Co-adaptive Rule Routing and Constraining for Structured Knowledge Retrieval

Co-adaptive Rule Routing and Constraining for Structured Knowledge Retrieval. By Zhisheng Qi and eleven others, from the University of Georgia, Michigan State University, and five other institutions. Published on the seventeenth of October. No code repository is linked to it.

Structured knowledge retrieval pulls task-relevant evidence out of structured knowledge sources so that a downstream task can use it. Existing methods take one of two routes. Some retrieve without constraint. Others rely on predefined supervision that pairs each kind of query with a rule, such as an author to paper rule for scientific discovery, or an anatomy to drug rule for precision medicine. That second route assumes someone has already written down which rule fits which query, and in practice such supervision is rarely available.

Co-Adaptive Rule Routing and Constraining Retrieval, or CoRE, learns both halves of that problem together. It learns which reasoning rule to assign to each query, and it learns how to retrieve evidence under that rule, without ever being given predefined query-rule pairs.

Two components share the work. A routing agent selects a rule for the query. The retriever belonging to that rule then traverses the knowledge base and ranks evidence that satisfies the rule's constraints. The two depend on each other: a good rule is only useful with a retriever that can follow it, and a retriever can only be trained once it knows which rule it is serving.

CoRE handles that dependence by alternating. The routing policy is trained with reinforcement learning, and the rule-constrained retriever with contrastive learning. Each supplies the other's training signal. The rules the router selects become supervision for the retriever, and how well the retriever performs becomes the reward for the router. That closed loop is what lets the two co-adapt without explicit query-rule supervision.

Experiments use STaRK-QA datasets spanning three domains: scientific discovery, biomedical reasoning, and e-commerce knowledge management, and the authors report that they demonstrate CoRE's effectiveness. Ablations show a benefit from co-adapting router and retriever rather than training them apart. Routing done by generation transfers across datasets, and it benefits from pretraining on multiple domains before fine-tuning on the target one.

Because the router has to pick a rule for every query, its choices form a record of which rules are used and where. The authors turn that record to a further purpose: analysing rule usage across queries informs automatic refinement of the rules and the discovery of new ones. The rule set, in other words, need not stay fixed at the one someone wrote down at the start.