Elamur Research⌕
☰
home / papers / 2012 / Implicit Discourse Relation Recognition by Selecting Typical Training Examples

Implicit Discourse Relation Recognition by Selecting Typical Training Examples

Summary

Recognizing implicit discourse relations (sentence pairs linked without an explicit connective) is difficult, and earlier systems trained on one of two sources. Some used manually annotated implicit relations. Others used "artificial" implicit examples, made by deleting connectives from explicit relations. Wang et al. argue that the better question is whether an example is typical, meaning its arguments carry clear cues for the relation, rather than whether it is real or artificial. They propose single centroid clustering (SCC), a bootstrapping-style procedure inspired by Yarowsky's algorithm. Starting from seed examples, it grows a set of typical examples for each relation, drawing on both natively implicit and connective-stripped explicit data. Classifiers trained only on this typical subset (Naïve Bayes and Decision Tree, without dedicated feature selection) perform comparably to strong prior systems on PDTB v2.0. They also beat training on all real, all artificial, or combined data. A portability test on RST-DT shows statistically significant gains. The work matters because it shows that curating training data can improve performance as much as feature engineering does.

Key points

  • Data: Experiments use the top-level PDTB v2.0 classes (Comparison, Contingency, Temporal, Expansion), set up as one-vs-other and 4-way tasks.
  • Main PDTB result: With the Decision Tree, SCC-selected training data raises Comparison-vs-Other F1 to 28.5, compared with 11.6 when training on real implicit examples alone. The reported 4-way accuracy is about 72.2.
  • Composition of the typical set: It contains roughly equal shares of implicit and explicit-derived examples. About 61.8% of the data was judged typical. Most misclassified test items are atypical.
  • Portability (RST-DT): The data was split into 347 training and 38 test articles, using Contrast, Background, Cause and Temporal. Cause-vs-Other F1 improved from 35.6 (IM baseline) to 54.8. Paired t-test p-values were about 0.057–0.059.
  • Future directions named by the authors: richer features, finer-grained relations, alternative distance measures in SCC, and use of unlabelled data.

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

Cited by

16 papers of our library cite it (of the 140673 whose bibliographies we read); 12 of them have a page here, 4 not yet.

Cite

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

@inproceedings{wang2012implicit,
  title = {{Implicit Discourse Relation Recognition by Selecting Typical Training Examples}},
  author = {Wang, Xun and Li, Sujian and Li, Jiwei and Li, Wenjie},
  booktitle = {Proceedings of COLING 2012},
  year = {2012},
  url = {https://aclanthology.org/C12-1168/}
}

References

20 references, 7 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Carlson, L., Marcu, D. and Okuro wski, M.E. (2003). Bu ild ing a discourse-tagged corpus in the framework of rhetorical structure theory. In Janvan Kuppelvelt and Ronnie Smith, editors, Current Directions in Discourse and Dialogue. Kluwer Academic Publishers. find
  2. Christiann, V. W. and Barnard., E. (2006). Data characteristics that determine classifier performance. In Proceedings of the Sixteenth Annual Symposium of the Pattern Recognition Association of South Africa, pp. 166–171, Parys, South Africa. find
  3. Feng, V. W. and Hirst, G., (2012). Text -level discourse parsing with rich linguistic features, In Proc.of ACL'12, pages 60-68. find our card
  4. Hernault, H., Prendinger, H., David, A., and Mitsuru I. (2010). HILDA: a discourse parser using support vector machine classification. Dialogue and Discourse, 1(3):1-33. find our card
  5. Lin, Z., Kan, M.– Y., and Ng, H. T. (2009). Recognizing imp licit d iscourse relations in the Penn discourse treebank, In Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, Singapore. find
  6. Mann, W. C. and Tho mpson, S. A. (1988). Rhetorical structure theory: towards a functional theory of text organization. Text, 8(3):243-281. find
  7. Marcu, D. and Ech ihabi, A. (2002). An unsupervised approach to recognizing d iscourse relations. In Proc. of ACL 2002, pages 368-375. find
  8. Pit ler, E., Raghupathy, M., Mehta, H., Nenkova, A., Lee, A. and Joshi, A. (2008). Easily identifiable d iscourse relations. In Proc. of the 22nd International Conference on Computational Linguistics (COLING08). pages 85-88. find
  9. Pit ler, E., Louis, A. and Nenkova, A. (2009). Automatic sense prediction for implicit d iscourse relations in text, In Proc. of the 47th ACL. pages 683-691. find
  10. Prasad, R., Dinesh, N., Lee, A., Miltsakaki, E., Robaldo, L., Joshi, A. and Webber, B. (2008). The Penn discourse treebank 2.0. In Proc. o f the 6th International Conference on Language Resources and Evaluation (LREC). Marrakech, Morocco. find our card
  11. RST_ DT. (2002). RST Discourse Treebank. Linguistic Data http://www.ldc.upenn.edu/Catalog/CatalogEntry.jsp?catalogId=LDC2002T07 ↗
  12. Saito, M., Yamamoto, K. and Sekine, S. (2006). Using phrasal patterns to identify discourse relations. In Proc. of the HLTCNA Chapter of the ACL. pages 133-136. find our card
  13. Sasha B.-G. (2007). Long-Answer Question answering and rhetorical-semantic relations. Ph. D. thesis, Columbia Unversity. find
  14. Soricut, R. and Marcu, D. (2003). Sentence level discourse parsing using syntactic and lexical information. In Proc. of HLT/NAACL 2003. pages 149-156. find our card
  15. Sporleder, C. and Lascarides. A. (2008). Using automatically labelled examp les to classify rhetorical relations: An Assessment. Natural Language Engineering, 14:369-416. find
  16. Wang, W., Su, J. and Tan C. L. (2010). Kernel based discourse relation recognition with temporal ordering information, In Proc. of ACL’10. pages 710-719. find our card
  17. Consortium, Wilson, T., W iebe, J. and Hoffmann, P. (2005). Recognizing contextual polarity in phrase-level sentiment analysis. In Proc. of the conference on Human Language Technology and Empirical Methods in Natural Language Processing, pp. 347-354. find
  18. Wellner, B., Pustejovsky, J., Havasi, C., Ru mshisky, A. and Sauri, R. (2006). Classification of discourse coherence relations: an exploratory study using multip le knowledge sources. In Proc. of the 7th SIGDIAL Workshop on Discourse and Dialogue. pages 117-125. find
  19. Yarowsky, D. (1995). Unsupervised word sense disambiguation rivaling supervised methods, In Proc. of the 33rd annual meeting on Association for Computational Linguistics, pages 189-196, Cambridge, Massachusetts. find
  20. Zhou, Z., Lan, M., Niu, Z. and Su, J. (2010). The effects of discourse connectives prediction on implicit discourse relation recognition, In Proc. of SIGDIAL 2010: the 11th Annual Meeting of the Special Interest Group on Discourse and Dialogue, pages 139–146. find our card

↑ back to top

This page in Markdown →