Elamur Research⌕
☰
home / papers / 2012 / Towards a Semantic Annotation of English Television News - Building and Evaluating a Constraint Grammar FrameNet

Towards a Semantic Annotation of English Television News - Building and Evaluating a Constraint Grammar FrameNet

Paper · 2012 · Shaohua Yang, Hai Zhao and Bao-liang Lu

Summary

Note: the requested title does not match the supplied text. The supplied text is instead a PACLIC 2012 paper by the same authors that applies machine translation to converting entire Chinese sentences from Pinyin into characters, and this summary covers that text.

The paper addresses Pinyin-to-character (PTC) conversion: turning a full sentence of toneless Pinyin into Chinese characters. The task is hard because roughly 410 Pinyin syllables must cover thousands of characters, so each syllable is highly ambiguous. Treating PTC as sequence labeling forces a classifier to choose among thousands of labels, which makes it slow and weak. The authors instead treat PTC as monotone phrase-based statistical machine translation, built on Stanford Phrasal. Because Pinyin and characters align one-to-one and never reorder, the costly alignment training step can be skipped, and phrases are extracted directly. They train on three data sizes (10K, 100K and 1M sentences) and compare against a Maximum Entropy (ME) labeler. SMT wins clearly on both character-level and whole-sentence accuracy and also trains much faster. The work shows that recasting a large-label-set tagging task as constrained translation can pay off in both quality and cost.

Key points

  • The system is monotone phrase-based SMT using Stanford Phrasal, a 3-gram language model, a maximum phrase length of 7, and phrase extraction directly from one-to-one Pinyin–character alignment.
  • Whole-sentence accuracy for ME vs SMT:
    • 10K: 0.075 vs 0.402
    • 100K: 0.169 vs 0.429
    • 1M: 0.302 vs 0.454
  • On 1M sentences, training took about a week for ME and under a day for SMT. Prior work reports that CRF training for PTC takes roughly 200 days without beating ME.
  • SMT decoding stays below 0.45s even for sentences longer than 120 characters. For typical 20–40 character sentences it is no slower than ME.
  • Prior PTC methods covered include language models, HMMs, SVM (92.94% accuracy reported by Jiang et al.), ME/MEMM, CRF, and an artificial immune network.

Interesting ideas

  • I1. When a tagging task has thousands of labels but the input-to-output order is strictly monotone, it can be cast as reordering-free translation so that phrase pairs carry context jointly.
  • I2. If the alignment between source and target units is known to be one-to-one, skip learning it and extract phrase pairs directly, which removes the most expensive training step.
  • I3. Phrase-level scoring resolved an error where per-character choices went wrong ("bang shou" read as "helper" instead of "top of the list") because the longer phrase outscored a split into shorter ones.
  • I4. Whole-sentence accuracy is reported alongside per-unit accuracy because one wrong unit costs the user a full correction, so the strict metric better reflects real usefulness.
  • I5. Decoding speed is compared at matched sentence lengths, which separates real speed differences from differences in length mix.

For Mova

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

  • I1 (tokens): lemmatization or morphological analysis of rare forms could be framed as monotone segment-to-segment mapping, with multiword chunks scored together instead of per-token labels.
  • I2 (back): when generating text from the meaning graph, aligned graph-node/word pairs from Mova's own analyses could feed a phrase table without a separate alignment model.
  • I3 (frames): when resolving ambiguous word senses for frames, scoring whole multiword constructions before single words could prevent locally plausible but globally wrong readings.
  • I4 (learn): parser and frame evaluations could report the share of fully correct sentences or trees next to LAS, exposing systems that are good on average but rarely completely right.
  • I5 (tree): benchmarking the perceptron parser against gparse should bucket timing by sentence length so the speed trade-off for long inputs is visible.

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

🎓Shanghai Jiao Tong UniversityShanghai · China63 works · 15,047 citations🎓Northeastern UniversityShenyang · China6 works · 1,810 citations🎓Michigan State UniversityEast Lansing · United States7 works · 635 citations

📚 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 English · 1,373 works Evaluation & benchmarks · 685 works Semantic role labeling · 132 works

Cite

Cite as Shaohua Yang, Hai Zhao and Bao-liang Lu. 2012. Towards a Semantic Annotation of English Television News - Building and Evaluating a Constraint Grammar FrameNet. Proceedings of the 26th Pacific Asia Conference on Language, Information, and Computation.

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

@inproceedings{yang2012semantic,
  title = {{Towards a Semantic Annotation of English Television News - Building and Evaluating a Constraint Grammar FrameNet}},
  author = {Shaohua Yang, Hai Zhao and Bao-liang Lu},
  booktitle = {Proceedings of the 26th Pacific Asia Conference on Language, Information, and Computation},
  year = {2012},
  url = {https://aclanthology.org/Y12-1036/}
}

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. McCallum Andrew, Pereira Fernando, and Lafferty John. 2001. Conditional random fields: probabilistic models for segmenting and labeling sequence data. In Proceedings of the Eighteenth International Conference on Machine Learning (ICML), pages 282–289, Williamstown, MA. find
  2. Adam L. Berger, Peter F. Brown, Stephen A. Della Pietra, Vincent J. Della Pietra, John R. GiUet, John D. Lafferty, Robert L. Mercer, Harry Printz, and Luboš Ureš. 1994. The candide system for machine translation. In Proceedings of the workshop on Human Language Technology, pages 157–162. Association for Computational Linguistics. find
  3. Adam L. Berger, Vincent J. Della Pietra, and Stephen A. Della Pietra. 1996. A maximum entropy approach to natural language processing. Computational linguistics, 22(1):39–71. ↗ our card
  4. Peter F. Brown, John Cocke, Stephen A. Della Pietra, Vincent J. Della Pietra, Fredrick Jelinek, John D. Lafferty, Robert L. Mercer, and Paul S. Roossin. 1990. A statistical approach to machine translation. Computational linguistics, 16(2):79–85. ↗
  5. Peter F. Brown, Vincent J. Della Pietra, Stephen A. Della Pietra, and Robert L. Mercer. 1993. The mathematics of statistical machine translation: Parameter estimation. Computational linguistics, 19(2):263–311. find
  6. Daniel Cer, Michel Galley, Daniel Jurafsky, and Christopher D.Manning. 2010. Phrasal: A statistical machine translation toolkit for exploring new model features. In Proceedings of the NAACL HLT 2010 Demonstration Session, pages 9–12, Los Angeles, California, June. Association for Computational Linguistics. find
  7. Zheng Chen and Kai-Fu Lee. 2000. A new statistical approach to chinese pinyin input. In Proceedings of the 38th Annual Meeting on Association for Computational Linguistics, pages 241–247, Hong Kong. Association for Computational Linguistics. find
  8. Stanley F. Chen. 2003. Conditional and joint models for grapheme-to-phoneme conversion. In Eighth European Conference on Speech Communication and Technology. find
  9. Ulrich Germann, Michael Jahr, Kevin Knight, Daniel marcu, and Kenji Yamada. 2001. Fast decoding and optimal decoding for machine translation. In Proceedings of the 39th Annual Meeting on Association for Computational Linguistics, pages 228–235. Association for Computational Linguistics. our card
  10. Jun Hatori and Hisami Suzuki. 2011. Japanese pronunciation prediction as phrasal statistical machine translation. In Proceedings of 5th International Joint Conference on Natural Language Processing, pages 120– 128. Asian Federation of Natural Language Processing. find
  11. Wei Jiang and Xiuli Pang. 2009. An artificial immune network approach for pinyin-to-character conversion. In Virtual Environments, Human-Computer Interfaces and Measurements Systems, 2009. VECIMS’09. IEEE International Conference on, pages 27–32. IEEE. find
  12. Wei Jiang, Yi Guan, Xiaolong Wang, and BingQuan Liu. 2007. Pinyin-to-character conversion model based on support vector machines. Journal of Chinese information processing, 21(2):100–105. find
  13. Pak kwong Wong and Chorkin Chan. 1996. Chinese word segmentation based on maximum matching and word binding force. In Proceedings of the 16th conference on Computational linguistics-Volume 1, pages 200–203. Association for Computational Linguistics. find
  14. Yue-Shi Lee. 2003. Task adaptation in stochastic language model for chinese homophone disambiguation. ACM Transactions on Asian Language Information Processing (TALIP), 2(1):49–62. ↗
  15. Haizhou Li, Min Zhang, and Jian Su. 2004. A joint source-channel model for machine transliteration. In Proceedings of the 42nd Annual Meeting on Association for Computational Linguistics, pages 159–166. Association for Computational Linguistics. our card
  16. Lu Li, Xuan Wang, Xiaolong Wang, and Yanbing Yu. 2009. A conditional random fields approach to chinese pinyin-to-character conversion. Journal of Communication and Computer, 6(4):25–31. find
  17. Bo Lin and Jun Zhang. 2008. A novel statistical chinese language model and its application in pinyin-to-character conversion. In Proceeding of the 17th ACM conference on Information and knowledge management, pages 1433–1434. ACM. find
  18. Bingquan Liu and Xaiolong Wang. 2002. An approach to machine learning of chinese pinyin-to-character conversion for small-memory application. In Machine Learning and Cybernetics, 2002. Proceedings. 2002 International Conference on, volume 3, pages 1287– 1291. IEEE. find
  19. Franz Josef Och and Hermann Ney. 2004. The alignment template approach to statistical machine translation. Computational Linguistics, 30(4):417–449. ↗
  20. Franz Josef Och. 2003. Minimum error rate training in statistical machine translation. In Proceedings of the 41st Annual Meeting on Association for Computational Linguistics-Volume 1, pages 160–167. Association for Computational Linguistics. find
  21. Xiaolong Wang, Qingcai Chen, and Daniel So Yeung. 2004. Mining pinyin-to-character conversion rules from large-scale corpus: a rough set approach. Systems, Man, and Cybernetics, Part B: Cybernetics, IEEE Transactions on, 34(2):834–844. ↗
  22. Xuan Wang, Lu Li, Lin Yao, and Waqas Wanwar. 2006. A maximum entropy approach to chinese pin yin-to-character conversion. In IEEE International Conference on Systems, Man and Cybernetics, volume 4, pages 2956–2959. IEEE. find
  23. Yan Zhang, Bo Xu, and Chengqing Zong. 2006. Rule-based post-processing of pinyin to chinese characters conversion system. In International Symposium on Chinese Spoken Language Processing. find
  24. Sen Zhang. 2007. Solving the pinyin-to-chinese-character conversion problem based on hybrid word lattice. CHINESE JOURNAL OF COMPUTERS-CHINESE EDITION-, 30(7):1145–1153. find
  25. Yibao Zhao. and Shenghe Sun. 1998. A word-self-made chinese phonetic-character conversion algorithm based on chinese character bigram [j]. ACTA ELEC-TRONICA SINICA, 10. find
  26. Xiaohua Zhou, Xiaohua Hu, Xiaodan Zhang, and Xiajiong Shen. 2007. A segment-based hidden markov model for real-setting pinyin-to-chinese conversion. In Proceedings of the sixteenth ACM conference on Conference on information and knowledge management, pages 1027–1030. ACM. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/yang-2012-semantic-annotation-english-television-news/