Learning When to Realize Retrieved Entity Information in Retrieval-Augmented End-to-End Task-Oriented Dialogue

Retrieval-augmented end-to-end task-oriented dialogue systems improve response generation by providing relevant database evidence to the generator. However, relevant evidence alone does not indicate whether entity information is required in the current response. This can lead to missing required entities or unnecessary entity mentions. To address this problem, we propose a response-level entity realization framework. An entity realization predictor combines the dialogue representation, the current attribute state, and newly activated attributes across turns to estimate whether entity information is needed. The predicted score controls an Entity Realization Adapter through a soft gate, adjusting the generator encoder representation at each dialogue turn. We also introduce an entity-focused generation loss to strengthen supervision at entity-related target positions. Although BLEU decreases slightly in the full knowledge-base setting, experiments on MultiWOZ 2.1 show that the proposed method achieves higher mean Entity Recall than the reproduced MAKER baseline under both condensed and full knowledge-base settings. These results indicate improved coverage of required entity information.