📄 The Utility of Discourse Parsing Features for Predicting Argumentation Structure

Summary

Argument mining research has long suspected a link between discourse structure and argumentation structure, but evidence has mostly relied on manual discourse annotations. This paper tests whether output from off-the-shelf discourse parsers is useful for predicting argumentation. The authors use part 1 of the English argumentative microtext corpus, which has parallel gold annotations for argumentation and RST. They chose an RST parser (Feng and Hirst, 2014) and a PDTB-style shallow parser (Lin et al., 2014) and ran both on gold-segmented text without retraining. They converted the parser output and the gold trees into dependency format, compared them, and derived new RST+ and PDTB features. These features were added to a re-implementation of the Evidence Graph model of Afantenos et al. (2018), which uses MST decoding. The PDTB parser recovered relatively few relations. Even so, the PDTB features improved results more than the RST features did. Combining the feature sets helped most for the function and attachment subtasks. The study suggests that automatic discourse parsing, especially shallow PDTB analysis, can benefit argumentation parsing in practice.

Key points

• Data: The corpus has 112 microtexts with 576 segments. The texts were originally German and professionally translated into English. The parsers received gold segmentation because their own segment boundaries often differed from the gold ones.
• Central claims: The argumentative central claim matched an RST most-nuclear segment in 67 texts (about 60%). With gold RST trees, the match rate is 85%.
• Parser coverage:
  • The RST parser scored poorly against gold RST trees: F1 of 0.338 for span, 0.264 for nuclearity and 0.115 for relation.
  • The PDTB parser found only 176 relations, compared with 547 argumentation relations. Of these, 73 aligned with argumentation edges.
• Relation correspondences: RST contrast aligned with undercut 28 times, and PDTB Comparison.Contrast aligned with undercut 21 times.
• Prediction results: Scores are macro-F1 on the full relation set, using 10×5-fold cross-validation.
  • Adding PDTB features to the Default features raised the scores for central claim (0.722→0.771), role (0.467→0.720), function (0.224→0.420) and attachment (0.673→0.691).
  • The full combination (Default, 2018, RST, RST+, PDTB) gave the best function (0.461) and attachment (0.732) scores.