Design And Implementation Of Classification System For Short Text User Comments

With the increasing popularity of the Internet,there were a growing number of comments on goods and services.Before purchasing a product,users usually reviewed the relevant comments and analyzed the advantages and disadvantages to get an objective evaluation about the product.Those comments provided useful help in determining whether the product suited users or not.Owing to the effectiveness,the amount of those reviews has taken an exponential increase making the analysis work more and more impractical.The short text classification system provided a convenient way for users and businesses to analyze products.Given a comparing classification results on products,users could have a general understanding of the practical use of products.Based on the above motivation,this thesis mainly completed the following work:1.The analysis of system's requirements determined the system's goals and function points.2.The system's preliminary design determined the system's five major modules.The detailed design of the system refined each major module.The main improvements of the core classification model are three points.(1)Proposing an improved Bi-RNN model.The existing Bi-RNN model had weakness in extracting historical information,resulting in bad performance on five-category problems.To solve this problem,this thesis added an attention layer before the GRU node,which enabled previous historical information to be considered and added dimensions to the output of each GRU.(2)Proposing a method of integrating the deep neural network model with the Attention mechanism to weight the importance degree of each word for the classification result.(3)Proposing a method for joint training both five-category and the two-category problems,with scoring predictions of the comments took into consideration.The experimental results showed that the classification model which combined the above three methods achieved the optimal performance.3.System used the crawler during the implementation processed and obtained 1.11 million user reviews for the purchase of mobile phones on the JD platform.After filtering a lot of spam comments,the entire system is implemented.Finally,functional tests and performance tests were performed on the system and its various modules.All modules met the design requirements in functional tests.In the performance test,the time for completing different functional points was compared.Experiments indicated that the system basically met the needs of its application scenario.