Sentiment Orientation Analysis Of Sentences With Modality

This paper studies sentiment orientation analysis of sentences with modality. The goal is to classify sentences as positive, negative or neutral based on the sentiment they expressed. As a primary research subject of opinion mining, sentiment orientation analysis has attracted numerous researchers’attention. They tried to improve the accuracy through redesigning features and machine learning methods. However, as these technologies are getting mature, it is hard to improve the accuracy greatly. One of the drawbacks of most existing methods is that they did not take account of some special sentence structures’characteristics. Recently, some researchers argued that it is scarcely possible to find a one-technique-fit-all method to deal with different types of sentences efficiently. A divide-and-conquer method is required. Handling different types of sentences specially based their characteristics is doable. Inspired by their work, we deal with sentiment orientation analysis of one special type of sentence-sentences with modality.Sentences with modality comprise a significant proportion in customer reviews. In our data set, there are almost18%of sentences with modality. For a sentence with modality contains at least one modal verbs, this type of sentence can be easily detected with the help of POS tagging. Our method can be naturally integrated into existing methods and helps to improve their accuracy.Sentiment orientations of sentences with modality are hard to be processed by general methods because of their unique particularities. Firstly, sentiment orientations of sentences with modality may be opposite to those conveyed by the sentiment words used. This is because modal verbs affect sentiment orientations of those sentences, for example,"should be nice" and "must be nice" always express different sentiment orientations. To solve this problem, we combine modal verbs and sentiment words that immediately follow them as modal features. We also release the problem of sparse modal features.Secondly, the categories of modality contribute to the sentiment orientations as well. For example, although sentences "I would recommend this phone to all my friends" and "If this vacuum had a more powerful turbo, I would recommend it to everyone" contain same modal features, they express different sentiments. We treat categories of modality as significant features. To capture these features, we predefine some patterns. With the help of POS tagging, we detect some types of modality such as subjective, deontic modality and dynamic modality.At last, we build feature vectors of sentences with modality using both general features and features related with modality, such as sentiment words, negation words, modal features and categories of modality. We use these feature vectors to train a SVM classifier, which is employed to process sentiment orientation analysis of sentences with modality. Experimental results show that the features we proposed helps a lot. The problem of sparse modal features is relieved by a certain degree by merging synonymous modal features.