In this project I re-implemented the model from the paper BERT Based Semi-Supervised Hybrid Approach for Aspect and Sentiment Classification. The paper addresses Task 5 in the SemEval-2016 workshop. This is the major dataset for aspect-based sentiment analysis (ABSA). The author's code was written in PyTorch; I re-wrote it in TensorFlow and was able to replicate the author's results. I also fixed a few bugs, and refactored parts.
I found the author's data augmentation technique quite interesting. Most aspect-based sentiment analysis tasks suffer from a lack of a large annotated dataset. Their semi-supervised labeling method is something I'd like to keep in mind for future NLP projects.
