Topic-adaptive Sentiment Analysis On Twitter Via Transfer Learning

Social network has been an irreplaceable tool for people,so there are more and more researches focusing on social network analysis including sentiment analysis on microblogs.In practical application,the target text of sentiment analysis task on microblogs usually under specific topic,the topic of target texts is change with the application requirement,and the term distribution and sentiment features of online social text like tweet is much different under diverse topics.Which situation make the sentiment analysis task gets difficult on microblogs,called topic adaption problem.This article focuses on the topic adaption problem,find that the essential reason of the problem is the lack of sentiment labeled microblogs data under possible specific topic.This article proposes two creative methods based on transfer learning for solving the problem.The main idea of proposed methods is importing sentiment labeled data from external source and reducing the difference between external source data and microblogs data by some transfer learning methods.The contribution of this article includes:An instance transfer method based on k-nearest neighbor algorithm.It thinks of a topic domain in original microblogs sample library special as a ‘topic sample',and the sentiment labeled instances is the label of ‘topic sample.The feature of ‘topic sample' and similarity between ‘topic samples' are defined by topic semantics,term distribution,and the length of instance.The k-nearest neighbor method is used for quickly locating relevant ‘topic sample' in sample library of specific topic of sentiment classification task and then outputs the label(sentiment labeled instances)of the relevant ‘topic sample'.An instance filtering method based on clustering algorithm.It puts the target microblogs instances and external source instances into same vector space,then run a clustering algorithm on both of target data and source data.After clustering,those external source instances whose cluster label are similar to target microblogs instances will be same to target microblogs instances and vice versa.The external source instances in same cluster also are ranking by their similarities to target microblogs instances.This method can reduce the difference between source and target data.A topic-adaptive sentiment classification system is implemented by implementing above two methods.Three experiments on data of eight difference topic is conducted for testing the effectiveness of proposed topic-adaptive method.The results of experiments show the proposed methods is effective,it can be used for boosting the performance of sentiment classification on microblogs under specific topic.