SentiVec is a method for distributional analysis, which is used to produce vector representations for words (also known as word embeddings). The method extends and outperforms popular Word2Vec by optimizing an auxiliary lexical objective. The lexical objective takes advantage of word classes, such as positive and negative opinion words, and enforces word embeddings of the same class to be closer to each other.

Try out SentiVec Demo.

Download Pre-trained Word Embeddings

Binary Classification Task Amazon
Amazon Sentiment 86.4% 82.6%
Rotten Tomatoes Sentiment 77.9% 75.8%
Objective-Subjective 90.8% 90.6%
News Topics 83.1% 83.5%

Source Code


Maksim Tkachenko, Chong Cher Chia, Hady W. LauwSearching for the X-Factor: Exploring Corpus Subjectivity for Word Embeddings, ACL, 2018.