Multi-theme sentiment analysis with sentiment shifting

Sentiment Analysis;Multi-theme;Natural language processing (NLP);Data Mining

Yu, Hongkun ; Han ; Jiawei

关键词: Sentiment Analysis;    Multi-theme;    Natural language processing (NLP);    Data Mining;   

Others  :  https://www.ideals.illinois.edu/bitstream/handle/2142/90799/YU-THESIS-2016.pdf?sequence=1&isAllowed=y

美国|英语

来源: The Illinois Digital Environment for Access to Learning and Scholarship

【 摘 要 】

Business reviews contain rich sentiment on multiple themes, disclosing more interesting information than the overall polarities of documents. When it comes to fine-grained sentiment analysis, given any segment of text, we are not only interested in overall polarity of such segment, but also the sentiment words play major effects. However, sentiment analysis at the word level poses significant challenges due to the complexity of reviews, the inconsistency of sentiment in different themes, and the sentiment shifting resulting from linguistic patterns---contextual valence shifters.To simultaneously resolve the multi-theme and sentiment shifting dilemma, a unified explainable sentiment analysis model, MTSA, is proposed in this paper, which enables both classification of sentiment polarity and discovery of quantified sentiment-shifting patterns. MTSA formulates multi-theme sentiment by learning embeddings (i.e., vector representations) for both themes and words, and derives the shifter effect learning algorithm by modeling the shifted sentiment in a logistic regression model. Extensive experiments have been conducted on Yelp business reviews and IMDB movie reviews. The improvement of sentiment polarity classification demonstrates the effectiveness of MTSA at rectifying word feature representations of reviews, and the human evaluation shows its successful discovery of multi-theme sentiment words and automatic effect quantification of contextual valence shifters.