UC Santa Cruz

Think tanks are constructed on the premise of research-based advocacy of a subject. Yet, it has been observed that the published opinions of think tanks are often devoid of unbiased objectivity. In my Masters’ thesis, I intend to establish that there exists an underlying bias in the articles published by a think tank, which is inferred from the choice of vocabulary used. For this, I used articles published between 2005 and 2019 by six prominent US-based think tanks to create a pool of about 100,000 articles. The think tanks are chosen randomly while ensuring that this selection represents the diversity of American political orientation. Since, the intrinsic nature of textual data poses the challenge of high-dimensional feature space, I chose to pre-emptively cluster the articles into prominent issue areas using word embeddings. This procedure of mapping articles to the maximum probable issue area is in part analogous to a probabilistic multi-label classification. Next, for each of the resulting subsets, I performed supervised multi-class classification and compared the output of various classifiers employed. The estimators deduced from these models help to construct a set of phrases that is characteristic of each think tank in each issue area. From this, I observed that for a classification rate in the range of 75 – 83%, these algorithms render similar feature sets for each think tank in a given issue area. These feature sets are further used to analyze the degree of overlap in the vocabulary used in the articles published by different think tanks. I observe that the data reflects a minimum overlap of approximately 20% between ideologically different think tanks, which increases to a maximum of 55% for certain ideologically aligned think tanks, thereby indicating that there exists an inherent bias in the analyses performed by think tanks, which can be primarily attributed to their founding ideology.