Joint Model of Intent Recognition and Slot Filling Based on Graph Neural Network fusion of external knowledge base
Hairui Huang, Xiwei Feng, Ziyue Wan · 2024
Pretrained language processing models have shown significant results in spoken language comprehension tasks, particularly in joint modelling of intention recognition and slot filling. However, these models lack external knowledge support for complex and precise reasoning tasks, which is a limitation compared to human language understanding. This paper proposes a joint model for spoken language comprehension that utilizes the BERT pre-trained language model and combines the graph attention mechanism as a coding layer to semantically represent the input text. BiLSTM is used to achieve the fusion of external knowledge. The model is based on BERT fusion of external knowledge base. The joint model captures the association between the tasks of intention recognition and slot filling through joint training. Additionally, an external knowledge base is introduced to enhance the model's performance. Empirical analyses support that the unified model achieves exceptional accuracy in both intent recognition and slot-filling areas, outperforming other benchmark models.