Distill-AER: Fine-Grained Address Entity Recognition from Spoken Dialogue via Knowledge Distillation

Fine-grained address entity recognition from spoken dialogue is an important but challenging task because there are multiple types of address entities distributed across the multi-round dialogue context. Existing work typically formulates this problem as a fine-grained named entity recognition task, which in our scenario suffers from a high cost of training data annotation. On the other hand, large-scale full standard addresses could be easily crawled from the web pages like Google Maps and annotated with fine-grained address tags with limited human effort. Leveraging this, we propose a distillation approach (Distill-AER) for transferring knowledge from the large-scale labeled full standard address dataset to the fine-grained address entity recognition task in a spoken dialogue context scenario. We further construct a labeled spoken dialogue dataset with address entities using the data augmentation paradigm we proposed, which could benefit future research. Experimental results show that Distill-AER significantly outperforms other competitive baselines.