Categorizing Sentiment Using Unstructured Text

• Activity Number: 304
• Type: Topic Contributed
• Date/Time: Tuesday, August 11, 2015 : 8:30 AM to 10:20 AM
• Sponsor: Section on Statistics in Defense and National Security
• Author(s): Wendy Martinez* and Lucilla Tan
• Companies: Bureau of Labor Statistics and Bureau of Labor Statistics
• Keywords: Contact history ; document clustering ; k-means clustering ; Survey management

Abstract

Interviewer observations about contacted sample units' initial reactions towards the survey request ("doorstep concerns") recorded in the Contact History Instrument (CHI) have been predictive of survey response in the Consumer Expenditure Interview Survey (CE). Previous studies employed an ad hoc definition of "themes" based on groupings of the 23 possible response options that describe the doorstep concerns in the CHI. Before applying this result in designing survey interventions, the meaning of these themes should be well understood. Interviewer comments recorded as unstructured text fields in the survey instrument at subsequent phases of the data collection process provide an opportunity to examine if the nature of these comments provide any corroboration of the sample units' initially observed concerns characterized by the doorstep concern themes. Our analysis consisted of creating corpora using subsets of the text fields in the CE. We pre-process the text and then encode the variables as a term-document matrix. We explore various clustering approaches, such as model-based clustering and k-means. We use the results of the clusters to verify the themes, or to suggest others.