Using Syntax to Disambiguate Explicit Discourse Connectives in Text

What it is about

Discourse connectives such as "once", "since" and "and" are often ambiguous in two ways: between a discourse use and a non-discourse use, and between discourse senses. The paper shows that syntax helps in both tasks. It uses gold Penn Treebank parses of the PDTB articles. The syntactic features are the category of the connective's node, of its parent, of its left and right siblings, and two flags on the right sibling (contains a VP, contains a trace). For discourse vs non-discourse usage, a maximum-entropy classifier with 10-fold cross-validation over sections 2 to 22 reaches F 75.33 with the connective string alone, 88.19 with syntax only, 92.28 with both, 93.63 with connective-syntax interactions, and 94.19 with syntax-syntax interactions. For the four top-level senses, a Naive Bayes classifier reaches 93.67% accuracy with the connective alone and 94.15% with syntax. The earlier surface-based result of Marcu (2000) was F 84.9, on a different corpus.

Ideas

I1. A gate for connectives in en parse output: an explicit decision list ("right sibling is a clause, so discourse use") with a reason stored per decision. Measure it on held-out data only, because the PDTB data are evaluation only.
I2. Keep a connective table in languages/en/36-discourse-relations.md as a BTreeMap from each connective to the syntactic contexts that it accepts. The paper says 11 connectives are discourse more than 90% of the time; start there.
I3. The paper uses constituency categories. Translate the features to UD: the deprel of the connective's head and of its right dependent. Report how much of the gain survives on the parser output of en/src/parse.rs, since the paper's numbers used gold trees.
I4. Negative control: replace the syntactic features with the labels of shuffled parents. The gain over the connective string must disappear.

For Mova

The code is en/src/parse.rs and en/src/tok.rs. The chapters are languages/en/36-discourse-relations.md, languages/en/38-dialogue-and-discourse-markers.md and languages/en/19-lexicon-mwe-register.md. The measure is accuracy and F on the discourse vs non-discourse decision, run twice on the same split, with gold and parser trees reported separately.

Library summary

The Research library's own summary of the paper, in our words — not the paper's text.

Written for the Research library (claude-opus-5-5) from the paper text, in our own words: the license "CC BY-NC-SA 3.0 (ACL Anthology, before 2016)" does not allow redistributing the text itself. Original — aclanthology.org/P09-2004.pdf

Summary

Words like "once", "since" and "and" can signal discourse relations, but they are ambiguous in two ways. They may or may not act as connectives at all. When they do, they can mark different kinds of relation. The authors ask whether syntactic context helps resolve both problems for explicit connectives in written English. They use the Penn Discourse Treebank (PDTB) with gold-standard Penn Treebank parses. From these they extract features about the connective's own constituent, its parent, and its left and right siblings. They also record whether the right sibling contains a VP or a trace. A maximum entropy classifier decides discourse vs. non-discourse usage, and a Naive Bayes classifier assigns one of four top-level senses. For usage detection, syntax alone beats the connective's identity alone. Combining both, plus interaction features, reaches an f-score of 94.19%. For sense labelling, adding syntax raises accuracy slightly to 94.15%, about the same as human inter-annotator agreement. The study shows that simple cues from the parse tree make explicit connectives reliable to identify, which is a basic step toward full discourse parsing.

Key points

Data: PDTB with 18,459 annotated instances of 100 explicit connectives. Sections 0–1 were used for feature development and sections 2–22 for ten-fold cross-validation. Only the top-level senses were used: Expansion, Comparison, Contingency and Temporal.
Only 11 of the 100 connectives function as connectives more than 90% of the time. The range is wide: "although" does so in 91.4% of cases, while "or" does in only 2.8%.
Discourse vs. non-discourse results (f-score / accuracy): connective string only 75.33 / 85.86, syntax only 88.19 / 92.25, both combined 92.28 / 95.04, and with connective–syntax and syntax–syntax interactions 94.19 / 96.26.
That is about 10 points better in f-score than Marcu (2000), who reported 84.9%. The comparison is not exact because the corpora differ and gold parses were used.
Sense classification: 93.67% accuracy with the connective alone and 94.15% with added syntax, against a reported inter-annotator agreement of 94%. Classifiers were built with MALLET.

Topics: Discourse analysis

Cite

The paper itself, not this card. BibTeX from the library's record of the work. DOI 10.3115/1667583.1667589

@inproceedings{pitler2009using,
  title = {{Using Syntax to Disambiguate Explicit Discourse Connectives in Text}},
  author = {Pitler, Emily and Nenkova, Ani},
  booktitle = {Proceedings of the ACL-IJCNLP 2009 Conference Short Papers},
  year = {2009},
  doi = {10.3115/1667583.1667589},
  url = {https://doi.org/10.3115/1667583.1667589}
}

References

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

R. Elwell and J. Baldridge. 2008. Discourse connective argument identification with connective specific rankers. In Proceedings of the International Conference on Semantic Computing, Santa Clara, CA. find
S. Fisher and B. Roark. 2007. The utility of parse-derived features for automatic discourse segmentation. In Proceedings of ACL, pages 488–495. find
A. Gravano, S. Benus, H. Chavez, J. Hirschberg, and L. Wilcox. 2007. On the role of context and prosody in the interpretation of ’okay’. In Proceedings of ACL, pages 800–807. find
J. Hirschberg and D. Litman. 1993. Empirical studies on the disambiguation of cue phrases. Computational linguistics, 19(3):501–530. find
D. Marcu. 2000. The rhetorical parsing of unrestricted texts: A surface-based approach. Computational Linguistics, 26(3):395–448. find our card
M.P. Marcus, B. Santorini, and M.A. Marcinkiewicz. 1994. Building a large annotated corpus of english: The penn treebank. Computational Linguistics, 19(2):313–330. find our card
E. Miltsakaki, N. Dinesh, R. Prasad, A. Joshi, and B. Webber. 2005. Experiments on sense annotation and sense disambiguation of discourse connectives. In Proceedings of the Fourth Workshop on Treebanks and Linguistic Theories (TLT 2005). find
E. Pitler, M. Raghupathy, H. Mehta, A. Nenkova, A. Lee, and A. Joshi. 2008. Easily identifiable discourse relations. In COLING, short paper. find
R. Prasad, N. Dinesh, A. Lee, E. Miltsakaki, L. Robaldo, A. Joshi, and B. Webber. 2008. The penn discourse treebank 2.0. In Proceedings of LREC’08. find our card
R. Soricut and D. Marcu. 2003. Sentence level discourse parsing using syntactic and lexical information. In HLT-NAACL. find our card
C. Sporleder and A. Lascarides. 2008. Using automatically labelled examples to classify rhetorical relations: An assessment. Natural Language Engineering, 14:369–416. find
B. Wellner and J. Pustejovsky. 2007. Automatically identifying the arguments of discourse connectives. In Proceedings of EMNLP-CoNLL, pages 92–101. find
B. Wellner, J. Pustejovsky, C. Havasi, A. Rumshisky, and R. Sauri. 2006. Classification of discourse coherence relations: An exploratory study using multiple knowledge sources. In Proceedings of the 7th SIGdial Workshop on Discourse and Dialogue. find