Elamur Research⌕
☰
home / papers / 2012 / Predicting Phrase Breaks in Classical and Modern Standard Arabic Text

Predicting Phrase Breaks in Classical and Modern Standard Arabic Text

Summary

The paper addresses automatic phrase break prediction for Arabic, a step in text-to-speech that decides where a speaker would pause. The authors use the Qur'an as a gold-standard boundary-annotated corpus, because its traditional recitation mark-up (Tajwīd, the system of stops and starts) already encodes graded pauses. Their dataset has 77,430 words, PoS tags from the Qur'anic Arabic Corpus (QAC), and boundaries reduced from eight Tajwīd degrees to two schemes: break/non-break, and major/minor/none. They train two NLTK sequence taggers, a trigram tagger and an HMM tagger. Each uses either a coarse three-way PoS set (noun, verb, particle) or QAC's ten categories, with or without the word itself as a feature. Non-breaks dominate the data, so a trivial baseline already reaches 85.56%. The trigram tagger beats this baseline in accuracy. The HMM tagger has lower accuracy but recovers more breaks, which gives it a higher Balanced Classification Rate (BCR). The work is groundwork for chunking tools for Modern Standard Arabic (MSA), which will require restoring short vowels first.

Key points

  • Corpus: 77,430 words in 8230 sentences, with PoS tags from QAC v2.0 (Dukes, 2010).
    • Training: 70,112 words, 7381 sentences.
    • Test: 7318 words, 849 sentences, containing 1057 breaks against 6261 non-breaks.
  • The test set is not random. It consists of the 21 chapters whose Meccan/Medinan origin is disputed, so it mixes both styles.
  • Two-class results:
    • The trigram tagger with (word, PoS) input scored 88.47% (3 tags) and 88.44% (10 tags), against the 85.56% baseline.
    • The HMM scored about 82.6% accuracy but reached BCR 0.72, versus 0.67 for the trigram tagger.
  • Three-class results: the trigram tagger reached 88.69% (3 tags) and 88.62% (10 tags).
  • Dropping the word feature often collapsed the taggers to the majority-class baseline. The coarse traditional {noun, verb, particle} tagset matched or beat the ten-category set.

Interesting ideas

  • I1. Report BCR (the mean of per-class correct rates) next to accuracy, so that a majority-class predictor cannot look good on skewed labels.
  • I2. Pick the held-out set to deliberately cover a stylistic split, here the chapters of disputed origin, so the test is fair to every style in the training data.
  • I3. A very coarse, linguistically traditional category set can carry as much signal as a finer one for a structural task.
  • I4. The study contrasts a greedy left-to-right tagger, which cannot revise earlier decisions, with a whole-sequence HMM. The greedy tagger was more accurate, while the sequence model caught more minority-class items.
  • I5. Fine graded labels (eight pause strengths) are collapsed into nested coarser schemes (2 and 3 classes), so the same data supports tasks at several granularities.

For Mova

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

  • I1 (learn): Use BCR as a standard negative control when evaluating rare-label decisions, such as rare dependency relations or rare frame roles.
  • I2 (learn): Build evaluation splits for tales and school texts that mix genres on purpose, instead of sampling at random.
  • I3 (tokens): Keep a coarse tag tier beside the full UPOS/XPOS tags and test whether downstream decisions really need the finer distinctions.
  • I4 (tree): Measure greedy and global decoding separately on precision versus recall of rare arcs, rather than on overall score alone.
  • I5 (back): When generating text from the meaning graph, model clause and phrase boundaries as graded break strengths that can be mapped to punctuation choices.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

Fields: Linguistics & languages

Topics Semitic languages

 Institutions

🎓University of LeedsLeeds · United Kingdom19 works · 1,290 citations🎓University of JordanAmman · Jordan4 works · 65 citations

Cite

Cite as Majdi Sawalha, Claire Brierley and Eric Atwell. 2012. Predicting Phrase Breaks in Classical and Modern Standard Arabic Text. Proceedings of the Eighth International Conference on Language Resources and Evaluation (LREC'12).

The paper itself, not this page. BibTeX from the library's record of the work.

@inproceedings{sawalha2012predicting,
  title = {{Predicting Phrase Breaks in Classical and Modern Standard Arabic Text}},
  author = {Majdi Sawalha, Claire Brierley and Eric Atwell},
  booktitle = {Proceedings of the Eighth International Conference on Language Resources and Evaluation (LREC'12)},
  year = {2012},
  doi = {10.63317/3u3bivt7k63h},
  url = {https://doi.org/10.63317/3u3bivt7k63h}
}

References

15 references, 1 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Atwell, E., Brierley, C., Dukes, K., Sawalha, M. and Sharaf, A.M. 2011. „An Artificial Intelligence Approach to Arabic and Islamic Content on the Internet.‟ National Information Technology Symposium (NITS). Riyadh, Saudi Arabia. find
  2. Bird, S., Klein, E. and Loper, E. 2009. Natural Language Processing with Python. Sebastopol, CA. O‟Reilly Media, Inc. find
  3. Brierley, C. 2011. ‘Prosody Resources and Symbolic Prosodic Feartures for Automated Phrase Break Prediction.’ PhD Thesis. School of Computing. University of Leeds. find
  4. Brierley, C. and Atwell, E. 2010. „ProPOSEC: a Prosody and PoS Spoken English Corpus.‟ In Proceedings of LREC 2010: Language Resources and Evaluation Conference. Valetta, Malta. May 2010. find
  5. Brierley, C., Sawalha, M. and Atwell, E. 2012. „Open-Source Boundary-Annotated Corpus for Arabic Speech and Language Processing.‟ In Proceedings of LREC 2012: Language Resources and Evaluation Conference. Istanbul, Turkey. May 2012. find
  6. Brierley, C., Sawalha, M. and Atwell, E. 2011. „Arabic Phonetics and Phonology for Text Analytics and Natural Language Processing Applications.‟ PowerPoint presentation for Arabic Phonetics and Phonology PG Workshop. York. find
  7. Denny, F.M. 1989. „Qur‟an Recitation: A Tradition of Oral Performance and Transmission.‟ In Oral Tradition. 4/1-2: 5-26 find
  8. Dukes, K. 2010. The Quranic Arabic Corpus (v. 2.0). Online. Accessed: August 2011. http://corpus.quran.com ↗
  9. Dukes, K. and Habash, N. 2010. „Morphological Annotation of Qur‟anic Arabic.‟ In Proceedings of LREC 2010: Language Resources and Evaluation Conference. Valletta, Malta. find
  10. Ingulfsen, T., Burrows, T. and Buchholz, S. 2005. „Influence of Syntax on Prosodic Boundary Prediction.‟ In Proceedings, INTERSPEECH 2005. 1817-1820. find
  11. Liberman, M.Y. and Church, K.W. 1992. „Text Analysis and Word Pronunciation in Text-to-Speech Synthesis.‟ In Advances in Speech Signal Processing. Furui S. and Sondhi, M.M. (eds.). New York. Marcel Dekker Inc. find
  12. Sawalha, M. 2011. ‘Open-Source Resources and Standards for Arabic Word Structure Analysis: Fine Grained Morphological Analysis of Arabic Text Corpora.’ PhD. Thesis. School of Computing. University of Leeds. find
  13. Sawalha, M. and Atwell, E. 2010. „Fine-Grain Morphological Analyzer and Part-of-Speech Tagger for Arabic Text.‟ In Proceedings of LREC'10: Language Resources and Evaluation Conference, Valetta, Malta. May 2010. our card
  14. Sharaf, A.M. 2011. „Macci and Madani Shurahs.‟ Online. Accessed: October 2011. http://www.textminingthequran.com/wiki/Makki_and_Madani_Surahs ↗
  15. Taylor, P. and Black, A.W. 1998. „Assigning Phrase-Breaks from Part-of-Speech Sequences.‟ In Computer Speech and Language. 12.2: 99-117. find

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/sawalha-2012-predicting-phrase-breaks-classical-modern/