Research On Punctuation Prediction Based On Deep Learning

With the continuous progress of computer technology,human-machine interaction develops from single word interaction to text,gesture and voice.The increase of interaction mode brings more convenience to human-machine communication,but also brings enormous challenges to computers.At present,the speech interaction in simple scene is the direct use of speech signal recognition.Meanwhile,in the vast majority of application scene,the speech signal is converted to text and then analyzed accordingly.But now the transcriptional text of the voice engine does not contain punctuation,and punctuation plays an important role in the expression of human emotion.The same sentence is marked with different punctuation marks,and the expression of emotion is often different.Therefore,marking punctuation correctly is very important for computer to understand human's real intention and achieve better man-machine interaction.In the early days of Natural Language Processing,artificial rule was the main way to deal with natural language.With the increase of the amount of data,the research method based on statistics has gradually become the mainstream.With the development of deep learning technology and the further increase of the amount of data,the depth learning technology is gradually applied to the field of natural language,and has achieved good results.But in the point of punctuation,most of the methods are stuck in the use of traditional methods or audio,tone and other information to predict,although to some extent can complete the punctuation prediction task,but the overall accuracy is not high,the effect is not satisfactory.In order to solve these problems and difficulties,this thesis proposes a deep learning method to solve the problem,and has achieved good results.The main innovations of this thesis are as follows:(1)This thesis applies the Long Short Term Memory Network(LSTM)model to the research of punctuation prediction in this thesis,aiming at the problem that the traditional punctuation Ngram(N element)language model has insufficient utilization of historical information.(2)A bidirectional Bi-directional LSTM model is proposed in this thesis.The model can simultaneously use forward and reverse learning text information to further improve the performance of LSTM model.To summarize,the thesis puts forward the model of punctuation prediction based on deep learning technology,and improves the performance of the model gradually through a large number of iterative optimization,and finally makes the model meet the expected performance.The model completed in this thesis has been put into use in real industrial scenarios,which fully proves the validity of the model proposed.