摘要
Hashtags of microblogs can provide valuable information for many natural language processing tasks. How to recommend reliable hashtags automatically has attracted considerable attention. However, existing studies assumed that all the training corpus crawled from social networks are labelled correctly, while large sample statistics on real social media shows that there are 8.9% of microblogs with hashtags having wrong labels. The notable influence of noisy data to the classifier is ignored before. Meanwhile, recency also plays an important role in microblog hashtag, but the information is not used in the existing studies. Some temporal hashtags such as World Cup will ignite at a particular time, but at other times, the number of people talking about it will sharply decrease. To address the twofold shortcomings above, the authors propose an long short-term memory-based model, which uses temporal enhanced selective sentence-level attention to reduce the influence of wrong labelled microblogs to the classifier. Experimental results using a dataset of 1.7 million microblogs collected from SINA Weibo microblogs demonstrated that the proposed method could achieve significantly better performance than the state-of-the-art methods.
作者
Jun Ma
Chong Feng
Ge Shi
Xuewen Shi
Heyang Huang
Jun Ma;Chong Feng;Ge Shi;Xuewen Shi;Heyang Huang(Beijing Engineering Research Center of High Volume Language Information Processing and Cloud Computing Applications, College of Computer Science, Beijing Institute of Technology University, 5 South Zhong Guan Cun Street of Haidian District, Beijing, People's Republic of China)
基金
The work was mainly supported by the National Key Research and Development Program of China （no. 2017YFB1002101） and the National Natural Science Foundation of China （no. U1636203）.
