Does ensemble really work when facing the twitter semantic classification?

With the rapid development of Internet social media, twitter has gradually become the most mainstream information release and information sharing platform. A large number of twitter users use the platform to express their views, emotions and opinions. However, it is still a challenge on twitter semantic classification based on the observation that Twitters are short, noisy, arbitrary, etc. Thus, we seek in the mainstream NLP algorithms to find out which algorithm performs best in this problem. After that, we analysis the ensemble methods on the former encode expand to get a better result. However, we find that it dosen’t work well as we expected. we analysis the reason and give the potential explain. The extensive experiments have shown that the LCF-BERT based model performs best over the mainstream algorithms and the ensemble model on the Twitter dataset.\