Correction Gated by Entropy and Committee Disagreement for Argument Component Classification

This thesis addresses argument component identification (ACI) and classification (ACC) for the ANR CIGAIA project, which studies argumentation about the war in Ukraine in news discourse. As the CIGAIA corpus is only partially annotated, experiments use MDAC, a 486-document corpus combining UN Security Council speeches, news editorials and persuasive essays. A two-stage pipeline, where one encoder-only transformer tags argumentative spans and a second classifies them as claims, premises or false positives, so that spans wrongly identified by the first stage can be rejected, is compared with a joint single-stage baseline. On top of it, a lightweight correction mechanism turns further spans into false positives using the ACI stage's predictive entropy and flips claim/premise labels using the votes of a five-member committee trained on different data folds, with thresholds selected on validation data only. All four RoBERTa/Longformer (ACI) and BERT/DeBERTaV3 (ACC) pipelines outperform the joint baseline (0.5776 test macro-F1) before correction, and the best corrected pipeline reaches 0.6218. The correction's gain, however, depends on the backbone pairing, from +0.18 to +1.36 points, and is significant at every seed only for RoBERTa+BERT; no tested variable explains this dependence. Future work includes applying the pipeline to CIGAIA data and to user-generated text.