This work investigates whether natural language inference can be performed using only interpretable graph‑based evidence representations. We propose a fully graph‑centric pipeline where the classifier never sees raw text. Sentences are first broken into atomic propositions, then converted into ConceptNet triples via constrained decoding, yielding three graphs for each pair: a premise graph, a hypothesis graph, and a retrieved ConceptNet subgraph. These graphs are fed to a fine‑tuned 0.8‑billion‑parameter language model for classification. On SNLI the system reaches 89.7% accuracy, just 1.9 points below an identically trained text‑based model. On ANLI it matches RoBERTa‑large on rounds R2 and R3 (≈50% accuracy) but lags 16 points on R1, resulting in an overall gap of 9‑14 points compared to its text counterpart. We call this the price of interpretability and show through ablations that the gap stems from representational limits rather than data scarcity. Additional studies reveal that graphs and text are complementary; their combination attains 92.1% accuracy on SNLI.
Review