Computational Public Opinion Measurement: A Systematic Review of Methods and Methodological Limitations

Computational Public Opinion Measurement (CPOM) uses natural language processing and machine learning to infer public attitudes from social media. However, the methodological foundations and validity constraints of CPOM remain inadequately documented. This systematic literature review examines the computational methods, sentiment representations, and measurement strategies employed in CPOM research, and identifies the structural and methodological limitations that constrain CPOM as an approach to measurement. This systematic review (PRISMA, 2020) searched seven databases, identifying 56 studies from 5,108 records (2008–2025). Methods shifted substantially across eras: lexicon-based approaches declined from 77.8% (2011–2015) to 28.6% (2021–2025) while deep learning grew from 0% to 40.0%. Despite this technical evolution, empirical validation remains rare. Only 11 of 56 studies (19.6%) computed a quantitative statistic against an external benchmark such as a survey or poll. In other words, methodological sophistication and validation rigour moved in opposite directions: none of the 14 deep learning studies computed a quantitative validation statistic, while all 11 that did used lexicon-based or classical machine learning. The broader field shows an even larger gap: 49.4% of all full-text-assessed papers were excluded for providing no external validation or representativeness analysis at all. The recognition-action gap widened over time: TSE awareness grew from 44.4% to 80.0% across eras while quantitative validation rates fell from 33% to 11%, showing that growing awareness has not translated into action. CPOM must move beyond technical sophistication toward systematic criterion validation, demographic adjustment, and transparent reporting.