Learning sentiment-inherent word embedding for word-level and sentence-level sentiment analysis

Vector-based word representations have made great progress on many Natural Language Processing tasks. However, due to the lack of sentiment information, the traditional word vectors are insufficient to settle sentiment analysis tasks. In order to capture the sentiment information, we extended Continuous Skip-gram model (Skip-gram) and presented two sentiment word embedding models by integrating sentiment information into semantic word representations. Experimental results showed that the sentiment word embeddings learned by two models indeed capture sentiment and semantic information as well. Moreover, the proposed sentiment word embedding models outperform traditional word vectors on both Chinese and English corpora.