📄 DM Doppelgangers: Implicit Connectives as eRST Signals

Summary

RST-style hierarchical discourse annotation and PDTB-style shallow annotation have mostly been studied separately. eRST now links RST relations to token-anchored signals such as explicit connectives, but it has no notion of an implicit connective, the unexpressed "but" or "so" that PDTB annotators insert between segments. This paper connects the two. The authors take the GUM corpus (v12), which carries eRST annotation, and its PDTB-style counterpart GDTB, which was semi-automatically converted from GUM and contains implicit relations. They design an alignment algorithm that does one of two things with each GDTB implicit connective:

• it attaches the connective as a new signal type on a compatible relation in the primary RST tree, or
• when no tree relation fits, it adds a tree-breaking secondary edge that the connective then signals.

The result is the first dataset of eRST relations aligned with implicit connectives, plus extra relations that PDTB covers but RST trees miss. The paper also gives a first analysis of how implicit connectives co-occur with other signal types. This matters because it brings deep and shallow discourse frameworks into one graph.

Key points

• Data: GUM Version 12, with 16 spoken and written English genres and over 290K tokens. It is described as the largest RST dataset annotated for both discourse markers and non-DM signals. All implicit connectives come from GDTB, whose implicit relations score about 0.94 F1 on argument spans and about 0.85 on exact PDTB labels.
• Alignment uses two devices to handle span mismatches between the frameworks:
  • Strong Nuclearity, which lets a relation between spans also hold between their nuclei;
  • conversion of multinuclear relations into right-to-left dependencies, following Li et al. 2014.
• Counts: an implicit connective marks 6,365 relations, while an explicit DM marks 8,300. Implicit connectives co-occur with non-DM signals in 57% of cases (3,627), against 26% (2,192) for explicit DMs.
• A chi-square test over seven signal categories is significant (χ²(6, N=5819)=301.46, p<.001, V=0.242). Implicit connectives come with reference signals more often and with morphological signals much less often.
• "then" and "in order" mostly attach to existing relations, while "in fact" more often produces new secondary edges. The eRST signaling toolkit (87.9% accurate) added non-DM signals to 30% of the new implicit secondary edges.

Interesting ideas

• I1. When an inferred connective does not fit the existing tree relation, the algorithm keeps the tree and adds a separate secondary edge for the local relation. This preserves both the global intent structure and local semantics without forcing a choice between them.
• I2. An implicit connective is a test: try inserting "but", "so" or "and" at a junction and see which one fits. This turns an abstract relation label into an explainable lexical paraphrase.
• I3. Candidate junctions for implicit relations are found from syntax trees: adjacent sentences in a paragraph, coordinated predications in one VP, and adverbial clauses without a conjunction.
• I4. Strong Nuclearity lets annotations with mismatched spans be aligned by comparing head nuclei instead of full spans. This is a general way to reconcile two segmentations of the same text.
• I5. Implicit relations rely more on reference cues (pronouns, demonstratives, comparatives), so anaphora is a strong hint that two segments are linked even without an overt marker.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (world): Event-order and causal links in the text world could keep one primary narrative structure plus extra "local" causal or result edges whenever an implied connective suggests a relation the main structure contradicts.
• I2 (frames): The pragmatics layer could justify each inter-clause relation by naming the connective it would insert, making implicature between sentences inspectable.
• I3 (tree): The UD tree can mark these junction types deterministically (e.g. conj between verbs, advcl lacking a mark), feeding a discourse step with well-defined slots.
• I4 (learn): When several witness parsers or teacher tiers segment clauses differently, agreement can be scored on head units rather than exact spans before building silver data.
• I5 (frames): Coreference links between adjacent sentences could raise the prior that a discourse relation holds there and prompt the system to pick an implied connective.