OEMG: OOV-Enhanced Multi-Task Generative Framework for Out-Of-Vocabulary Problems in Slot-Filling Task

Recently, the application of generative frameworks in dialogue systems has attracted considerable attention. However, in practical dialogue systems, existing slot-filling methods often encounter Out-of-Vocabulary (OOV) problems, resulting in a breakdown of semantic understanding in utterances and consequently diminishing the model's generalization capability. To address this limitation, we propose an OOV-enhanced multi-task generative framework named OEMG. Specifically, in the OOV pre-training stage, we introduce two auxiliary tasks for the OOV-word problem namely OOV-word correction (OC) and OOV-word boundary enhancement (OBE) to boost the performance on the OOV-word problem from word and span levels. As for the OOV fine-tuning stage, we design the OOV-slot context enhancement (OCE) task to explicitly enhance contextual semantics. Our approach achieves state-of-the-art results on both OOV-word and OOV-slot problems in two benchmarks.