The Research Of Spoken Language Understanding Based On Few-shot Learning

As the first process of task-based dialogue systems,spoken language understanding has great research significance in the field of dialogue systems.At present,the research of spoken language understanding is mostly based on deep learning.A large amount of labeled dialogue data is needed in each dialogue field to support the training of models,resulted in considerable data costs.The few-shot learning proposed for data problems is mainly used in image classification tasks and is still in its infancy.This paper applies the few-shot learning to spoken language understanding and presents models for small sample data for the subtasks of spoken language understanding-intent determination and slot filling.Combining the two models,a joint training model for spoken language understanding is proposed.The few-shot intent determination model proposed in this paper adopts the model structure of encoder-induction-relationship,applying Bi-LSTM and Attention to encode the sample semantic vector,and the principle of capsule network to the model algorithm with the use of dynamic routing for class prototype vector calculation.The model is trained and tested on Few Joint,the latest few-shot data set.The intent determination accuracy of this model has reached 78.04%,demonstrating a substantial improvement compared to the baseline model.This paper proposes a few-shot slot filling model.Based on the principle of CRF algorithm,it adopts the structure of transition probability and emission probability.It proposes an "essential transition matrix" to obtain cross-domain shared transition probability information on the entire source domain,and uses Tap Net to Calculate the emission probability score.The model was trained and tested on the same data set,with final result of 67.11% of the slot filling F1 value,greatly exceeded the performance of the baseline model on the same data set.Finally,the author combines the few-shot intent determination model with the few-shot slot filling model,and proposes a few-shot spoken language understanding joint model and a model structure of common encoder-respective induction-fusion scoring,and at the same time,innovatively merging the intent determination results of the query samples into the similarity calculation of the slot filling task.Consequently,this model achieved 79.74% accuracy of intent determination and 72.57% of slot filling F1 value on the data set demonstrating that joint training can effectively improve the performance of the these tasks.