Research On Slot Filling Method Based On Deep Ensemble Learning

The human-computer dialogue system usually consists of three parts: dialogue understanding,dialogue management and dialogue generation.Slot filling is an important task of dialogue understanding.Its purpose is to extract key information related to the task from the dialogue and form a structured semantic representation in the form of "slot-value".The quality of the slot filling has an important impact on the quality of the human-machine dialogue system.In recent years,it has been widely concerned by researchers in various fields,and has gradually become a hot spot in the field of human-machine dialogue system research.Traditional slot filling methods are mostly based on rule matching and statistical learning.Rule-based matching methods add semantic and grammatical rules on the basis of string matching,and extract relevant attributes through rule matching.These rules usually require domain experts Come to formulate,have a strong dependence on background knowledge in the professional field.Methods based on statistical learning usually use artificially labeled data sets for model training and adjust the model parameters by optimizing the loss function.The disadvantage of such methods is that a large number of manually labeled data samples are required.In recent years,deep learning has been widely used in slot filling tasks and has achieved many meaningful results,but due to the complexity of the problem itself,there are still many problems in the field that need to be solved,such as the slot filling model cannot be well expressed The meaning of things changing with time,insufficient sample learning during training,gradient disappearance and gradient explosion.In view of the above problems,this paper proposes a slot filling method based on deep integrated learning.By using the encoder-decoder model that introduces the attention mechanism,it learns different distributed samples,and uses the idea of integrated learning to improve the performance of the model.Experimental results show that,compared with other methods,the method proposed in this paper has more performance advantages in the slot filling task and can be effectively applied to the slot filling task.The main contributions of this article are as follows:1.Propose a slot filling method based on attention mechanism and encoder-decoder structure.Aiming at the problems of gradient disappearance,gradient explosion and inability to deal with time series well in RNN and its variants,an encoder-decoder model with attention mechanism is proposed.The model uses Bi GRU as the encoder and unidirectional GRU as the decoder to extract sequence features with comprehensive semantic information.At the same time,an attention mechanism is introduced to highlight the key features of the input sequence,and then a decoder model based on the attention mechanism is constructed to obtain better performance in slot filling.2.Propose a strategy of slot filling integration based on spectral clustering.Aiming at the problems that the single model is not enough for sample learning and the slot filling performance needs to be improved,a dynamic classifier integration selection method is proposed based on spectral clustering.This method uses a encoder-decoder model based on the attention mechanism to train differently distributed samples to generate a set of classifiers with different sample characteristics.Based on the spectral clustering method,a suitable classifier is dynamically selected from the classifier set.Each type of sample in the test set is predicted,and the performance of slot filling is further improved by means of voting fusion.The existing method and the method proposed in this paper are compared on the ATIS data set and Sql query statement data set.The experimental results show that compared with some existing methods,the method proposed in this paper can further improve the performance of slot filling.