Research On Evaluation Collocation Extraction Method Based On Capsule Network

In the age of big data,a huge amount of comment data is produced on the Internet every day,those comment data contains valuable information,that can be useful to marketing decisions and public opinion monitoring and so on.With the improvement of computing power of computer and the rise of machine learning techniques,people can extract useful information from these massive comment data.Opinion mining is widely concerned,and become a hot spot in the field of natural language processing.Evaluation collocation extraction is a basic task of opinion mining,and it is to extract the opinion targets and opinion-bearing words in the comment text data,Opinion-bearing words can contain information such as the emotional tendency and attitude of the comment text.Opinion targets is an aspect of the product.The extraction of evaluation collocation is conducive to the realization of fine-grained emotional analysis and other upper-level applications.At present,the methods of utilizing these grammatical structures or other manual features are still being used.Therefore,it is meaningful to use the neural network model to automatically extract text structure information and semantic information and further improve the accuracy of the method of evaluation collocation extraction.Capsule network is a new type of neural network,it use a capsule vector to represent a class of features,instead of a scalar used by the traditional neural network,Capsule vector can improve the network's feature fitting ability,and the dynamic routing algorithm in the capsule network,can make full use of the relationship between features,the high-level capsule can gather the same class of features from the lower-level capsule,In this paper,the dynamic routing algorithm of capsule network is used to optimize the evaluation collocation extraction method.In this paper,the evaluation collocation extraction process is divided into two stages: the joint extraction of opinion targets and opinion-bearing words,and the recognition of evaluation collocation.A network model based on Caps Bi GRU is proposed for the stage of opinion targets and opinion-bearing words extraction,the model extract different context information features by using multiple Bi GRU in parallel,capsule vectors is constructed based on these extracted features.Then dynamic routing algorithm is used to aggregated vector features.Finally,the aggregated features were used in Conditional Random Field to complete the sequence labeling.Experiment on three data sets are carrying out and the results show that Caps Bi GRU achieves the best F1 value in contrast to several existing models.In the evaluation collocation recognition stage,two composite models are proposed.One is a network model combining Bi GRU and dynamic routing algorithm,which uses Bi GRU to extract context features,and dynamic routing algorithm is used to identify the correlation between the semantic features and location features of opinion targets and opinion-bearing words.The other is a joint model based on Bi GRU and Capsule Network,convolution is used to obtain more advanced features on the basis of features extracted by Bi GRU,and using dynamic routing algorithm to aggregation features.Three data sets for evaluating collocation recognition are established by modifying the original data sets,and two model achieves the best F1 value on the built data set.