Elamur Research⌕
☰
home / papers / 2011 / Pointwise Prediction for Robust, Adaptable Japanese Morphological Analysis

Pointwise Prediction for Robust, Adaptable Japanese Morphological Analysis

Summary

The paper asks whether Japanese morphological analysis (MA), meaning joint word segmentation and POS tagging of unsegmented text, really needs structured sequence models such as CRFs. The authors propose a two-step pointwise method. First, a binary classifier decides at every gap between characters whether a word boundary exists. Then each word gets a POS from a multi-class classifier. Both classifiers are logistic regression models that use only features read directly off the surface string: character n-grams, character-type n-grams and dictionary-match features. Because no predicted neighboring labels are used, each decision is independent, and partially annotated sentences can be used for training.

On BCCWJ, the method outperforms MeCab, the standard CRF-based joint analyzer, both in-domain and on web text. It nearly matches a CRF that uses the same features while training about three times faster. Combined with active learning that annotates only single uncertain words, it adapts to a new domain much faster than sentence-level annotation. This makes cheap, targeted domain adaptation practical.

Key points

  • Data: BCCWJ is split into two domains.

    • General domain: whitepaper, news and books sections, with 782k training words and 87.5k test words.
    • Target domain: web text, with 153k training words and 17.3k test words.

  • Systems compared:

    • JOINT: MeCab, an implementation of the Kudo et al. (2004) CRF method.
    • 2-LR: pointwise logistic regression trained with LIBLINEAR.
    • 2-CRF: a CRFSuite model using the same features as 2-LR.

  • Results (Word/POS F-measure):

    SettingJOINT2-LR2-CRF
    General → general97.31%98.03%98.08%
    General → web94.57%95.13%95.39%
    General + target → web96.45%96.82%96.91%

  • Training time: 2-LR trained in 16m44s, while 2-CRF needed 51m19s. JOINT's errors come largely from mis-segmenting unknown words because it relies heavily on the dictionary.

  • Active learning setup: 100 words were annotated per iteration, for 100 iterations.

    • Word-level partial annotation (PART) with 2-LR reached in 15 iterations the gains that sentence-level selection (SENT) needed 100 iterations for.
    • After those 15 iterations, PART with 2-LR overtook 2-CRF.

Interesting ideas

  • I1. Restrict features to the raw input. If no feature depends on predicted boundaries or neighboring tags, every decision point becomes a standalone training example. Partial annotations can then be used directly by simply skipping unlabeled points. This could suit any sequence task where labels are expensive.
  • I2. Treat segmentation as a per-gap yes/no decision. Classifying each character gap as boundary or no boundary turns segmentation into ordinary classification. This could carry over to tokenizing other unsegmented scripts, or to tasks like compound splitting.
  • I3. Let dictionary features describe the context. The dictionary features signal whether known words begin, end or span the current position, without committing to a segmentation. The lexicon informs the classifier but does not dictate the output, which helps with unknown words in new domains.
  • I4. Annotate single uncertain points, not whole sentences. The system queries the word or boundary with the smallest margin between its top two candidates, and annotates each unique word only once per iteration. Annotator effort goes only to genuinely hard cases, which is valuable when adapting a model to a new domain on a small budget.
  • I5. Count retraining speed as a practical cost. The authors note that a human annotator waits while the model retrains between rounds. Fast linear models can therefore beat marginally more accurate structured ones in interactive annotation workflows.

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: Natural language processing

Cited by

133 papers of our library cite it (of the 140884 whose bibliographies we read); 6 of them have a page here, 127 not yet.

Cite

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

@inproceedings{neubig2011pointwise,
  title = {{Pointwise Prediction for Robust, Adaptable Japanese Morphological Analysis}},
  author = {Neubig, Graham and Nakata, Yosuke and Mori, Shinsuke},
  booktitle = {Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies},
  year = {2011},
  url = {https://aclanthology.org/P11-2093/}
}

References

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

  1. Masayuki Asahara and Yuji Matsumoto. 2000. Extended models and tools for high-performance part-of-speech tagger. In Proceedings of the 18th International Conference on Computational Linguistics, pages 21–27. find
  2. Yee Seng Chan and Hwee Tou Ng. 2007. Domain adaptation with active learning for word sense disambiguation. In Proceedings of the 45th Annual Meeting of the Association for Computational Linguistics. find
  3. Gerard Escudero, Lluı́s Màrquez, and German Rigau. 2000. An empirical study of the domain dependence of supervised word sense disambiguation systems. In Proceedings of the 2000 Joint SIGDAT Conference on Empirical Methods in Natural Language Processing and Very Large Corpora. find
  4. Rong-En Fan, Kai-Wei Chang, Cho-Jui Hsieh, Xiang-Rui Wang, and Chih-Jen Lin. 2008. LIBLINEAR: A library for large linear classification. Journal of Machine Learning Research, 9:1871–1874. find
  5. Canasai Kruengkrai, Kiyotaka Uchimoto, Jun’ichi Kazama, Yiou Wang, Kentaro Torisawa, and Hitoshi Isahara. 2009. An error-driven word-character hybrid model for joint Chinese word segmentation and POS tagging. In Proceedings of the 47th Annual Meeting of the Association for Computational Linguistics. find
  6. Taku Kudo, Kaoru Yamamoto, and Yuji Matsumoto. 2004. Applying conditional random fields to Japanese morphological analysis. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, pages 230–237. find our card
  7. Taku Kudo. 2006. MeCab: yet another part-of-speech and morphological analyzer. http://mecab.sourceforge.net. ↗
  8. Percy Liang, Hal Daumé III, and Dan Klein. 2008. Structure compilation: trading structure for features. In Proceedings of the 25th International Conference on Machine Learning, pages 592–599. find
  9. Kikuo Maekawa. 2008. Balanced corpus of contemporary written Japanese. In Proceedings of the 6th Workshop on Asian Language Resources, pages 101–102. find
  10. Masaaki Nagata. 1994. A stochastic Japanese morphological analyzer using a forward-DP backward-A ∗ Nbest search algorithm. In Proceedings of the 15th International Conference on Computational Linguistics, pages 201–207. find
  11. Tetsuji Nakagawa. 2004. Chinese and Japanese word segmentation using word-level and character-level information. In Proceedings of the 20th International Conference on Computational Linguistics. find
  12. Graham Neubig and Shinsuke Mori. 2010. Word-based partial annotation for efficient corpus construction. In Proceedings of the 7th International Conference on Language Resources and Evaluation. find
  13. Hwee Tou Ng and Jin Kiat Low. 2004. Chinese part-of-speech tagging: one-at-a-time or all-at-once? word-based or character-based. In Proceedings of the Conference on Empirical Methods in Natural Language Processing. find our card
  14. Naoaki Okazaki. 2007. CRFsuite: a fast implementation of conditional random fields (CRFs). http://www.chokkan.org/software/crfsuite/. ↗
  15. Fuchun Peng, Fangfang Feng, and Andrew McCallum. 2004. Chinese segmentation and new word detection using conditional random fields. In Proceedings of the 20th International Conference on Computational Linguistics. find our card
  16. Piyush Rai, Avishek Saha, Hal Daumé III, and Suresh Venkatasubramanian. 2010. Domain Adaptation meets Active Learning. In Workshop on Active Learning for Natural Language Processing (ALNLP-10). find
  17. Eric Ringger, Peter McClanahan, Robbie Haertel, George Busby, Marc Carmen, James Carroll, Kevin Seppi, and Deryle Lonsdale. 2007. Active learning for part-of-speech tagging: Accelerating corpus annotation. In Proceedings of the Linguistic Annotation Workshop, pages 101–108. find
  18. Manabu. Sassano and Sadao Kurohashi. 2010. Using smaller constituents rather than sentences in active learning for Japanese dependency parsing. In Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 356–365. find
  19. Manabu Sassano. 2002. An empirical study of active learning with support vector machines for Japanese word segmentation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, pages 505–512. find
  20. Burr Settles and Mark Craven. 2008. An analysis of active learning strategies for sequence labeling tasks. In Conference on Empirical Methods in Natural Language Processing, pages 1070–1079. find
  21. Yuta Tsuboi, Hisashi Kashima, Hiroki Oda, Shinsuke Mori, and Yuji Matsumoto. 2008. Training conditional random fields using incomplete annotations. In Proceedings of the 22th International Conference on Computational Linguistics, pages 897–904. find

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/neubig-2011-pointwise-prediction-robust-adaptable-japanese/