In this session, I'm going to talk more about sentiment analysis techniques that are implemented in white text minor, which are Stanford CoreNLP, LingPipe, and SentiWordNet. Stanford CoreNLP uses several supervised learning algorithms for sentiment analysis. Most general one is recursive neural tensor that classifiers, which is part of deep learning algorithms. And it is widely acknowledged to be a top performing sentiment classifier. It stores sentences in a parsed tree format, rather than the typical bag-of-words approach. This allows the classifier to take the sentence structure into account when classifying sentence. As shown in the example, Sentiment Treebank, or it's called cementic Treebank, knows of a binary tree of each sentence including the root node of each sentence a given sentiment score. In this semantic treebank at every level of parse tree, stands for sentiment classifier and annotates the sentiment of the phrase it subsumes. It uses 5-class scheme --, is minus two, is minus one, is zero is in the middle scale, + is one and ++ is two. This classifier is particularly useful for a more robust to unseen input. LingPipe sentiment classification is a hierarchy classifier, by composing basic sentiment analysis algorithms such as logistic regression. Basic sentiment classifiers handle two level classification tasks. First task is separating subjective from objective sentences. Second task is separating positive from negative movie reviews. LingPipe classifier is basically re-implement the basic classifiers and hierarchical classification techniques, which are described in both Pang and Lilian Lee's 2004 Acer paper. It's shown in the architecture of LingPipe Classifier, once the input text is fed into the classifier, it first eliminates objective sentences. Then use remaining sentences to classify document priority. By doing this way, it can reduce noise sentences which gain performance improvement. LingPipe implements a technique that reduced a review to 5 to 25 sentences. In other words, there are the 5 most subjective sentences as ranked by conditional probability of the subjectivity model. As well as up to 20 more sentences if 50% or more highly likely to be subjective. Let's take a closer look at the algorithm of LingPipe Classifier. At first, use unigram features, extracted from movie review data. It assumes that adjacent sentences are likely to have similar subjective-objective polarity. It uses a min-cut algorithm to efficiently extract subjective sentences. After that, it iterates over the sentences and classify them with the subjectivity classifier. Let me move onto the last classifier, which is a supervised learning based symptom analysis based upon SentiWordNet. SentiWordNet was constructed based on WordNet synsets. WordNet can be downloadable at Princeton site. SentiWordNet consists of approximately 1.7 million polar expression words. SentiWordNet's synsets consist of three labels, which are Obj, which is objective. Pos, which is positive. And Neg, which is negative. This is why SentiWordNet is based on ternary classifier. Synset scores are determined by eight ternary classifiers. As shown in figure on previous slide, scores are calculated as proportion of classifiers assigning the three respective labels. Specifically speaking, Sentiment scores are calculated by subtracting negativity from positivity scores, yielding scores ranging from -1, which is very negative, to 1, which is very positive. Classifiers differ in training data, which is expansion of seed set using WordNet relations.and learning approaches. To create the SentiWord Lexicon, Support Vector Machines, Rocchio classifiers were trained on training set, is used a four values of k to create eight classifiers with different precision and recall characteristics. Note that, as k increases, PD creation are increases. So, let's take one sentence and apply SentiWordNet scores for the sentence. Suppose that we have the sentence, very comfortable, but strap go loose quickly. In this case, SentiWordNet matches tours comfortable and loose. For the word comfortable, its positive score is 0.75, objective score is 0.25, and negative score is 0. For the word loose, its positive score is 0, objective score is 0.375 and negative score is 0.625. Therefore, overall the sentence is positive because the sum of positive score is 0.075. And sum of objective score is 0.625, and negative score is 0.625. There are several studies adopted SentiWordNet for different purpose. SentiWordNet score was used to select features. For example, Zhang and Zhang in a study, 2006, used words in corpus with subjectivity score of 0.5 or greater. Another use of SentiWordNet was combining positive, negative, objective scores to calculate document-level score. For example, Devitt and Ahmad in 2007, conflate polarity scores with a Wordnet-based graph representation of documents to create predictive metrics.
