Semantic Filtering in Social Media for Trend Modeling

Applications that utilize publically available content from the web have been successful in tracking major events across a number of areas. We have developed a method of filtering to characterize trends of consumer behavior in relationship to specific products using the Twitter messaging system. Our process considers semantics at three successive levels to determine a demand signal. This begins with the establishment of ground truth keywords followed by word-level and category-level empirical keywords. Next, semantic categories of humor, emotion and negation are considered. Following, a classifier is applied for additional filtering to further support the characterization of consumer behavior. We apply this procedure to the goal of modeling vehicle purchase behavior with data acquired from Twitter. Results present strong correlation to sales data, allowing for contributions to forecasting efforts as well as Customer Relationship Management (CRM).