Live Comments Automatic Generation Based On External Knowledge

In recent years,with the rapid development of Internet technology and the widespread use of mobile terminals,live streaming has rapidly emerged as an emerging profession and penetrated many fields such as culture and entertainment,commercial sales,and education science.As an indispensable and important element of the live streaming industry,live comments play a decisive role in the live streaming environment and live broadcast effect.Automatically generated live comments by artificial intelligence method can save labor costs,purify the network environment and other multiple effects,and has good economic and social benefits.Therefore,automatic live streaming comments generation method has gradually attracted the attention of practitioners and operators in the live streaming industry.The automatic live comments generation task refers to generating a natural language comments through the current popular artificial intelligence and neural network methods with the help of the rich multi-modal data information in the live broadcast room.At present,the automatic live comments generation on the live streaming platform is still in the initial and exploratory stage,it still has great research value and the possibility of progress.The complete automatic live comments generation technology can solve the problem of the lack of live comments in the tail of live streaming room.In this way,automatic live comments generation method can increase user stickiness anchors improve the quality of live broadcasts,enhances the audience's experience of watching live streaming,and strengthens the ability of the live streaming platform to attract flow.It can also purify the network environment by generating many anthropomorphic live comments to dilute the vulgar comments.However,the current research on automatic live comments generation still faces many difficulties and problems,which can be summarized as follows:(1)How to use multi-modal data to model the category information of the live streaming;(2)Given a period of time,the live comments comes from different groups of conversations,and they are mixed together,how to model the relationship between each live comment;(3)There is a lack of large-scale datasets to support deep model training for automatic live streaming comments generation task.The above difficulties have restricted the development of the automatic generation method of live barrage.In this paper,we propose a live streaming comments generation model based on knowledge enhancement to solve the above problems.Our model uses the title and description information from the live streaming room to characterize the live streaming topic,while using external knowledge and multi-modal information to model the center of audience talk into a multi-modal graph.Then a graph neural network is used to extract live topic expression feature through the previous graph.To support the experiment,we also independently constructed a live barrage data set,which contains 1,131 videos,234237 barrages,and 15190 data fragments after segmentation.Experiments have demonstrated the superiority of our model over several state-of-theart baseline.