TL;DR
This paper explores how to automatically infer implicit discourse relations, common in Rhetorical Structure Theory (RST), from existing shallow, explicit discourse relation annotations (like those in the Penn Discourse Treebank - PDTB). The goal is to enrich existing annotations and bridge the gap between different discourse parsing frameworks, using a German news commentary corpus.
What Problem Does This Paper Address?
Discourse parsing aims to understand how sentences and clauses in a text connect to form a coherent whole. Several frameworks exist for this, notably Rhetorical Structure Theory (RST) and the Penn Discourse Treebank (PDTB). While both capture coherence, they differ in their segmentation of text and the types of relations they annotate. RST uses a hierarchical tree structure with explicit and implicit relations, while PDTB focuses on explicit relations marked by connectives. Creating annotated data for discourse parsing is expensive, leading to relatively small corpora. This paper addresses the challenge of leveraging existing, differently structured discourse annotations to enrich each other. Specifically, it investigates the feasibility of using RST trees to automatically identify and annotate implicit discourse relations in texts that only have shallow, explicit PDTB-style annotations. This would provide a richer understanding of text coherence without the cost of manual annotation for implicit relations.
Methods
The core of this work is a mapping procedure between RST trees and PDTB-style annotations. The authors use the Potsdam Commentary Corpus (PCC), which has both RST and PDTB-style annotations on the same German news commentary articles.
- Data Preparation: They filter RST trees to include only non-complex relations, as complex relations involve multiple EDUs and complicate direct mapping. This results in 2,111 non-complex RST relations compared to 1,110 explicit PDTB-style relations in the corpus.
- Segment Matching: The primary method involves comparing Elementary Discourse Units (EDUs) from the RST layer with arguments (arg1 and arg2) from the PDTB layer. A "complete match" is defined as a significant token overlap (greater than 75%) between an EDU and a PDTB argument.
- Scenario Categorization: Three scenarios are distinguished:
- Complete Match: An RST EDU matches both arg1 and arg2 of a PDTB relation.
- Partial Match: One EDU matches one argument, but the other EDU does not match the other argument.
- No Overlap: No significant overlap between any EDU and any PDTB argument.
- Focus on Implicit Relations: The paper prioritizes cases with "no overlap" as the most promising candidates for automatically inferring implicit relations, as they represent coherence structures present in RST but not explicitly marked in PDTB.
Datasets
- Potsdam Commentary Corpus (PCC): A German corpus of news commentary articles. It contains approximately 33,000 words, with 176 RST trees (3,018 EDUs) and 1,110 explicit PDTB-style relations. The RST layer uses a modified version of Mann and Thompson's (1988) constraints, and the PDTB layer annotates explicit relations based on Pasch et al. (2003).
Key Insights and Ideas
- Bridging Frameworks is Feasible: The study demonstrates that it's possible to map between RST and PDTB annotations, offering a path to enrich shallow annotations with implicit relations.
- Implicit Relations are Abundant: The PCC corpus, like others, shows a significant number of implicit relations in RST (63% of RST relations are implicit, as noted from prior work on the same corpus), highlighting the value of inferring them.
- Segmentation Differences are Crucial: A key challenge is the inherent difference in how RST and PDTB define segments (EDUs vs. arguments). EDUs are fundamental to RST segmentation, while PDTB arguments are defined by explicit connectives and a "minimality principle." This discrepancy can lead to situations where segments don't align perfectly, even if they refer to related discourse units.
- "No Overlap" Cases are Goldmines: The "no overlap" category, where RST EDUs and PDTB arguments don't align, is identified as the most fertile ground for discovering implicit relations. These are cases where RST captures a connection that is not explicitly signaled in the PDTB layer.
- Partial Overlap Requires Careful Handling: Cases where one EDU matches one argument but not the other are complex. They might represent different relations or simply different ways of segmenting the same underlying coherence.
Key Results
- Out of 2,111 non-complex RST relations analyzed, 305 resulted in a "complete match" between RST EDUs and PDTB arguments.
- 323 cases fell into the "partial match" category.
- The largest category, 1,483 cases, showed "no overlap" between RST EDUs and PDTB arguments. These are the primary candidates for inferring implicit relations.
- A manual check on 17 documents (10% of the corpus) found 64 RST relations of the "no overlap" type. Of these, many (21) showed partial overlap below the 75% threshold, indicating segmentation differences.
- The paper identifies 27 cases where there was no PDTB-style annotation at all, making them prime candidates for automated addition of implicit relations.
Limitations and Future Work
- Exploratory Nature: The current approach is described as "high-precision, but relatively low-recall" for finding implicit relations.
- Partial Overlap Complexity: The paper acknowledges that partially overlapping relations require further investigation, as they might represent entirely different discourse relations.
- Segmentation Discrepancies: The fundamental differences in segmentation between EDUs and implicit relation arguments need more in-depth study.
- Future Work:
- Investigating the 323 "partial match" cases.
- Developing methods to handle cases where EDUs and arguments don't align perfectly.
- Exploring more complex mapping algorithms, such as those incorporating the Strong Nuclearity hypothesis.
- The authors suggest that a theory-neutral standard for discourse segmentation could be highly beneficial for cross-theory annotation augmentation.
Summary updated