Research And Application Of Modified Neural Turing Machine With Deep Reinforcement Learning

As the rapidly development of research on machine learning and artificial intelligence,machines can understand more complex knowledge.Natural language Processing(NLP)is a hot research field,which aims to make machines understand native human languages.Neural network is very effective algorithm in NLP,and as a kind of modified model of the neural network,Neural Turing Machine(NTM)makes many significant improvements.On the other hand,Deep Reinforcement Learning(DRL)gives machines the ability to learn from failures so that the algorithm can optimize continuously,and the DRL models even defeat humans in some tasks.Considering the advantages of the above two models,and focusing on the shortcomings of the traditional NLP models,this thesis studies the attention based NTM model,and the application of the DRL model in NLP area.The main three contents of the research are list as below:Firstly,this thesis proposes the Attention based Neural Turing Machine model.aiming at the problem of the poor memory in long sentences and the lack of core words' influence.By using the NTM to combine the sentence sequence,the model can handle the logical and long-term-memory task.At the same time,the attention layer over the hidden features of the sentence,which calculates weights at the word level,can strengthen the influence of the keywords and extract their information.Secondly,this thesis proposes the Deep Reinforcement Learning based Sentence Structure Generation model,aiming at the problem of the difficulties in sentence structure adjusting.By using the DRL to optimize the generation strategy,the model filters the words in sentences,and generates the sentence structure according to specific contexts and NLP tasks.On the other hand,the generation model is trained based on the prediction of the NTM model,so the two models are combined as a dual form,which can be trained and improved together.Finally,this thesis combines the proposed two models,and implements a public opinion analysis system based on the proposed DRL-NTM model.The model is reconstructed and packaged to fit the application.The system can perfectly complete the collecting,processing,analyzing and persisting of the public opinion data.