Learning Probabilistic Graphs from Natural Language

Natural-language descriptions contain knowledge about uncertain systems, but using that knowledge in probabilistic models requires an explicit account of variables, states, dependencies, and probabilities. Recovering these elements together is difficult because errors in one component can change the meaning of the others. We introduce DUAL-BN, a two-stage framework for learning this translation through bidirectional supervision and structured feedback. A shared language model first learns to extract probabilistic graph components from text and reconstruct descriptions from graphs. Group Relative Policy Optimization (GRPO) then refines jointly generated components using reference-graph fidelity, semantic consistency of the text–graph–text cycle, and a penalty for component counts exceeding the reference. The probability targets are parent-specific pairwise conditional tables. On PRISM-BN backbones unseen during training, DUAL-BN increases Node F1 from 0.77 to 0.91 and mapped Edge F1 from 0.68 to 0.84 over supervised fine-tuning, while reducing forward KL divergence on matched pairwise tables from 0.87 to 0.66. The trained policy is evaluated through staged component queries with reference-assisted filtering. The gains accompany a decrease in conditional State F1. A separate study of composition from supplied partial graphs examines how such representations can support larger probabilistic backbones. On ten descriptions rated by three evaluators, mean overall reconstruction quality rises from 3.4/5 for SFT to 4.1/5 for DUAL-BN. The results support combining token-level supervision with feedback on the structure and recoverable information of graph predictions as a step toward building probabilistic models from language.