ARI’s relation inference architecture is a layered cascade of deterministic logic and probabilistic models. It combines symbolic structure with learned similarity to infer high-quality mappings between schemas.
ARI inference architecture
Starting from heterogeneous source and target schemas, ARI first normalizes structure into shape graphs (shaping). It then generates candidate correspondences, computes lexical, semantic, and structural features, and scores candidates using learned models. Higher-order components—such as cross-encoders, CRFs, and graph-based methods—promote structural consistency, while a constraint solver produces a globally optimal, consistent mapping. Human feedback and ontology signals continuously refine both the models and the output.
ARI trains a unified model over four structured layers of training data, progressing through pretraining, supervised alignment, and fine-tuning. Each layer is incorporated with controlled weighting, scheduling, and task-specific objectives.
Training dataset layers & training stages
Pretraining establishes structural and semantic priors using specifications, ontologies, and synthetic data. Supervised alignment refines the model using verified mappings and curated hard negatives. Fine-tuning incorporates human feedback and mined negatives to align ranking behavior with real-world usage. Across stages, ARI dynamically balances data sources and objectives to maintain global structural consistency while improving both recall and precision.
Training data from mappings, user feedback, and synthetic generation is unified into a canonical stream of training examples. These examples are normalized, enriched with lightweight features, and projected into model-specific datasets for each component of the inference stack.
Training data flow
All training signals, whether derived from production mappings, human feedback, or synthetic generation, are first captured as structured training examples. This representation provides a consistent foundation for feature computation, including lexical, structural, ontological, and embedding-based signals. A projection layer then transforms examples into the formats required by each model, allowing a single pipeline to support heterogeneous learning objectives across the system.
The following diagram depicts the internal data model of the ARI training system.
ARI Training Data Model