Semi-supervised Probabilistic Sentiment Analysis:Merging Labeled Sentences with Unlabeled Reviews to Identify Sentiment

摘要

Document level sentiment analysis, the task of determiningwhether the sentiment expressed in a document is positiveor negative, is commonly performed by supervisedmethods. As with all supervised tasks, obtaining trainingdata for these methods can be expensive and timeconsuming.Some semi-supervised approaches have beenproposed that rely on sentiment lexicons. We propose anovel supervised and a novel semi-supervised sentimentanalysis method that are both based on a probabilisticgraphical model, without requiring any lexicon. Our semisupervisedmethod takes advantage of the numerical ratingsthat are often included in online reviews (e.g., 4 out of 5stars). While these numerical ratings are related tosentiment, they are noisy and hence, by themselves, theyare an imperfect indicator of reviews’ sentiments. Weincorporate unlabeled user reviews as training data bytreating the reviews’ numerical ratings as sentiment labelswhile modeling the ratings’ noisy nature. Our empiricalresults, utilizing a corpus of labeled sentences from hotelreviews and unlabeled hotel reviews with numerical ratings,show that treating reviews’ ratings as noisy and utilizingthem to augment a small amount of labeled sentencesoutperforms strong existing supervised and semi-supervisedclassification-based and lexicon-based approaches.