Master's Thesis: Weakly Supervised Learning for Compositional Sentiment Recognition
Michael Haas: "Weakly Supervised Learning for Compositional Sentiment Recognition in German", 2014. Supervised by Dr. Yannick Versley
My master's thesis is concerned with compositional sentiment recognition. I propose a weakly supervised technique to create sentiment lexicons from product reviews. In a cross-lingual setting, I use automatic alignments between syntax trees to leverage extensive English resources for sentiment analysis in German on the phrase level. Furthermore, I show how semi-supervised methods and ensemble learning can improve sentiment analysis performance in a resource-scarce setting. A crowdsourced German sentiment treebank is used to for training and evaluation. On this compositional dataset, I further investigate the relevance of compositionality for sentiment analysis.
The full text of my thesis is available (PDF). The code is available on Code on GitHub. The README file describes how to reproduce my results. The required data currently is available on request.