A Fully Automated Approach for Classifying Marketer-Generated Text on Social Media
Tracking marketer-generated text by topic on social media has many useful applications in marketing research and practice. However, traditional approaches to automated text classification (such as lexicon-based and machine learning approaches) have many limitations, in particular when applied to social media, which tends to have short posts that contain rapidly evolving linguistic features, and for sparse topics, which are difficult to curate training data for. As such, many topics of interest to marketers are infeasible or cost-ineffective to track using extant methods. We propose a fully automated approach that requires only a single keyword of input to statistically train an up-to-the-minute classifier on a topic of the researcher's choice. The method uses an “exemplar" approach to automatically create proportionately labeled sets of posts from live accounts on Twitter, followed by feature selection to create a transparent lexicon-based keyword classifier. We test the approach's performance for four topics on tweets from a diverse set of brands, and find a consistently strong performance, with F1 scores averaging 0.81. We also test performance on Facebook posts and find a similarly strong performance. These results suggest that this method can provide a reliable, flexible, transparent, and cost-effective means for classifying MGC on social media.
Meeting ID: 948 5159 1217