Elamur Research⌕
☰
home / papers / 2015 / Spectral Semi-Supervised Discourse Relation Classification

Spectral Semi-Supervised Discourse Relation Classification

Summary

Implicit discourse relations, where no connective word signals the link between two text spans, are among the hardest parts of discourse parsing. Most systems depend only on costly human annotation. Fisher and Simmons propose a semi-supervised approach that combines the Penn Discourse Treebank (PDTB) with a large pool of unlabeled news text from the New York Times Annotated Corpus. Relations between adjacent EDUs are treated as a sequence and modeled with an HMM-style latent variable model, learned with spectral methods following Hsu et al. (2012). This makes training fast, statistically consistent and free of local optima.

Unlabeled data enters through bootstrapping. After each round, automatically labeled sequences that the model is confident about and that resemble the labeled corpus are added to the training set. On the PDTB implicit relations task, the method reaches an F1 of 0.485, about 7–9 points above purely supervised systems. It also matches supervised accuracy using roughly half the labeled data. This suggests discourse parsers could be moved to new domains with less annotation effort.

Key points

  • Data: PDTB (Wall Street Journal) supplies the labeled data. The paper reports about 16K implicit relations (16,224 in one place, 16,053 in another), restricted to 11 level-2 relation classes. The unlabeled data comes from the New York Times Annotated Corpus, with about 1.8M articles from 1987–2007. PDTB sections 2–22 were used for training and section 23 for testing.
  • Model: A spectral HMM with one latent state per relation is learned in an SVD-derived subspace. Features come from Feng and Hirst (2014) and are augmented with 200 LDA concepts computed over Wikipedia.
  • Semi-supervised selection: Unlabeled sequences are labeled with Viterbi and ranked by density-weighted certainty sampling (DCS). DCS combines prediction entropy with a KDE-based density computed over 100-dimensional LDA vectors. Training ran for 7 bootstrapping rounds.
  • Results: Trained on labeled data only, the model scores F1 41.1%, comparable to Lin et al. (2014). With unlabeled data it reaches F1 0.485, a gain of 7–9 points, and it matches the supervised score using about 50% of the labeled data.
  • Efficiency: Spectral optimization keeps training tractable even with more than one million unlabeled documents.

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: Linguistics & languages

Topics: Discourse analysis · Treebanks

Cited by

6 papers of our library cite it (of the 140673 whose bibliographies we read); 4 of them have a page here, 2 not yet.

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.3115/v1/p15-2015

@inproceedings{fisher2015spectral,
  title = {{Spectral Semi-Supervised Discourse Relation Classification}},
  author = {Fisher, Robert and Simmons, Reid},
  booktitle = {Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing (Volume 2: Short Papers)},
  year = {2015},
  doi = {10.3115/v1/p15-2015},
  url = {https://doi.org/10.3115/v1/p15-2015}
}

References

17 references, 4 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Pranjal Awasthi, Delip Rao, and Balaraman Ravindran. 2006. Part of speech tagging and chunking with hmm and crf. Proceedings of NLP Association of India (NLPAI) Machine Learning Contest 2006. find
  2. Byron Boots and Geoffrey J Gordon. 2010. Predictive state temporal difference learning. arXiv preprint arXiv:1011.0041. — [arXiv:1011.0041](https://arxiv.org/abs/1011.0041) find
  3. Shay B Cohen, Karl Stratos, Michael Collins, Dean P Foster, and Lyle Ungar. 2014. Spectral learning of latent-variable pcfgs: Algorithms and sample complexity. The Journal of Machine Learning Research, 15(1):2399–2449. find
  4. Paramveer S Dhillon, Jordan Rodu, Michael Collins, Dean P Foster, and Lyle H Ungar. 2012. Spectral dependency parsing with latent variables. In Proceedings of the 2012 joint conference on empirical methods in natural language processing and computational natural language learning, pages 205–213. Ass… find
  5. Vanessa Wei Feng and Graeme Hirst. 2012. Text-level discourse parsing with rich linguistic features. In Proceedings of the 50th Annual Meeting of the Association for Computational Linguistics: Long Papers-Volume 1, pages 60–68. Association for Computational Linguistics. find our card
  6. Vanessa Wei Feng and Graeme Hirst. 2014. A linear-time bottom-up discourse parser with constraints and post-editing. In Proceedings of The 52nd Annual Meeting of the Association for Computational Linguistics (ACL 2014), Baltimore, USA, June. find our card
  7. Robert Fisher, Reid Simmons, Cheng-Shiu Chung, Rory Cooper, Garrett Grindle, Annmarie Kelleher, Hsinyi Liu, and Yu Kuang Wu. 2014. Spectral machine learning for predicting power wheelchair exercise compliance. In Foundations of Intelligent Systems, pages 174–183. Springer. find
  8. Hugo Hernault, Danushka Bollegala, and Mitsuru Ishizuka. 2011. Semi-supervised discourse relation classification with structural learning. In Computational Linguistics and Intelligent Text Processing, pages 340–352. Springer. find
  9. Daniel Hsu, Sham M Kakade, and Tong Zhang. 2012. A spectral algorithm for learning hidden markov models. Journal of Computer and System Sciences, 78(5):1460–1480. find
  10. Shafiq R Joty, Giuseppe Carenini, Raymond T Ng, and Yashar Mehdad. 2013. Combining intra-and multi-sentential rhetorical parsing for document-level discourse analysis. In ACL (1), pages 486–496. find
  11. Ziheng Lin, Min-Yen Kan, and Hwee Tou Ng. 2009. Recognizing implicit discourse relations in the penn discourse treebank. In Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing: Volume 1-Volume 1, pages 343–351. Association for Computational Linguistics. find our card
  12. Ziheng Lin, Hwee Tou Ng, and Min-Yen Kan. 2014. A pdtb-styled end-to-end discourse parser. Natural Language Engineering, pages 1–34. find
  13. Ankur Parikh, Shay B Cohen, and Eric Xing. 2014. Spectral unsupervised parsing with additive tree metrics. In Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics: Long Papers. Association for Computational Linguistics. find
  14. Natalia Ponomareva, Paolo Rosso, Ferrán Pla, and Antonio Molina. 2007. Conditional random fields vs. hidden markov models in a biomedical named entity recognition task. In Proc. of Int. Conf. Recent Advances in Natural Language Processing, RANLP, pages 479–483. find
  15. Rashmi Prasad, Nikhil Dinesh, Alan Lee, Eleni Miltsakaki, Livio Robaldo, Aravind K Joshi, and Bonnie L Webber. 2008. The penn discourse treebank 2.0. In LREC. Citeseer. find our card
  16. Evan Sandhaus. 2008. The new york times annotated corpus ldc2008t19. Linguistic Data Consortium. find
  17. Bonnie Webber, Markus Egg, and Valia Kordoni. 2012. Discourse structure and language technology. Natural Language Engineering, 18(4):437–490. find

↑ back to top

This page in Markdown →