Honors Presentation: Miles Bratton
Unsupervised Emotional Classification Using Vector Space Models
Thursday, May 1, 2014 at 11:30 a.m.
Classification of emotions from audio and text has broad applications such as sentiment analysis for security, business, and health purposes. For example, websites that work to immerse users, especially social networking websites, may utilize the user’s emotions to personalize its product offerings. This honors thesis is concerned with the classification of emotion from text. The classification is done using a categorical linguistic lexical database and an extension to that database which classifies words based on emotions, WordNet and WordNet-Affect. After word classification, non-negative matrix factorization and other vector space models are applied to identify the different emotional clusters, allowing the unsupervised categorization of large amounts of documents. The datasets being evaluated are the SemEval2007 and ISEAR databases which contain emotionally charged documents.