Research On Key Technologies Of Suggestion Mining And Generation

Nowadays,the opinions of users on various social media platforms are growing exponentially with the rapid development of the Internet.Some of these opinions include the suggestions of users on various products,services and even consumers.These suggestions have high commercial value that enterprises and public departments can improve the quality of services or products.Consumers can also decide their own consumption behavior according to these suggestions.Therefore,suggestion mining is a meaningful research.At the same time,if it can automatically generate suggestion sentences,it will bring more convenience to consumers and enterprises.In this paper,suggestion mining and generation tasks have the following points:First of all,we researched the specific domain suggestion mining technology.In this task,training set,validation set and test set all came from the software development suggestion forum.It used two-category methods to classify the given sentence into two classifications:suggestion or non-suggestion.For specific domain suggestion mining task,this paper used the BERT pre-training model to obtain context sensitive text representation,and then combined Convolutional Neural Networks and Bidirectional Gated Recurrent Unit to extract the temporal and n-gram features of the text,finally the features were input into the full connection layer to get the results.Secondly,we researched the cross domain suggestion mining technology.In this task,the training set came from the software development suggestion forum,while the validation set and test set came from hotel reviews.It was also modeled as a twocategory classification task.For this task,based on the above model,this paper combined with the idea of adversarial learning,and then added a domain classifier,which maximizes the loss to eliminate its ability to distinguish the text domain,so as to realized the function of cross domain.Then,the related technologies of suggestion generation were studied.This task was derived from the uneven distribution of suggestion and non-suggestion sentences in datasets.The main function of suggestion generation was to generate suggestion sentences.For this task,this paper used BERT pre-training model to obtain the word vector,and then added a Convolutional Neural Networks classifier on the Variational Auto-Encoder model.The classifier was used to control the generated sentences to belong to the suggestion class.Then,Beam Search algorithm was introduced to improve the logic of the generated sentences.Finally,a sampling method based on penalty factor to remove repeated n-gram was proposed to solve the problem of repeated fragments caused by Beam Search.Finally,a suggestion mining and generation system was designed and implemented.The system was divided into two modules:suggestion mining module and suggestion generation module.The suggestion mining module called the proposed corss domain suggestion mining model interface to determine whether the sentence entered by the user belongs to suggestion class.The suggestion generation module called the proposed suggestion generation model interface to generate sentences of the suggestion class.Users can understand the research content and results of suggestion mining and suggestion generation more intuitively through the system.