--- title: "Collection and Evaluation of Broadcast News Data for Arabic" description: "The paper describes a fast way to build Arabic broadcast news speech data and a first recognizer from material found on the web. 4 references." authors: ["Mohamed Afify", "Ossama Emam"] year: 2004 venue: "Proceedings of the Fourth International Conference on Language Resources and Evaluation (LREC'04)" source: "https://aclanthology.org/L04-1170/" licence: "CC BY-NC 4.0 (ELRA, LREC proceedings)" licence_url: "https://creativecommons.org/licenses/by-nc/4.0/" summary_licence: "Elamur Cosmic License" canonical: "https://research.elamur.ai/free/en/papers/afify-2004-collection-evaluation-broadcast-news-data/" --- [Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/afify-2004-collection-evaluation-broadcast-news-data/) [home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2004](https://md.research.elamur.ai/free/en/papers/2004/page/2/) / Collection and Evaluation of Broadcast News Data for Arabic # Collection and Evaluation of Broadcast News Data for Arabic **Authors:** Mohamed Afify and [Ossama Emam](https://md.research.elamur.ai/free/en/authors/ossama-emam/) · **Year:** 2004 · **Venue:** [Proceedings of the Fourth International Conference on Language Resources and Evaluation (LREC'04)](https://md.research.elamur.ai/free/en/venues/lrec/) ([LREC 2004](https://md.research.elamur.ai/free/en/venues/lrec/2004/)) · **Publisher:** [European Language Resources Association (ELRA)](https://md.research.elamur.ai/free/en/publishers/european-language-resources-association-elra/) · **Licence:** [CC BY-NC 4.0 (ELRA, LREC proceedings)](https://creativecommons.org/licenses/by-nc/4.0/) **Links:** [Publisher's page](https://aclanthology.org/L04-1170/) · [PDF](https://aclanthology.org/L04-1170.pdf) · [OpenAlex](https://openalex.org/W2396463464) **Citations:** 0 citations (OpenAlex, 2026-10-14) ## Summary The paper describes a fast way to build Arabic broadcast news speech data and a first recognizer from material found on the web. The authors downloaded about 66 hour-long conversational programmes with their transcripts. These recordings are too long for standard training tools, and cutting them by hand is slow. So only about 10 hours were segmented manually, and a bootstrapped model was then used to cut the remaining audio automatically. Standard Arabic writing omits short vowels, which a phonetic lexicon needs. To fill them in, the authors combined dictionary lookup on stems with a statistical vowel predictor. Adapting an existing telephony model by retraining on the 10 manual hours cut the error rate from about 60% to about 36%. Adding automatically segmented data brought almost no further gain. The work matters as a practical recipe for under-resourced languages. It shows that a small hand-labelled seed goes a long way, and it is candid about where automatic bootstrapping falls short. ## Key points - The corpus has 66 programmes (about 66 hours) of lossy-compressed 8 kHz audio and roughly 300K words of transcripts. Hand-segmenting 10 hours by speaker turn took one transcriber over a month. A development set of about 15 minutes (around 3K words) was held out. - The language model is a trigram over a 50K-word vocabulary with no morphological splitting, reaching a perplexity of about 500. Adding other text corpora did not help. - The lexicon uses 38 phonemes. Short vowels are restored in steps: - first, prefix stripping plus lookup of vowelized stems; - if that fails, a maximum-likelihood model, argmax P(V|L)P(V), with a vowel trigram, a ±2-letter context window and A* search producing N-best candidates; - finally, a manual pass to fix gross errors. - Acoustic features are 39-dimensional (13 MFCC plus first and second derivatives) with utterance-level cepstral mean normalization. MLLR, MAP and full retraining were compared, and retraining won. Error rates were 60.7% for the telephony model, 36.4% after 10 hours of broadcast data, and 36.1% after 22 hours. - Automatic segmentation used Viterbi alignment anchored on correctly decoded words, with a small per-programme vocabulary and language model (perplexity about 10). Of 30 auto-processed hours, only about 12 survived after very short and very long segments were dropped. ## Interesting ideas - Long unlabeled streams can be split by first finding spots where the system's output exactly matches the reference, then aligning only the stretches between those anchors; overly long leftover chunks reveal where recognition failed. - A small, expensive hand-labelled seed (10 hours) beat the much larger auto-labelled addition, and the authors reported that null result openly instead of hiding it. - Missing information is restored through a cascade: dictionary lookup, then affix stripping with stem lookup, then a statistical model offering N-best candidates, then a human check. - Restricting the vocabulary and language model to the words of one specific document dropped perplexity from about 500 to about 10, though it made the system fragile when the input matched poorly. - Mechanically filtering auto-produced segments by length removed most bad outputs cheaply, at the cost of discarding over half the data. ## For Mova How these ideas could serve Mova, our small open language model — its files, stages and tests: - **I1** (learn): When aligning silver parses with reference annotations across long documents, anchor on tokens where all witness parsers agree, and treat long unanchored stretches as low-confidence regions to exclude or send to the LLM teacher. - **I2** (learn): Always measure silver additions against a gold-only baseline on a fixed dev set, so that volume of agreed silver data is never assumed to be a gain without evidence. - **I3** (tokens): Out-of-vocabulary lemmatization could follow the same lookup → affix-stripping → statistical-guess chain, recording which stage produced each lemma so every decision stays explainable. - **I4** (world): Entity and referent resolution could first search only candidates mentioned in the current text before falling back to the full knowledge graph, keeping that fallback because the narrow search can fail. - **I5** (learn): Simple structural sanity filters, such as sentence length or tree depth bounds, are a cheap first gate on silver trees, but the discard rate should be tracked as a signal of upstream parser quality. *Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.* ## Institutions - [Cairo University](https://md.research.elamur.ai/free/en/universities/cairo-university/) — University, Giza, Egypt · 12 works · 730 citations - [IBM (Egypt)](https://md.research.elamur.ai/free/en/companies/ibm-egypt/) — Company, Giza, Egypt · 2 works · 3 citations 📚 Shelves (5): [🖥️ Computer Science](https://md.research.elamur.ai/free/en/shelves/computer-science/) · 9,086 works · [🤖 Artificial Intelligence](https://md.research.elamur.ai/free/en/shelves/computer-science/artificial-intelligence/) · 8,889 works · [🎭 Arts and Humanities](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/) · 9,296 works · [🗣️ Language and Linguistics](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/language-and-linguistics/) · 9,206 works · [🔊 Signal Processing](https://md.research.elamur.ai/free/en/shelves/computer-science/signal-processing/) · 213 works 🧭 Topics (2): [📏 Evaluation & benchmarks](https://md.research.elamur.ai/free/en/topics/evaluation-benchmarks/) · 685 works · [Semitic languages](https://md.research.elamur.ai/free/en/topics/semitic-languages/) · 1,484 works ## Cite Cite as: Mohamed Afify and Ossama Emam. 2004. Collection and Evaluation of Broadcast News Data for Arabic. Proceedings of the Fourth International Conference on Language Resources and Evaluation (LREC'04). The paper itself, not this page. BibTeX from the library's record of the work. ```bibtex @inproceedings{afify2004collection, title = {{Collection and Evaluation of Broadcast News Data for Arabic}}, author = {Mohamed Afify and Ossama Emam}, booktitle = {Proceedings of the Fourth International Conference on Language Resources and Evaluation (LREC'04)}, year = {2004}, doi = {10.63317/2pmqwv2r3dun}, url = {https://doi.org/10.63317/2pmqwv2r3dun} } ``` - [← previous Chapter 3: Minimalist assumptions on negation in Biblical Hebrew syntax](https://md.research.elamur.ai/free/en/papers/snyman-2004-chapter-3-minimalist-assumptions-negation/) - [next → Extraposition in Modern Hebrew](https://md.research.elamur.ai/free/en/papers/bar-2026-extraposition-modern-hebrew/) - [↑ back to papers of 2004](https://md.research.elamur.ai/free/en/papers/2004/page/2/) ## References 4 references, 0 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet. 1. J. Billa et al., ”Audio indexing of Arabic broadcast news,” in Proceedings of ICASSP 2002, Orlando, Florida. [find](https://scholar.google.com/scholar?q=J%2E+Billa+et+al%2E%2C+%E2%80%9DAudio+indexing+of+Arabic+broadcast+news%2C%E2%80%9D+in+Proceedings+of+ICASSP+2002%2C+Orlando%2C+Florida%2E) 2. K. Kirchoff et al., ”Novel approaches to Arabic speech recognition-final report from the JHU summer workshop,” Tech. report, John-Hopkins University, 2002. [find](https://scholar.google.com/scholar?q=K%2E+Kirchoff+et+al%2E%2C+%E2%80%9DNovel+approaches+to+Arabic+speech+recognition%2Dfinal+report+from+the+JHU+summer+workshop%2C%E2%80%9D+Tech%2E+report%2C+John%2DHopkins+University%2C+2002%2E) 3. L.R. Bahl, P.V. deSouza, P.S. Gopalakrishnan, D. Nahamoo, and M. Picheny, ”Robust methods for context dependent features and models in a continuous speech recognizer,” in Proc. ICASSP’94, Adaleide, Australia, April 1994. [find](https://scholar.google.com/scholar?q=L%2ER%2E+Bahl%2C+P%2EV%2E+deSouza%2C+P%2ES%2E+Gopalakrishnan%2C+D%2E+Nahamoo%2C+and+M%2E+Picheny%2C+%E2%80%9DRobust+methods+for+context+dependent+features+and+models+in+a+continuous+speech+recognizer%2C%E2%80%9D+in+Proc%2E+ICASSP%E2%80%9994%2C+Adaleide%2C+Au) 4. M. Afify, ”Automatic vowelization of Arabic text,” unpublished research report, IBM Cairo Technology Development Center, December 2002. [find](https://scholar.google.com/scholar?q=M%2E+Afify%2C+%E2%80%9DAutomatic+vowelization+of+Arabic+text%2C%E2%80%9D+unpublished+research+report%2C+IBM+Cairo+Technology+Development+Center%2C+December+2002%2E) *Updated 24 October 2026*