A machine-learning based approach to measuring constructs through text analysis

Purpose: This paper aims to develop a novel and generalizable machine-learning based method of measuring established marketing constructs through passive analysis of consumer-generated textual data. The authors term this method scale-directed text analysis. Design/methodology/approach: The method first develops a dictionary of words related to specific dimensions of a construct that is used to assess textual data from any source for a specific meaning. The method explicitly recognizes both specific words and the strength of their underlying sentiment. Findings: Results calculated using this new approach are statistically equivalent to responses to traditional marketing scale items. These results demonstrate the validity of the authors’ methodology and show its potential to complement traditional survey approaches to assessing marketing constructs. Research limitations/implications: The method we outline relies on machine learning and thus requires either large volumes of text or a large number of cases. Results are reliable only at the aggregate level. Practical implications: The method detail provides a means of less intrusive data collection such as through scraped social media postings. Alternatively, it also provides a means of analyzing data collected through more naturalistic methods such as open-response forms or even spoken language, both likely to increase response rates. Originality/value: Scale-directed text analysis goes beyond traditional methods of conducting simple sentiment analysis and word frequency or percentage counts. It combines the richness of traditional textual and sentiment analysis with the theoretical structure and analytical rigor provided by traditional marketing scales, all in an automatic process.