Predicting Discourse Connectives for Implicit Discourse Relation Recognition

Summary

Implicit discourse relations are hard to classify because no connective (such as "but" or "because") signals the sense. On PDTB 2.0, the authors found that mapping each gold implicit connective to its most frequent sense gives about 91.8% average F1 over the four top-level senses. However, those connectives come from annotation and do not appear in raw text. The authors therefore predict them automatically. They build synthetic sentences by placing each candidate connective either before Arg1 or between the arguments. A language model trained on unannotated text scores these sentences by perplexity, and the best-ranked connectives are kept.

The predicted connectives are used in two ways:

As extra features in a supervised classifier with nine linguistically motivated feature types.
As the sole signal, with each predicted connective mapped to its most frequent sense.

The feature-based approach raises average F1 by roughly 3 points over a strong baseline. The connective-only approach is competitive with much richer systems. The study shows that unlabeled data can supply useful connective knowledge for discourse parsing.

Key points

Setup: Experiments use PDTB 2.0, with sections 2–20 for training, 21–22 for testing and 0–1 for development. There are four one-vs-rest binary tasks (Comparison, Contingency, Expansion, Temporal), and training sets are balanced.
Supervised features: The nine feature types include verb classes (Levin), context, polarity, modality, General Inquirer tags, first/last words, production rules, and cross- and intra-argument word pairs.
Feature-based gains: Adding predicted connectives improves F1 by 1.07 (Comp.), 1.78 (Cont.), 4.16 (Exp.) and 3.84 (Temp.) points. Expansion reaches about 70% F1.
Connective-only results: F1 is 26.02 (Comp.), 35.72 (Cont.), 64.94 (Exp.) and 13.76 (Temp.). The Comparison score is about 5 points above Pitler et al. (2009a).
Upper bound: Using gold implicit connectives yields F1 of 94.08 (Comp.), 98.19 (Cont.), 97.79 (Exp.) and 77.04 (Temp.).

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

58 papers of our library cite it (of the 140673 whose bibliographies we read); 36 of them have a page here, 22 not yet.

Cite

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

@inproceedings{zhou2010predicting,
  title = {{Predicting Discourse Connectives for Implicit Discourse Relation Recognition}},
  author = {Zhou, Zhi-Min and Xu, Yu and Niu, Zheng-Yu and Lan, Man and Su, Jian and Tan, Chew Lim},
  booktitle = {Coling 2010: Posters},
  year = {2010},
  url = {https://aclanthology.org/C10-2172/}
}

References

18 references, 10 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

J. Baldridge and A. Lascarides. 2005. Probabilistic head-driven parsing for discourse structure. Proceedings of the Ninth Conference on Computational Natural Language Learning. find
L. Carlson, D. Marcu, and Ma. E. Okurowski. 2001. Building a discourse-tagged corpus in the framework of rhetorical structure theory. Proceedings of the Second SIG dial Workshop on Discourse and Dialogue. find our card
B. Dorr. LCS Verb Database. Technical Report Online Software Database, University of Maryland, College Park, MD,2001. find
R. Girju. 2003. Automatic detection of causal relations for question answering. In ACL 2003 Workshops. find
S. Blair-Goldensohn. 2007. Long-Answer Question Answering and Rhetorical-Semantic Relations. Ph.D. thesis, Columbia Unviersity. find
M. Lapata and A. Lascarides. 2004. Inferring Sentence-internal Temporal Relations. Proceedings of the North American Chapter of the Assocation of Computational Linguistics. find our card
Z.H. Lin, M.Y. Kan and H.T. Ng. 2009. Recognizing Implicit Discourse Relations in the Penn Discourse Treebank. Proceedings of the 2009 Conference on EMNLP. find our card
D. Marcu and A. Echihabi. 2002. An Unsupervised Approach to Recognizing Discourse Relations. Proceedings of the 40th ACL. find our card
E. Pitler, A. Louis, A. Nenkova. 2009. Automatic sense prediction for implicit discourse relations in text. Proceedings of the 47th ACL. find our card
E. Pitler and A. Nenkova. 2009. Using Syntax to Disambiguate Explicit Discourse Connectives in Text. Proceedings of the ACL-IJCNLP 2009 Conference Short Papers. find our card
M. Porter. 1980. An algorithm for suffix stripping. In Program, vol. 14, no. 3, pp.130-137. find
R. Prasad, N. Dinesh, A. Lee, E. Miltsakaki, L. Robaldo, A. Joshi, B. Webber. 2008. The Penn Discourse TreeBank 2.0. Proceedings of LREC’08. find our card
M. Saito, K.Yamamoto, S.Sekine. 2006. Using Phrasal Patterns to Identify Discourse Relations. Proceeding of the HLTCNA Chapter of the ACL. find our card
R. Soricut and D. Marcu. Sentence Level Discourse Parsing using Syntactic and Lexical Information. Proceedings of HLT/NAACL 2003. find our card
C. Sporleder and A. Lascarides. 2008. Using automatically labelled examples to classify rhetorical relations: an assessment. Natural Language Engineering, Volume 14, Issue 03. find
P.J. Stone, J. Kirsh, and Cambridge Computer Associates. 1966. The General Inquirer: A Computer Approach to Content Analysis. MIT Press. find
B. Wellner , J. Pustejovsky, C. H. R. S., A. Rumshisky. 2006. Classification of discourse coherence relations: An exploratory study using multiple knowledge sources. Proceedings of the 7th SIGDIAL Workshop on Discourse and Dialogue. find our card
F. Wolf, E. Gibson, A. Fisher, M. Knight. 2005. The Discourse GraphBank: A database of texts annotated with coherence relations. Linguistic Data Consortium. find

↑ back to top

📄 Predicting Discourse Connectives for Implicit Discourse Relation Recognition