Elamur Research⌕
☰
home / papers / 2006 / Story Segmentation of Broadcast News in English, Mandarin and Arabic

Story Segmentation of Broadcast News in English, Mandarin and Arabic

Paper · 2006 · Andrew Rosenberg and Julia Hirschberg

Summary

The paper addresses automatic detection of story boundaries in broadcast news audio. Within SRI's NIGHTINGALE question-answering system, search results must come in coherent, interpretable chunks. The authors treat segmentation as binary classification over candidate positions. Each hypothesized sentence boundary from upstream components (ASR, speaker diarization, sentence segmentation) is labeled as a story boundary or not. The classifier is the rule learner JRip (a RIPPER implementation in Weka), trained on the TDT-4 audio in English, Mandarin and Arabic. It uses three feature families, all extracted automatically:

  • Lexical: TextTiling and LCseg scores, plus cue-words selected by χ²
  • Acoustic/prosodic: F0, intensity, speaking rate and final-vowel lengthening
  • Speaker-dependent: features derived from diarization

Separate models are trained per news show. Combining all feature sets works best. Results appear competitive with prior work, though the corpora differ. These are claimed as the first reported Arabic broadcast news segmentation results. The most notable finding is that prosodic cues useful in English and Mandarin carry essentially no signal for Arabic. The authors also argue that precision matters more than recall for downstream processing.

Key points

  • Data: TDT-4 audio, comprising 312.5 h of English (450 shows), 134 h of Mandarin (205 shows) and 88.5 h of Arabic (109 shows), all aired between Oct 2000 and Jan 2001. TDT-5 is text-only and was not used.
  • Upstream ceiling: Story boundaries are only allowed at automatic sentence boundaries. This holds for just 47% of English reference boundaries (average offset 9.88 words), 69% for Mandarin and 62% for Arabic.
  • English results: The full L+A+S configuration reaches F1 .421 (P .67, R .31), Pk .194, WindowDiff .318 and Cseg .0670, with k = 5 for the window metrics. Mandarin L+A+S reaches F1 .592.
  • Show-specific modeling: Training a model per show reduced errors by roughly 10–30% across metrics.
  • Speaker features: These help noticeably even though diarization is noisy, especially for Mandarin. In China National Radio, speaker changes strongly track story changes.

Interesting ideas

  • I1. Measuring how many true boundaries even fall on upstream candidate positions gives a hard ceiling that explains downstream errors before any model is blamed.
  • I2. Cue-words can be mined with a χ² test in fixed windows before and after known boundaries, then counted at several window sizes as features.
  • I3. The authors deliberately favor precision over recall, because a spurious boundary damages anaphora resolution and summarization more than a missed one.
  • I4. The authors use pronoun density as a hint of story position, on the intuition that pronouns cluster late in a story once entities are established.
  • I5. Per-feature-family ablation across languages revealed that one cue type simply failed to transfer, rather than merely weakening.

For Mova

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

  • I1 (learn): Evaluation could report the fraction of gold structures reachable given tokenizer/sentence-splitter output, separating pipeline loss from parser or frame errors.
  • I2 (world): The same statistic could induce explainable discourse-transition markers that start new episodes or scenes when ordering events in tales.
  • I3 (frames): Coreference could treat segment breaks conservatively, cutting antecedent search only at high-confidence boundaries, so chains are not wrongly severed.
  • I4 (frames): Pronoun-to-full-NP ratios per sentence could serve as a cheap, interpretable signal of discourse position for salience ranking in coreference.
  • I5 (learn): Ablating feature groups per genre (fables, school tests, word problems) would expose cues that do not transfer, instead of hiding them in an averaged score.

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

🎓Columbia UniversityNew York · United States254 works · 56,747 citations

 Fields, topics and concepts

📚 Shelves 5The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksArtificial Intelligence · 17,965 worksArts and Humanities · 17,450 worksLanguage and Linguistics · 17,296 worksSignal Processing · 493 works

🧭 Topics 2Research subjects and languages the work is about.English · 2,379 worksSemitic languages · 3,066 works

Cite

Cite as Andrew Rosenberg and Julia Hirschberg. 2006. Story Segmentation of Broadcast News in English, Mandarin and Arabic. Proceedings of the Human Language Technology Conference of the NAACL, Companion Volume: Short Papers.

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

@inproceedings{rosenberg2006story,
  title = {{Story Segmentation of Broadcast News in English, Mandarin and Arabic}},
  author = {Andrew Rosenberg and Julia Hirschberg},
  booktitle = {Proceedings of the Human Language Technology Conference of the NAACL, Companion Volume: Short Papers},
  year = {2006},
  doi = {10.3115/1614049.1614081},
  url = {https://doi.org/10.3115/1614049.1614081}
}

References

26 references, 3 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. R. Barzilay, M. Collins, J. Hirschberg, and S. Whittaker. 2000. The rules behind roles: Identifying speaker role in radio broadcasts. In AAAI/IAAI, 679–684. ↗
  2. D. Beeferman, A. Berger, and J. Lafferty. 1999. Statistical models for text segmentation. Machine Learning, 31:177–210. ↗
  3. P. Boersma. 2001. Praat, a system for doing phonetics by computer. Glot International, 5(9-10):341–345. find
  4. L. Chaisorn, T. Chua, C. Koh, Y. Zhao, H. Xu, H. Feng, and Q. Tian. 2003. A two-level multi-modeal approach for story segmentation of large news video corpus. In TRECVID. find
  5. W. Cohen. 1995. Fast effective rule induction. In Machine Learning: Proc. of the Twelfth International Conference, 115–123. ↗
  6. G. Doddington. 1998. The topic detection and tracking phase 2 (tdt2) evaluation plan. In Proccedings DARPA Broadcast News Transcription and Understanding Workshop, 223–229. find
  7. M. Franz, J. S. McCarley, T. Ward, and W. J. Zhu. 2000. Segmentation and detection at ibm: Hybrid statstical models and two-tiered clustering. In Proc. of TDT-3 Workshop. Mandarin Pk WD .179 .245 .178 .252 .200 .278 .277 .354 .309 .391 .291 .377 Arabic Pk .264 .333 .286 .257 .283 .332 F1(p,r) .300(.65,.19) .0487(.81,.03) .285(.68,.18) .284(.64,.18) .286(.68,.18) .0526(.81,.03) WD .353 .426 .372 .344 .349 .422 C seg .0679 .0680 .0742 .0886 .0979 .0904 C seg .0850 .0999 .0884 .0851 .0849 .0996 find
  8. M. Galley, K. McKeown, E. Fosler-Lussier, and H. Jing. 2003. Discourse segmentation of multi-party conversation. In 41st Annual Meeting of ACL, 562–569.
  9. M. A. Hearst. 1997. Texttiling: Segmenting text into multi-paragraph subtopic passages. Computational Linguistics, 23(1):33–64. find
  10. J. Hirschberg and C. Nakatani. 1998. Acoustic indicators of topic segmentation. In Proc. of ICSLP, 1255–1258. find
  11. J. H. Hsieh, C. H. Wu, and K. A. Fung. 2003. Two-stage story segmentation and detection on broadcast news using genetic algorithm. In Proc. of the 2003 ISCA Workshop on Multilingual Spoken Document Retrieval (MSDR2003), 55–60. find
  12. W. Hsu, L. Kennedy, C. W. Huang, S. F. Chang, C. Y. Lin, and G. Iyengar. 2004. News video story segmentation using fusion of multi-level multi-modal features in trecvid 2003. In ICASSP. find
  13. H. Kozima. 1993. Text segmentation based on similarity between words. In 31st Annual Meeting of the ACL, 286–288. find
  14. G. A. Levow. 2004. Assessing prosodic and text features for segmentation of mandarin broadcast news. In HLT-NAACL. find
  15. M. T. Maybury. 1998. Discourse cues for broadcast news segmentation. In COLING-ACL, 819–822. find
  16. D. D. Palmer, M. Reichman, and E. Yaich. 2004. Feature selection for trainable multilingual broadcast news segmentation. In HLT/NAACL. find
  17. R. J. Passonneau and D. J. Litman. 1997. Discourse segmentation by human and automated means. Computational Liun-guistics, 23(1):103–109. ↗
  18. L. Pevzner and M. Hearst. 2002. A critique and improvement of an evaluation metric for text segmentation. Computational Linguistics, 28(1):19–36. ↗
  19. M. Porter. 1980. An algorithm for suffix stripping. Program, 14(3):130–137. ↗
  20. E. Shriberg, A. Stolcke, D. Hakkani-Tür, and G. Tür. 2000. Prosody based automatic segmentation of speech into sentences and topics. Speech Comm., 32(1-2):127–154. ↗
  21. N. Stokes. 2003. Spoken and written news story segmentation using lexical chains. In Proc. of the Student Workshop at HLT-NAACL2003, 49–53. find
  22. S. Strassel and M. Glenn. 2003. Creating the annotated tdt-4 y2003 evaluation corpus. http://www.nist.gov/speech/tests/tdt/tdt2003/papers/ldc.ppt. ↗
  23. S. Strassel, M. Glenn, and J. Kong. 2004. Creating the tdt5 corpus and 2004 evalutation topics at ldc. http://www.nist.gov/speech/tests/tdt/tdt2004/papers/LDC-TDT5.ppt. ↗
  24. G. Tür, D. Hakkani-Tür, A. Stolcke, and E. Shriberg. 2001. Integrating prosodic and lexical cues for automatic topic segmentation. Computational Linguistics, 27:31–57. ↗
  25. C. L. Wayne. 2000. Multilingual topic detection and tracking: Successful research enabled by corpora and evaluation. In LREC, 1487–1494. find
  26. I. Witten, E. Frank, L. Trigg, M. Hall, G. Holmes, and S. Cunningham. 1999. Weka: Practical machine learning tools and techniques with java implementation. In ICONIP/ANZIIS/ANNES, 192–196. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/rosenberg-2006-story-segmentation-broadcast-news-english/