Research On Speech Recognition Method Of Railway Traffic Term Based On Natural Language Processing

In order to prevent the safety accidents caused by the unskilled operation of the traffic attendant from happening again,China Railway Corporation issued three consecutive documents,which clearly mentioned that the station needs to be equipped with corresponding simulation drilling equipment,so as to ensure that the quality of the on duty train attendant is up to the standard and has the certificate to work,so a set of training platform consistent with the site is urgently needed to train and assess the train attendant.However,the existing traffic simulation training system has not completed the operation training under abnormal conditions.Under abnormal conditions,the traffic attendant needs to conduct voice interaction with other types of work to eliminate the fault.The introduction of speech recognition can realize speech interaction,but the problems also follow.The existing speech recognition software,such as iFLYTEK,ThinkIT,Baidu and so on,does not aim at the corpus of railway industry,resulting in the recognition accuracy of railway traffic term is only 50%,which is far from meeting the needs of speech interaction.Therefore,dissertation proposes a traffic term speech recognition method based on natural language processing,which aims to break through the limitations of traditional traffic attendant training,virtualize related types of work,realize the voice interaction between the middle class and each post,and recognize it as text information for system assessment.Based on the natural language processing,this method checks and corrects the speech text recognized by iFLYTEK,and then analyzes the semantics,so as to improve the recognition accuracy of railway traffic term and complete the human-computer interaction between the training personnel and the system.The research content includes the following parts:First of all,combined with the related theories of the goal of dissertation and natural language processing,the most suitable method is selected and applied to the following text.Secondly,the method of traffic speech text correction and error detection is designed,and the joint error detection model and text error correction algorithm are established.When the conditional random field is used in the error checking model,mutual information is used to make up for the lack of the internal relationship between the damaged words and words.Different error correction methods are used for different types of errors.The redundancy errors are directly deleted,the missing errors are corrected by language model,and the substitution errors are corrected by homonym dictionary.Finally,the semantic analysis of the corrected text is carried out to complete the system identification.Dissertation proposes an improved forward maximum matching segmentation algorithm.Through word bank preprocessing,all words with the same initial word and the same number of words are stored in the same chain list,which can quickly locate keywords and shorten the running time of the algorithm.This topic can also be used for texting recording documents such as message,approval document and dispatching order of on-site traffic service and separating the record of maintenance account from manual work to form special format document,which is convenient for fault analysis and expert guidance.