📄 A Two-Stage Parsing Method for Text-Level Discourse Analysis

Summary

The paper addresses RST discourse parsing at the document level. Earlier transition-based parsers predicted span, nuclearity and relation together through one classifier. That design produces a large and sparse action set, which is hard to learn from the small RST-DT corpus. The authors split the task into two pipelined stages. First, a shift-reduce parser with only four actions builds an unlabeled ("naked") tree that already encodes nuclearity. Second, each internal node of that tree receives a relation label from one of three separate classifiers, chosen by whether the two children lie within one sentence, span sentences of the same paragraph, or cross paragraph boundaries. These relation classifiers can use features derived from the finished tree, such as node depth and the nuclearity of ancestors. All components are linear SVMs over hand-crafted features. On RST-DT, the system achieved the best span (86.0) and nuclearity (72.4) scores among the parsers compared, and its relation score (59.7) was competitive. This suggests that a carefully reduced action inventory makes transition-based discourse parsing viable again.

Key points

• Evaluation uses the RST Discourse Treebank (RST-DT): 385 Wall Street Journal documents, split into 347 for training and 38 for testing. The setup follows standard conventions:
  • 18 coarse-grained relations;
  • right-branching binarization;
  • gold EDU segmentation;
  • Stanford CoreNLP for preprocessing.
• The parser's action set is Shift, Reduce-NN, Reduce-NS and Reduce-SN, compared with more than 40 reduce actions in unified designs. In the training data, the least frequent action (Reduce-SN) still occurs 3065 times.
• Final F-scores are S 86.0, N 72.4 and R 59.7, against human agreement of 88.7 / 77.7 / 65.8. Ji and Eisenstein (2014) remain stronger on relations, with 61.6.
• The ablation results show where the gains come from:
  • Switching from a single-stage parser to the two-stage design raises scores from 84.4 / 70.7 / 57.7 to 86.0 / 72.4 / 58.6.
  • Adding level-specific classifiers raises the relation score to 59.4.
  • Adding tree features raises it further to 59.7.
• Attribution (F 0.87) and Same-Unit (F 0.83), which occur mostly within sentences, are the easiest relations. Compared with the single-stage baseline, the rarer and mostly cross-paragraph relations improve noticeably: Textual-Organization by about 20% and Topic-Comment by about 8%.

Interesting ideas

• I1. Shrink the label space of the structure-building classifier. The parser predicts only structure and closely tied attributes, and defers fine-grained labels to a later step. This trick helps any low-resource structured prediction task where combined labels become too sparse to learn.
• I2. Treat nuclearity like arc direction. Nuclearity is folded into the reduce action, much as left-arc and right-arc encode head direction in dependency parsing. Asymmetric properties that are intrinsic to the structure can often be predicted along with it at little extra cost.
• I3. Label after the full structure exists. Once the complete tree is built, global features become available for labeling, such as depth, height and the grandparent's nuclearity. A greedy one-pass parser cannot see these, so a pipeline can exploit global context that a joint greedy model lacks.
• I4. Route instances to separate classifiers by granularity. Relations are sent to different classifiers depending on whether they hold within a sentence, across sentences, or across paragraphs, and each classifier gets its own feature subset. For example, dependency features are used only within sentences. This could carry over to document-level tasks where phenomena differ by scope, such as coreference or argument mining.
• I5. Build an incremental ablation ladder. The authors add one design component at a time (stage, then level, then tree features). This cleanly attributes the gains to each choice and makes for a cheap, transparent evaluation template.