Analysis Of Mobile Game Public Opinion Based On Extended LDA With Word2Vec And Optimized SKM

With the rapid iteration of smart phones and networks,the global mobile game industry is changing with each passing day.Especially,with the blessing of the vigorous development of the "home economy" in 2020,major mobile game manufacturers have achieved substantial growth in revenue and new users.In the lifetime of a game,game manufacturers urgently need to understand the user's experience of the game product,so that they can design game products that meet user preferences,thereby improving user long-term retention and reputation.At present,the player communities and game review areas have become the main places for users to post game comments.These places carry a large amount of public opinion information about games.How to make full use of the user's public opinion and deal with negative events is the problem that manufacturers urgently should solve.Employing ML methods to analyze text is a hot issue in the current network public opinion analysis.In order to obtain the hot topics and the users' emotional tendency towards game products without manual intervention,this paper establishes two main tasks: topic detection and sentiment analysis.In the task of topic detection,the topic distribution obtained by extending the LDA model with the word vector training of Word2Vec's Skip-gram is proposed as a text representation,which is used as a feature of the SKM algorithm for clustering.The first-order difference of the distance cost function value is used to optimize the selection of the number of clusters in SKM.Multiple text clusters are obtained as topics of game public opinion,and the content of each topic is understood through keyword extraction and word cloud visualization.When comparing the differences between the experimental results of each group and manual annotation,this method achieves 80% in F1 score.In the tasks of sentiment analysis,this paper proposes to use the text sentiment scores obtained by Boson NLP to correct the original sentiment labels users given.Finally,a three-layer front is constructed.It realized the three classification tasks of text sentiment,and achieved 76% recognition accuracy on the test set.Compared with the method based on a single How Net and Boson NLP sentiment dictionary,the recognition accuracy was improved by 24.2% and 2%,respectively.In order to verify the effectiveness of the topic detection and sentiment analysis methods,this paper display the whole process in the empirical research.This paper uses a directional crawler to obtain the "Heartstone" game product on major platforms.After preprocessing the crawled original text data,the EDA was carried out to initially understand the data and determine the value of its further analysis;then the overall topic of the text was detected,and the hot topic of "game mechanism and environment" was discovered,at the same time,conduct sub-platforms to check whether there are differences in their respective topics,and give targeted game improvement suggestions to different platforms;finally,analyze the overall sentiment tendency of the text to identify negative comments that belong to the topic of "mechanism and environment",Sub-topics such as "dropped in the round and lost the game","bad card balance","deteriorating ladder environment" and other sub-topics of the topic are detected again,and game optimization suggestions are given for the content of the sub-topics.