Can We Do Interpretable NLI with Graphs Based on Atomic Propositions?

Summary

This paper asks whether natural language inference can rest entirely on interpretable structures. Sentences are decomposed into atomic propositions (MPropositionneur-V2-large), converted into ConceptNet triples by Qwen3.5-9B under a constrained JSON schema with a 28-relation enum, and rendered as three graphs per pair — premise, hypothesis, and a retrieved ConceptNet subgraph — serialised in a fixed order for a fine-tuned Qwen3.5-0.8B-Base classifier that never sees raw text. On SNLI the pipeline reaches 89.7% (three-seed mean ± 0.006), 1.9 points below an identically trained text model; combining both modalities gives 92.1%. On ANLI (R1–R3, with SNLI replay) it trails its text counterpart by 9–14 points, though it matches published RoBERTa-large on R2 and R3. A scaling experiment — adding roughly 600,000 MNLI and FEVER-NLI training pairs yields only +0.6 points on ANLI — supports the claim that the gap is representational rather than a data problem. Ablations show the retrieved knowledge is neutral on SNLI. Main limitations: single-seed text baselines and ANLI runs, two interrupted graph seeds (the 1.9-point gap sits within run-to-run variation), no human validation of interpretability, no direct evaluation of graph quality, and the no-K ablation untested on ANLI.