Multilevel Cascaded Analysis for Sentiment Extraction from Movie Reviews:

A great amount of research is devoted to identifying and extracting the sentiment from texts. However, the classification accuracy remains quite low in comparison to other applications of natural language processing, such as topic extraction. We believe that a method capable of extracting and combining information from different text-levels, such as words and sentences, would be able to capture the sentiment of documents more accurately than existing techniques. In this thesis we propose three improved methods for cascaded sentiment analysis (semantic center, codebook and mixed-error autoencoders). Also, we introduce a new feature extraction approach for sentiment analysis, the multilevel cascaded sentiment analysis. We experimentally evaluate the proposed methods using two movie reviews datasets. Multilevel cascaded sentiment analysis was able to exceed the state-of-the art accuracy in one dataset by 1.1% and achieve the second best result in the other. We also provide statistically significant evidence that our method can lead to accuracy gains when combined with other document-level classifiers.