Turning Tweets into Stances

Little did we know how much our students would be competitive when we first presented them this year’s SemEval. However, as we joined forces and started passionately working on the task of detecting tweets’ stances, hardly anyone could stop them from taking the reins and bringing us some amazing results!

TakeLab is no stranger at SemEval competitions — this is our third time competing at this renowned competition. At SemEval 2012, we’ve tackled the problem of semantic textual similarity, while at SemEval’s 2015 edition we’ve tried our best to detect paraphrases in tweets. In both editions we have achieved respectable results.

Considering that this year’s SemEval included the task of detecting stances in tweets, it comes as no surprise that we got so eager to put our skills to the test once again. This time we even got some help from our students! Filip Čulinović, Paula Gombar, Ivan Paljak, and Ivan Sekulić all recognized this as a great opportunity to do something challenging, yet tremendously rewarding. And they definitely weren’t wrong!

During this endeavor we’ve organised two day-long hackathons at which our students produced most of the code (and gobbled down most of the junk food as well!). Even the night before the deadline was spent optimising classifier hyperparameters and fixing bugs.

Our model brought us the amazing 3rd place (competing against 20 teams from all over the world),  missing the top rank by the smallest of margins. This outcome, though a bit unfortunate, only encourages us to try even harder at future SemEval editions, at which we will definitely join forces with our students once again! The system description paper is available here.

Big thanks to our awesome students and Martin who devoted his time to steer the final efforts in the right direction.

SemeEval 2016 Team: Ivan Sekulić, Paula Gombar, Filip Čulinović, Ivan Paljak, Filip Boltužić, Domagoj Alagić, Jan Šnajder, Mladen Karan, Martin Tutek