Stance Detection with Bidirectional Conditional Encoding.

Stance detection is the task of classifying theattitude expressed in a text towards a targetsuch as “Climate Change is a Real Concern”to be “positive”, “negative” or “neutral”. Previouswork has assumed that either the targetis mentioned in the text or that training data forevery target is given. This paper considers themore challenging version of this task, wheretargets are not always mentioned and no trainingdata is available for the test targets. Weexperiment with conditional LSTM encoding,which builds a representation of the tweet thatis dependent on the target, and demonstratethat it outperforms the independent encodingof tweet and target. Performance improveseven further when the conditional model isaugmented with bidirectional encoding. Themethod is evaluated on the SemEval 2016Task 6 Twitter Stance Detection corpus andachieves performance second best only to asystem trained on semi-automatically labelledtweets for the test target. When such weaksupervision is added, our approach achievesstate–of-the-art results.