Ensembling Heterogeneous NLU Models via Knowledge‐Driven Constraints for Task‐Oriented Dialogue Systems

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Intent classification and slot filling are essential tasks in natural language understanding. The latest approaches consider these tasks jointly in order to leverage their interdependence but use single models for training and inference, which struggle to capture fine semantic distinctions and contextual dependencies in diverse input patterns. Recently, the integration of expert‐defined constraints into these models has gained traction as a promising approach, effectively bridging the gap between data‐driven methodologies and expert knowledge. In this paper, we advance this hybrid intelligence approach by employing model ensembles that combine the outputs from two models using expert‐defined constraints. The proposed approach was evaluated using combinations of four state‐of‐the‐art models (Bi‐model, DCA‐Net, E2EMG‐CRF and AGIF) across three widely used benchmark datasets (ATIS, SNIPS and NLU‐Benchmark). The results consistently indicate substantial improvements in all metrics, including intent accuracy, slot F1‐score and semantic accuracy. The magnitude of the improvement depends on the dataset and ensemble configuration, with average absolute improvements in semantic accuracy ranging from approximately 0.7% to 1.8% points across the evaluated datasets and the largest improvement exceeding 6% points.