Emo2Vec: Learning Emotional Embeddings via Multi-Emotion Category

摘要

Sentiment analysis or opinion mining for subject information extraction from the text has become more and more dependent on natural language processing, especially for business and healthcare, since the online products and service reviews affect the consuming behaviors. Word embeddings that can map the words to low-dimensional vector representations have been widely used in natural language processing tasks. But the word embeddings based on context such as Word2Vec and GloVe fail to capture the sentiment information. Most of existing sentiment analysis methods incorporate emotional polarity (positive and negative) to improve the sentiment embeddings for the emotion classification. This article takes advantage of an emotional psychology model to learn the emotional embeddings in Chinese first. In order to combine the semantic space and an emotional space, we present two different purifying models from local (LPM) and global (GPM) perspectives based on Plutchik's wheel of emotions to add the emotional information into word vectors. The two models aim to improve the word vectors so that not only the semantically similar words but also the sentimentally similar words can be closer than before. The Plutchik's wheel of emotions model can give eight-dimensional vector for one word in emotional space that can capture more sentiment information than the binary polarity labels. The obvious advantage of the local purifying model is that it can be fit for any pretrained word embeddings. For the global purifying model, we can get the final emotional embeddings at once. These models have been extended to handle English texts. The experimental results on Chinese and English datasets show that our purifying model can improve the conventional word embeddings and some proposed sentiment embeddings for sentiment classification and multi-emotion classification.