| In recent years,human-computer interaction has slowly entered people's lives,intelligent audio,intelligent car navigation and other products have brought great convenience to people's lives.The technology behind human-computer interaction,in addition to speech recognition technology that converts speech to text,also includes the spoken language understanding that translates text into user's intentions.Therefore,the accuracy of spoken language understanding directly affects the quality of the human-computer interaction experience.So it is very helpful to study the spoken language understanding to improve the human-computer interaction experience.Spoken language understanding task is essentially a typical sequence labeling problem that each word in a sentence is labeled as a semantic tag.The recurrent neural network is particularly suitable for dealing with sequence problems,so spoken language understanding is modeled by recurrent neural network in this paper.In order to improve the accuracy of the spoken language understanding,this paper researches from three aspects of data augmentation,improvement of recurrent neural network model and grammar information fusion.The main work and contributions are:1.Because the public data set for the study of spoken language understanding is relatively small,it is easy to cause overfitting when training the model.In order to alleviate the problem of overfitting when the dataset is too small,a method of fine-grained data augmentation based on word is proposed.The method can use the redundant information in the sentence,and can effectively increase the amount of data by removing randomly the words that do not affect the semantic expression.The principle of this method can effectively alleviate the over-fitting problem is that on the one hand,it only removes the words in the sentences that do not affect the semantic expression to achieve data augmentation,on the other hand,at the same time it brings some random noise to the sentences,which can be effective to alleviate the over-fitting problem.2.A method to enhance the context information of the recurrent neural network is proposed,it is called a windowed recurrent neural network in this paper.Although the recurrent neural network has the ability to memorize historical information,it cannot record long history information due to a serious problem of gradient decenting.In order to improve this problem,a new windowing method for the hidden layer of recurrent neural network to enhance its memory and use of contextual information is proposed.This method is not only effective for simple recurrent neural networks,but also effective for long short-term memory networks with stronger memory.3.A method of fusing grammatical information into the spoken language understanding system is proposed.Grammatical information plays a crucial role in understanding the semantic expression of a sentence.Although the grammatical structure information is extracted implicitly when constructing a model through a neural network,this method does not make full use of grammatical information.At the same time,considering that in the previous studies of semantic understanding of spoken language,grammatical information was not explicitly added to the model of spoken language understanding,so this article will use the parallel,serial and mixed three different fusion methods to integrate the grammatical structure information into the model of spoken language understanding.Experiments show that the mixed fusion method can effectively improve the accuracy of spoken language understanding. |
