Research On Garbage Comments Recognition Algorithm Based On Deep Learning

At present,the most popular way of shopping in contemporary society is online shopping.At the same time,shopping platforms generate a large number of merchandise reviews all the time.Reviews are very significant for consumers to decide whether to buy or not.However,existing spam comments usually mislead consumers’behavior,which will not only disrupt the order of online shopping,but also greatly waste information resources.Therefore,it is very necessary to design an effective method to identify spam comments.To improve the identification of spam comments,this thesis puts forward a series hybrid model.The model adds the residual structure to the Convolutional Neural Networks（CNN）,and connects the CNN with the Bidirectional Grated Recurrent Unit,then added the improved attention mechanism to learn the important information,and then used softmax to get the final results.In the comparative experiment,the accuracy rate,the recall rate and the₁F-score of the model are 93.6%,91.1%and 0.923 respectively.Compared with other recognition algorithms,the serial hybrid model proposed in this thesis has better recognition effect.However,the recognition efficiency of the hybrid model is not high,and the training time and testing time are long.Aiming at the application scenarios requiring fast recognition,this thesis proposes a parallel hybrid neural network model to improve the efficiency of spam comment recognition while ensuring a better recognition effect.The model is divided into two channels,CNN and Attention are combined as the upper channel,BiGRU and Attention are combined as the lower channel.By integrating the output results of the above two channels,softmax is used to get the final recognition conclusion.In the comparative experiment,compared with the CNN-BiGRU-A model and Res CNN-BiGRU-A model,the recognition efficiency of this model is increased by 14.6%and 24.8%respectively.And the accuracy rate,the recall rate and the₁F-score of the model are 92.7%,90.6%and 0.916 respectively.Although the effect of the model is proposed in this thesis series hybrid model is shown slightly inferior,but parallel hybrid neural network model has better advantages in the recognition efficiency,shorter training time and testing time.The serial recognition model proposed in this thesis can be applied to the application scenarios requiring high recognition effect and small number of comments.The parallel recognition model can be applied to the application scenarios requiring fast recognition and large number of comments.