The act of persuasion has great power in influencing people's attitudes and activities and has gained growing importance in the age of media. Recently, understanding the persuasiveness of arguments is crucial for us to better understand how information is expressed on social media and has attracted more and more interest in Psychology and Natural Language Processing (NLP) communities. 

While recent works in NLP commonly focus on studying the factors from human characters, language style, and the audience, a prior study in Psychology has shown how the topic would significantly influence the speakers' performance and audience's favorites. Therefore, we hypothesize that studying the relationships between topics and arguments or speakers would help better predict the persuasiveness. 

In this paper, we first use a novel embedding approach to encode the relations among topics. We then propose to jointly utilize the semantics of the topics as well as the previously studied speakers' and argument information to predict the persuasiveness. 

Experiments demonstrate that our system that taking the topics into the consideration can significantly improve the performance. Further analysis indicates that representing the connections between topics through embedding methods can also enhance the generalization ability over unseen topics.