C-BERT: Factorized Causal Relation Extraction

Existing causal relation extraction (CRE) systems prioritize binary (cause, effect) tuples, discarding strength distinctions. We introduce C-BERT, a multi-task transformer that extracts fine-grained relations as (C, E, I) triples, where I E [-1, +1] encodes signed influence. Our key contribution is a factorized architecture decomposing influence into three linguistically motivated classification heads – role, polarity, and salience – coupled with a negation-aware annotation framework that distinguishes indicator-based, propositional, and object negation. On German biodiversity discourse (2,391 relations), the factorized model consistently outperforms unified classification across five random seeds (0.768 ± 0.009 vs. 0.744 ± 0.007 reconstructed accuracy). Decomposed error analysis shows that factorization reduces multi-head error cascades and concentrates failures in single, interpretable subtasks. We release both model variants, code, and a data subset of German parliamentary debates.