152-2008: Coming to a Theater Near You! Sentiment Classification Techniques Using SAS® Text Miner

Many Web sites, including blogs, online stores, and some database Web sites, give users the ability to state their opinions about the products they buy and use. This is a serious concern for many companies because of the potential damage of reputation. Companies want to monitor and gauge customers’ opinions, but they do not have the time or resources to manually gather statistics on this ever-expanding content source. Can companies reliably predict whether a customer comment is positive or negative without actually reading each posted comment? In this paper, we demonstrate how different techniques for classifying movie reviews can be implemented in SAS Enterprise MinerTM and SAS Text Miner. Movie reviews are classified not by category, but by sentiment. INTRODUCTION A large amount of textual content is subjective and reflects opinions. With the rapid growth of the Web, more people write online reviews for all types of products and services. It is becoming a common practice for a consumer to learn why others like or dislike a product before he or she buys it, or for a manufacturer to track customer opinions on its products to improve customer satisfaction. However, as the number of reviews for a product grows, it becomes harder to understand and evaluate customer opinions about a specific product. Sentiment classification, also referred to as polarity, tone, or opinion analysis, can track changes in attitudes toward a brand or product, compare the attitudes of the public between one brand or product and another, and extract examples of types of positive or negative opinions. In this paper, we explore several techniques for improving sentiment classification with SAS Text Miner. Standard techniques in SAS Text Miner such as weightings, synonyms, and part-of-speech tagging are presented, along with more complex extraction and feature manipulation techniques. In the next section, we provide the background on sentiment classification approaches, a description of the data set that we use for our experiments, and some technical information necessary to understand our techniques. In the next sections, we present several methods for improving the effectiveness of sentiment classification: 1. Baseline SAS Text Miner – running SAS Text Miner with the default settings 2. Baseline SAS Text Miner with MI – running SAS Text Miner with the default settings and using mutual information (MI) for weighting terms 3. Sentence Filtering – building a model using relevant sentences and removing noisy sentences 4. Synonym Lists – assisting the prediction with a knowledge base of positive and negative terms 5. Sentiment Reversal of Terms– manipulating features to track the occurrence of “not” and “n’t” in text Following these sections are experimental results applied to the data set described in the background section. Then, we present our conclusions. BACKGROUND SENTIMENT CLASSIFICATION APPROACHES There are two basic approaches to sentiment classification. The first approach is to use a model that is based on the frequency of occurrence of each term in each document. Documents are represented as vectors of term frequencies, and the information about the sequence in which the terms appear is disregarded. The terms themselves become features. Techniques such as stop lists, SVD, and term weightings are used to select and enhance the features. This statistical-based approach is at the heart of SAS Text Miner and has significant advantages because the algorithms can learn the important characteristics from the collection as a whole. Surprising and unexpected terms might actually prove useful for prediction accuracy. Unfortunately, while this bag-of-words approach can be effective for classifying documents according to a topic or category, it is often less successful in classifying sentiment. Data Mining and Predictive Modeling SAS Global Forum 2008