Media Cloud fellow Rebecca Weiss is presenting a paper at the International Communications Association 2014 Conference, in the “Advances in Measurement and Methodology” track. The paper is titled “A Case Study in Computational Content Analysis: Comparisons of Sentiment Analysis Methods on News Media“.
The advent of the Internet has been a boon for the field of communication. The digital mediums of online communication platforms hold promise for a new era of quantitative, data-driven analyses of communicative behaviors. Content analysis specifically stands to benefit greatly from advances in computer science, particularly text as data approaches. However, not all computational methods are equally suited to the content analysis domains typically studied in communication. In this paper, I survey existing text as data methods as they pertain to sentiment analysis, including the assumptions that are foundational to each method (and which are often ignored in the pursuit of automated content analysis). I review the practicalities of each of these approaches as they pertain specifically to the types of content analysis performed in social science as opposed to computer science. Finally, I present an analysis of the relationship between average daily news sentiment for 24 news outlets on presidential favorability ratings that allows for the comparison of each of these methods against each other. I find that the daily average sentiment computed by these models produce point estimates that are substantively and significantly different from each other. Additionally, I find evidence that the use of one model over another can lead to substantively different inferences regarding the relationship of news sentiment to public opinion.