Speaker: SEAS Postdoctoral Fellow Oren Tsur

Title: Don’t Let Me Be #Misunderstood: Linguistically Motivated Algorithm for Predicting the Popularity of Textual Memes

Abstract:

Prediction of the popularity of online textual snippets gained much attention in recent years. In this talk I investigate some of the factors that contribute to popularity of specific phrases such as Twitter hashtags. I define two prediction tasks and propose a linguistically motivated algorithms for accurate prediction of hashtag popularity. These prediction algorithms successfully models the interplay between various constraints such as the length restriction, typing effort and ease of comprehension. Controlling for network structure and social aspects we get a glimpse into the processes that shape the way we produce language and coin new words. In order to learn the interactions between the constraints we cast the problem as a ranking task. We adapt Gradient Boosted Trees for learning ranking functions in order to predict the hashtags/neologisms to be accepted. Our results outperform several baseline algorithms including SVM-rank, while maintaining higher interpretability, thus our model’s prediction power can be used for better crafting of future hashtags. Finally, I'll discuss possible causes for some errors in the prediction and show how social forces such as canonization and institutionalization inject ``noise'' to the system.  

Biography:

Oren Tsur is a postdoctoral fellow at Harvard University (SEAS & IQSS) jointly with Lazer's lab at Northeastern University. He earned his PhD. in Computer Science from the Hebrew University and his research combines Natural Language Processing and Network Science. Oren received the 2014 NSF fellowship for research of Political Networks. Him and his colleagues were recognized by by Time Magazine as one of the 50 Best Inventions of 2010 for their work on sarcasm detection. Here is pop sci talk [HEB].

For more information, Oren's homepage is available here.