An Explicit-Memory Few-Shot Joint Learning Model

There are two difficulties in existing spoken language understanding models. The first problem is that it is difficult to extract the implicit relationship information between the intention and the slot in the utterance for the inference process, and the inference effect is not ideal; the second problem is that the training data is scarce, and the existing models cannot learn from the small amount of training data. Get more useful information. To address these two challenges, this paper proposes an Explicit-Memory Few-shot join learning model. In order to solve the first problem, a multi-layer model structure from coarse to fine is adopted to train the hidden semantic relationship and hidden state information between intentions and slots in the utterance; in order to solve the second problem, using the Siamese BERT metric learning method to jointly train the model. We use the Snips and ATIS datasets to train the model, and the test results show better results. In the case of a small amount of data, the model can also obtain stronger inference ability.