Comparing qualitative content analysis and semi automatic text analysis through Topic Modeling: the discourse on Italian steel privatizations

In this paper we analyze a corpus of texts and we compare frames inductively elicited with topics elicited through topic modeling (TM). TM provides a semi-automated way to code the content of a corpus of texts into a set of topics, which must be interpreted by the researcher. This technique has been increasingly used for text analysis, as it permits to deal with big corpora of texts, and permits to more easily reproduce results. Topic modeling is now widely used by researchers interested in how meanings are constructed through words, but studies lack that confront traditional qualitative content analysis and this technique. We compare both kinds of coding by analyzing 858 statements extracted from 390 articles dealing with the privatization of state steel in Italy. First, we use qualitative content analysis to extract frames used in the public debate. Then, we use the same data to automatically extract topics, and we inductively analyze them. Finally, we compare frames and topics qualitatively and through Multiple Correspondence Analysis and we critically reflect on both techniques. In a nutshell, we think that the results of Topic Modeling and of our qualitative analysis are similar, but not exactly identical. They do not contradict each other; rather they seem to complement each other, thus enriching our interpretation.