Researchers propose ReaORE, a framework for open relation extraction that utilizes large reasoning models to achieve reliable generalization to unseen relation types. The method addresses limitations of current clustering and direct generation approaches through a coarse-to-fine reasoning process.
- The framework consists of two stages: relation filtering and relation prediction.
- Relation filtering reasons over multiple aspects to understand relations and instances, yielding an initial set supplemented by embedding-based similarity.
- Relation prediction uses fine-grained comparative reasoning to distinguish easily confused relations from the filtered set.
- Extensive experiments on two widely used OpenRE datasets demonstrate that ReaORE outperforms existing baselines.
The authors consider this approach important because it effectively overcomes the poor generalization of clustering techniques and the discriminative capacity issues of direct LLM generation in real-world applications.