The Use of Inverse Regression, Topic Models, and Sentiment Engines to Tone and Categorize Comments

• Activity Number: 240
• Type: Contributed
• Date/Time: Monday, July 30, 2012 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistics in Marketing
• Author(s): Kurt Pflughoeft*+ and Felix Flory and Carrie Lu
• Companies: Maritz and evolve24 and evolve24
• Keywords: Sentiment engines ; topic models ; inverse regression

Abstract

Statistical topic models are currently the state of the art in extracting topics from a corpus of documents. The popular unsupervised LDA (Latent Dirichlet Allocation) proposed by Blei et al. has successfully been improved to learn the unknown number of topics, the variation of topics over time and the correlation between the extracted topics. Supervised models such as Inverse Regression for text and Supervised LDA have also been developed to study the relationship between topics and dependent variables (such as sentiment). In this paper, we use both unsupervised and supervised models to extract topics and sentiment from textual comments and ratings collected from hotel review websites. The results are compared to values generated by sentiment engines and scores assigned by humans. The comparison is summarized by several statistics for inter-rater agreement (e.g. Cohen's kappa, Krippendorff's alpha). Furthermore, we compare the results of unsupervised topic model to those of rule based NLP approaches.