Multi-domain adaptation for cross-domain semantic slot filling

Slot filling is a crucial sub-task in the field of Spoken Language Understanding and aims to match the corresponding semantic slot for each word in the sequence. Slot prediction in an unknown domain requires a large amount of data in the domain for training, but in reality, there is often a lack of trainable samples in the unknown domain, which makes it difficult for the model to predict new domains. This is the biggest challenge of the cross-domain slot filling task. In recent years, the idea of transfer learning has been applied to cross-domain slot filling tasks. The current training method directly mixes the source domain data samples without considering the differences between the various domains in the source domain, which ignores the domain-invariant features contained in the source domain. In this paper, we proposed a cross-domain slot filling model based on multi-domain adaptation. First, we used the domain-adaptive domain projection layer to let the feature learner classify the domain-invariant information and domain-exclusive information into the specified dimension part of the vector, so as to realize the extraction of domain-invariant feature information, and then used the trainable linear transformation matrix to relieve the generalization burden of the feature learner. Experimental results show that our proposed models significantly outperform other methods on average F1-score.