📄 Automatically Identifying the Arguments of Discourse Connectives

Summary

The paper tackles a core step in discourse parsing: given an explicit discourse connective such as "because", "however" or "and", find its two arguments (Arg1 and Arg2) in the text. Working with the Penn Discourse TreeBank (PDTB), the authors avoid predicting exact argument spans, which are often discontinuous and hard to delimit. Instead, they recast the task as finding the head word of each argument.

The system has two stages:

• Candidate ranking. For each connective, separate log-linear ranking models score candidate head words for Arg1 and for Arg2.
• Joint re-ranking. A log-linear re-ranker then picks the best (Arg1, Arg2) pair from the top candidates, capturing dependencies between the two arguments.

Features are lexical, syntactic and dependency-based. They are drawn from Penn Treebank constituency trees, including dependency paths derived from those trees, and they are conditioned on connective type. Arg2, which is syntactically bound to the connective, is found far more reliably than Arg1. Joint re-ranking improves the accuracy of getting both arguments right to roughly 74%. The work shows that a head-based, syntax-driven formulation makes argument identification for discourse connectives tractable.

Key points

• Data. The PDTB annotates explicit connectives with two arguments over Wall Street Journal text. Its annotation layers on top of Penn Treebank syntax.
• Head-based targets. Arguments are represented by their syntactic head words rather than full spans. This simplifies both learning and evaluation.
• Asymmetric difficulty. Arg2 is usually local to the connective and identified with high accuracy. Arg1 can sit far away, sometimes in an earlier sentence, especially with adverbial connectives, which makes it much harder.
• Pipeline.
  • Independent log-linear rankers propose head candidates for each argument.
  • A re-ranker scores N-best argument pairs jointly.
  • Joint accuracy of both arguments reaches about 74%.
• Feature analysis. The paper compares feature families (lexical, constituency, dependency, connective-type). Dependency-path features derived from constituency trees are especially helpful.

Interesting ideas

• I1. Replacing hard-to-delimit argument spans with argument heads turns a fuzzy segmentation task into a cleaner ranking task over words, and spans can be recovered later from the tree.
• I2. Rank candidates for each slot independently first, then re-rank the top combinations jointly so that interactions between the slots can be modelled.
• I3. Conditioning features on connective type (coordinating, subordinating, adverbial) reflects that each class places its arguments differently, so one generic model would underperform.
• I4. Paths through a dependency representation derived from constituency trees proved more useful than raw constituency features for linking distant words.
• I5. Reporting Arg1 and Arg2 accuracy separately, alongside joint accuracy, exposes where the real difficulty lies instead of hiding it in one number.

For Mova

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

• I1 (frames): Discourse relations between frames could be anchored on the head predicates of clauses in the UD tree, linking frame to frame without guessing span boundaries.
• I2 (world): Linking events by causal or contrastive connectives could use a two-step choice, where each event argument is shortlisted separately and the pair is then checked for consistency before it is written into the event order.
• I3 (frames): Prag could keep small per-class rules or features for connectives, for example looking for an adverbial's Arg1 in the previous sentence, keeping the decision explainable.
• I4 (tree): Mova's UD trees already provide such paths, so the dependency path from connective to candidate head is a natural, transparent feature for any discourse-linking step.
• I5 (learn): Evaluations of discourse or relation linking should break down scores per argument slot and per connective class, so that self-induced repair rules can target the weak slot.