicon, or words only partly recognized. The authors build a robust, partial-analysis pipeline for SARF, a voice server for Tunisian railway information.

The pipeline has three steps:

• A semantically tagged utterance is cut into conceptual segments: illocutionary, referential and filler segments, plus a special Disfluent segment.
• Disfluent segments are repaired with reparandum/editing-term/alteration patterns.
• Words that do not fit the word class expected by their segment are flagged as OOV and reassigned or corrected.

On a Wizard-of-Oz evaluation corpus, the module reduced semantic-frame filling failures from 18.54% to 12.63%. This shows that shallow, meaning-driven repair can make a speech understanding system much more tolerant of noisy input, even for an under-resourced language.

Key points

• In the study corpus, 25.24% of utterances contained disfluencies and 38.22% contained OOV words.

• The evaluation corpus has 2823 spontaneous, automatically transcribed requests (34,726 words): 1291 of type A, 1434 of type D and 98 of type X.

• Detection and correction rates:

  	Correctly detected	Well corrected
  OOV words	78.04%	87.36%
  Disfluencies	61.29%	81.07%

• Overall results were recall 79.23%, precision 74.09% and F-Measure 76.57%. Processing a 12-word utterance took about 0.394 s on average.

• The module is implemented in Java (JBuilder 2007) and outputs semantic frames as XML for the dialogue manager.

• Most errors trace back to wrong conceptual segmentation. Complex disfluencies that mix repetition and self-correction remain unhandled.

Interesting ideas

• I1. A word can be in the lexicon and still be misrecognized: the system catches this when the word's class clashes with the type the slot expects (a month name where a destination city should be). Type expectations thus act as an error detector, not just a filter.
• I2. A repair is written as reparandum, optional editing term, interruption point and alteration, with each pair of words labelled as identical, same-role substitute, or neutral insertion. Only the alteration is kept, and the cleaned segment is then re-classified from scratch.
• I3. Suspected disfluencies are first quarantined in a dedicated Disfluent segment, so later stages handle them explicitly instead of letting them corrupt ordinary analysis.
• I4. A generic Number tag is refined into specific readings such as Fare, Hour or Minute using its position and neighbouring words.
• I5. When an OOV word sits equally close to two candidates, the system declines to guess and hands the ambiguity to the dialogue manager. Detection and correction are also scored separately.

For Mova

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

• I1 (frames): VerbNet selectional restrictions on roles could flag fillers whose type clashes with the role, marking a likely tagging or parse error instead of silently accepting it.
• I2 (tokens): a small explainable pattern layer could strip self-corrections such as "to Boston, uh, no, to Denver" from transcribed or chat-style English before tagging, and log which pattern fired.
• I3 (tree): material that is probably broken could be wrapped as a flagged span for the parser, giving traceable treatment rather than spurious dependency arcs.
• I4 (world): bare numerals could be typed into measures, values or times from local context before entering the world model, which helps maths word problems.
• I5 (answer): readers could abstain or name the competing candidates when evidence ties, and evaluation could report "found the problem" and "fixed it" as separate numbers.