Abstract
We introduce a method for learning to assign suitable sentiment ratings to review articles. In our approach, reviews are transformed into collections of n-gram and semantic word class features aimed at maximizing the probability of classifying them into accurate ratings. The method involves automatically segmenting review articles into sentences and automatically estimating associations between features and sentiment ratings via machine learning techniques. At run-time, a simple weighting strategy is performed to give extra weights to features in potential evaluative sentences (e.g., the first, the last sentences and sentences with adverbs) from others. Experiments show that word class information alleviates data sparseness problem facing higher-level n-grams (e.g., bigrams and trigrams) and that our model using both training-time n-gram and semantic features and run-time weighting mechanism outperforms a strong baseline with surface n-gram features by 2.5% relatively. © 2009 by Chung-chi Huang, Meng-chiech Lee, Zhe-nan Lin, and Jason S. Chang.