Research On The Instruction Understanding In Voice Manipulation For Unmanned Vehicles

With the development of social economy and scientific technology,unmanned vehicles plays an important role in the future intelligent transportation.The users can be liberated from the hard driving task,and efficient,friendly and accurate man-unmanned vehicle manipulating method is the key link to satisfy the operational requirements of users.Among many interaction methods,voice manipulation is the most direct and efficient way.Voice manipulation is mainly divided into two modules.Firstly,speech recognition gets text instruction corresponding to voice instruction.Then intention understanding module finds intention category corresponding to text instruction.How to make unmanned vehicles correctly judge the category of intention corresponding to the text instruction is an import part of voice manipulation.This paper studies the method of intention understanding of text instruction.First of all,a single intention corpus and a complex intention corpus are constructed by analyzing all the possible user's intention requirements of unmanned vehicle voice manipulation;secondly,based on character-level features in English text classification,considering the differences between Chinese Pinyin and English words,a feature representation method using consonant and vowel in Pinyin for Chinese text classification is proposed;thirdly,traditional Recurrent Neural Network(RNN)units are replaced by Gated Recurrent Units(GRU)for the problem of difficulties in capturing long-term dependencies.To extract high-level features,shorten the length of feature sequences and increase convergence rate of the model,a deep learning model combining Convolutional Neural Network(CNN)with GRU-RNN is established.The output layer of the model uses softmax to classify the instructions,then the intention of instruction is understood.To evaluate the performance of the model on short and long sequence task,10-fold cross validations are implemented on corpuses for two tasks respectively,then comparisons and analysis are carried out against other classical natural language understanding methods based on vector space model,word embedding and character-level feature based on single character table.The result shows that the proposed model can significantly improve the accuracy of classification for the intention texts.