Relation and target prediction in multilingual argument mining

Improve relation and target prediction for argumentative components in Swiss4Arg, including identifying the correct direction of argumentation and the precise target segment, even when component labels are provided correctly.

Background

Swiss4Arg evaluates argument component classification and joint relation-plus-target prediction across German, French, Italian, and Romansh using GPT-5.4 and SwissBERT. Although both systems perform relatively well on component classification, relation and target prediction remains substantially more difficult, including under oracle component labels where errors from component classification are removed.

The unresolved challenge concerns both determining whether an argumentative unit supports or attacks another unit and identifying the specific target segment. The reported exact-match performance is no higher than 0.485 under correctly classified components, indicating that relation prediction requires document-structure understanding beyond component identification.

References

The relation+target task remains an open challenge even under oracle component labels, that is, only evaluating relation+target labels if the component prediction is correct, (exact match $\leq 0.485$), highlighting that identifying the correct direction of argumentation and the precise target segment requires a deeper understanding of document structure than GPT-5.4 currently exhibits in a zero-shot setting.

— Swiss4Arg: A Parallel Dataset for Multilingual Argument Mining across Swiss National Languages  (2610.17894 - Conrad et al., 13 Oct 2026) in Section 5, Baseline Argument Classification, subsection “Discussion”