Elamur Research⌕
☰
home / papers / 2012 / Rescoring N-Best Hypotheses for Arabic Speech Recognition: A Syntax- Mining Approach

Rescoring N-Best Hypotheses for Arabic Speech Recognition: A Syntax- Mining Approach

Paper · 2012 · Dia AbuZeina, Moustafa Elshafei, Husni Al-Muhtaseb et al.

Summary

The paper tests whether part-of-speech patterns can fix syntactically wrong output from an Arabic speech recognizer. The KFUPM baseline is built on CMU PocketSphinx and trained on 7.57 hours of MSA broadcast news. Its transcription corpus was tagged with the Stanford Arabic tagger. WEKA-3-6-5 then mined 3000 association-style rules over tag positions, and these rules may link non-adjacent words. To rescore a hypothesis, the system tags it and counts how many mined rules it satisfies. The candidate with the most matches wins, and ties go to the one with the better acoustic rank. Starting from a baseline WER of 16.04%, the rescoring helped some test files but hurt others, with no significant overall gain. The authors give two causes. First, tagger errors pollute the mined rules. Second, diacritized N-best lists are filled with near-duplicates that differ only in short vowels. Their takeaway is a proposed language-model add-on: mined long-distance word rules that complement n-grams. The study matters mainly as an honest negative result with a clear diagnosis.

Key points

  • The baseline has a dictionary of 17,236 vocabulary items (28,682 words and variants), 6146 audio files and a 40-phoneme set. It reaches 16.04% WER on 400 test utterances (about 0.57 h, 3585 diacritized words).
  • The Stanford Arabic tagger uses 29 tags, and its reported accuracy ranges from 80% to 96%.
  • WEKA extracted the top 3000 rules over absolute tag positions in sentences of up to 30 words. The best rules have accuracy around 0.956, for example: a CD at position 4 and a DTNN at position 6 imply an IN at position 5.
  • In one 300-best list, only 16 hypotheses were actually distinct at the word level. Many entries differed only in diacritics, which pushed the search away from the high-scoring candidates.
  • The future work the authors recommend has three parts:
    • bringing linguistic knowledge into the decoder itself;
    • better tagging of diacritized Arabic;
    • word-level long-distance rules as a second LM module.

Interesting ideas

  • I1. Rescore candidates by how many mined structural rules they satisfy, and fall back to the original model's ranking on ties; every choice can then be justified by the list of matched rules.
  • I2. Before rescoring, check how diverse the candidate list really is. Many top candidates may differ only along a dimension the rescorer cannot see, which wastes the N-best budget.
  • I3. Rules mined from automatically tagged text inherit the tagger's mistakes, and the authors trace their failure to exactly this upstream noise.
  • I4. Association rules with gaps can capture constraints between words far apart in a sentence, which n-grams cannot reach. Such rules could serve as a complementary scoring module rather than a replacement.

For Mova

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

  • I1 (tree): rerank gparse/perceptron k-best parses by counting matched treebank-mined dependency patterns, logging which patterns tipped the decision.
  • I2 (learn): when witness parsers or LLM teachers propose alternatives, collapse candidates that differ only in irrelevant details before voting, so agreement statistics are not inflated.
  • I3 (learn): mine repair or reranking rules only from gold or multi-witness-agreed silver data, and test them against negative controls before trusting them.
  • I4 (tokens): use mined long-distance context rules, ideally keyed on relative rather than absolute positions, as explainable tie-breakers for the tagger's ambiguous lemma/POS decisions.

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

📚 Shelves 4 Computer Science · 9,086 works Artificial Intelligence · 8,889 works Arts and Humanities · 9,296 works Language and Linguistics · 9,206 works

🧭 Topics 3 Semitic languages · 1,484 works Speech & spoken language · 233 works Syntax & grammar formalisms · 161 works

Cite

Cite as Dia AbuZeina, Moustafa Elshafei, Husni Al-Muhtaseb and Wasfi Al-Khatib. 2012. Rescoring N-Best Hypotheses for Arabic Speech Recognition: A Syntax- Mining Approach. Fourth Workshop on Computational Approaches to Arabic-Script-based Languages.

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

@inproceedings{abuzeina2012rescoring,
  title = {{Rescoring N-Best Hypotheses for Arabic Speech Recognition: A Syntax- Mining Approach}},
  author = {Dia AbuZeina, Moustafa Elshafei, Husni Al-Muhtaseb and Wasfi Al-Khatib},
  booktitle = {Fourth Workshop on Computational Approaches to Arabic-Script-based Languages},
  year = {2012},
  url = {https://aclanthology.org/2012.amta-caas14.8/}
}

References

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

  1. Bing Xiang, Bowen Zhou and Martin Cmejrek. 2009. Advances in syntax-based Malay-English speech translation. Proceedings of the 2009 IEEE International Conference on Acoustics, Speech and Signal Processing, IEEE Computer Society: 4801-4804. find
  2. Dan Jurafsky and Martin J. 2009. Speech and Language Processing, second edition, Pearson. find
  3. Fatma Al-Shamsi and Ahmed Guessoum. 2006. A Hidden Markov Model –Based PS Tagger for Arabic, CiteSeerX. find
  4. Guihong Cao, Jian-Yun Nie, and Jing Bai. 2005. Integrating word relationships into language models. Proceedings of the 28th annual international ACM SIGIR conference on Research and development in information retrieval. Salvador, Brazil, ACM: 298-305. find
  5. High Performance Computing (HPC) Center, 2011. http://hpc.kfupm.edu.sa/Home.htm ↗
  6. Luis R. Salgado-Garza, Richard M. Stern and Juan A. Nolazco F. 2004. N-Best List Rescoring Using Syntactic Trigrams ,MICAI 2004: Advances in Artificial Intelligence. find
  7. Machine Learning Group at University of Waikato, 2011.http://www.cs.waikato.ac.nz/ml/WEKA/ ↗
  8. Mansour Alghamdi , Moustafa Elshafei , and Husni Almuhtasib . 2009. Arabic broadcast news transcription system, International journal of speech and technology, 10: 183–195 find
  9. Mansour Alghamdi , Husni Almuhtasib, and Moustafa Elshafei . 2004. Arabic Phonological Rules, King Saud University Journal: Computer Sciences and Information. Vol. 16, pp. 1-25 find
  10. Maria Ruiz-casado , Enrique Alfonseca , and Pablo Castells. 2007. Automatising the learning of lexical patterns: an application to the enrichment of WordNet by extracting semantic relationships from Wikipedia. Data Knowledge and Engineering, 61 3 , pp. 484–499. ↗
  11. Mohammed Albared, Nazlia Omar, Mohd.Aziz, and Mohd Ahmad Nazri. 2010. Automatic part of speec tagging for Arabic: an experiment using Bigram hidden Markov model, RSKT'10 Proceedings of the 5th international conference on Rough set and knowledge technology Mohamed Ali, Moustafa Elshafei, Mansour Alghamdi, Husni Almuhtaseb and Atef Alnajjar, 2009. Arabic Phonetic Dictionaries for Speech Recognition. Journal of Information Technology Research, Volume 2, Issue 4, pp. 67-80. find
  12. Mona Diab, Kadri Hacioglu, and Daniel Jurafsky. 2004. Automatic tagging of Arabic text: from raw text to base phrase chunks, 5th Meeting of the North American Chapter of the Association for Computational Linguistics/Human Language Technologies Conference. our card
  13. Moustafa Elshafei. 1991. Toward an Arabic Text-to-Speech System, The Arabian Journal of Science and Engineering, Vol. 16, No. 4B, pp.565-583. find
  14. Moustafa Elshafei , Husni Almuhtasib and Mansour Alghamdi. 2002. Techniques for High Quality Text-to-speech, Information Science, 140 (3-4) 255-267. find
  15. Moustafa Elshafei., Husni Al-Muhtaseb, and Mansour Alghamdi. 2006 . Machine generation of Arabic diacritical marks. In Proceedings of the 2006 international conference on machine learning; models, technologies, and applications (MLMTA’06), USA. find
  16. René Beutler. 2007. Improving Speech Recognition through Linguistic Knowledge, Doctoral Dissertation ,ETH Zurich. ↗
  17. Stanford Log-linear Part-Of-Speech Tagger, 2011. http://nlp.stanford.edu/software/tagger.shtml ↗
  18. Tobias Scheffer . 2005. Finding association rules that trade support optimally against confidence. Intell. Data Anal. 9(4): 381-3 ↗
  19. Wen Wang, Yang Liu, and Mary P. Harper. 2002. Rescoring effectiveness of language models using different levels of knowledge and their integration. Acoustics, Speech, and Signal Processing (ICASSP), 2002 IEEE International Conference on. find
  20. Yahya El Hadj, Imad Abdulrahman Al-Sughayeir and Abdullah Mahdi Al-Ansari. 2009. Arabic Part-Of-Speech Tagging using the Sentence Structure, Proceedings of the Second International Conference on Arabic Language Resources and Tools, The MEDAR Consortium. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/abuzeina-2012-rescoring-n-best-hypotheses-arabic/