thunlp/NSC

Implements three neural architectures (NSC, NSC+LA, NSC+UPA) that incorporate user and product attention mechanisms across semantic levels for document-level sentiment classification. Built on Theano with pre-trained word embeddings, the models are evaluated on IMDB and Yelp datasets, achieving state-of-the-art accuracy and RMSE metrics through hierarchical attention over document structure.

287 stars. No commits in the last 6 months.