Abstract

An increasingly common problem in data analysis is to infer the relationship between text and characteristics of the speaker or document source. Various modifications of the multinomial bag-of-words model are most prominent among approaches designed specifically for text regression, although many generic high-dimensional pattern recognition techniques are also applicable. We investigate one such generic technique, partial least-squares (PLS), which is commonly used in engineering and physical sciences. This inquiry is motivated by the discovery that ``slant-measure’’, a heuristic from political economics for regressing ideology onto text, is just the first PLS direction. Our goal is to provide a Bayesian analysis scheme for text regression which takes advantage of the mechanics (and initial economic motivation) of PLS , and to this end we devise model-based interpretations of the algorithm and adapt these to account for the specifics of text-count covariate matrices. Results are provided in the motivating application of ideology analysis for the 109th US Congress.

http://faculty.chicagobooth.edu/matt.taddy/