摘要: This paper presents a simple unsupervised learning algorithm for classifying reviews as recommended (thumbs up) or not down). The classification of review is predicted by the average semantic orientation phrases in that contain adjectives adverbs. A phrase has positive when it good associations (e.g., "subtle nuances") and negative bad "very cavalier"). In this paper, calculated mutual information between given word "excellent" minus "poor". classified if its positive. achieves an accuracy 74% evaluated on 410 from Epinions, sampled four different domains (reviews automobiles, banks, movies, travel destinations). ranges 84% automobile to 66% movie reviews.