Entity disambiguation links mentions in text to the entities they refer to in a knowledge base. Recent cross-entity interaction models improve candidate comparison by jointly encoding multiple candidates, but incorporating entity metadata into this joint encoding is not straightforward. Appending metadata to every candidate forces joint encoding to process longer inputs with detailed distractor information, which can weaken the contrastive cues that make cross-entity interaction effective. We propose DuCEI, a dual-stage framework that avoids this broad metadata interaction by reducing the candidate set before introducing entity metadata. DuCEI first compares all candidates using title-based cross-entity interaction, and then performs metadata-enriched interaction only among the remaining candidates. This design makes metadata use selective at the candidate level, applying detailed evidence only where it is likely to help. Experiments on six standard benchmarks show that DuCEI achieves the best overall performance among models trained solely on AIDA-CoNLL. Further analyses show that confidence-based selective refinement preserves nearly all of the full two-stage performance while applying Stage 2 to only 27.80% of mentions, reducing inference cost from 1.90× to approximately 1.27× that of Stage 1 alone. Our code is available at https://github.com/jinwoomin/DuCEI.

This paper describes the Jeonbuk National University (JBNU) system for the 2020 shared task on Cross-Framework Meaning Representation Parsing at the Conference on Computational Natural Language Learning. Among the five frameworks, we address only the abstract meaning representation framework and propose a joint state model for the graph-sequence iterative inference of (Cai and Lam, 2020) for a simplified graph-sequence inference. In our joint state model, we update only a single joint state vector during the graph-sequence inference process instead of keeping the dual state vectors, and all other components are exactly the same as in (Cai and Lam, 2020).