Due to the abstractness of the framing concept, one of the most prominent challenges in media frame researches is the development of methodological procedures that can provide more validity and reliability to analyses. One way of dealing with this challenge is to use data-reduction techniques, i.e. divide frames into representative categories, for a more systematical research operationalization. Studies break down frames into elements by two means: from terms or expressions — syntactical/automatic perspective — or from a priori defined elements — semantic/manual analysis. The automatic perspective is criticized because it tends to identify topics rather than frames. Nonetheless, it is more reliable, since the codification process itself is done by a machine. On the other hand, the manual approach is more valid in analytical terms because it tends to actually identify frames and not correlated concepts. However, its coding step requires many strategies to ensure reliability to results. Considering the contributions of each perspective, we exploit both approaches in a mixed-method that particularly copes better with large data sets. Two different combinations are proposed: (1) automatic detection of some categories in the manual analysis; and (2) adoption of manual procedures in a sample of the data set, with a subsequent automatic step that uses the generated knowledge from the manual identification to analyze the largest part of the data. We adopt data mining techniques, namely clustering and classification tools, to materialize a quantitative analysis, which has yet a semantic level as well. We apply our methodology for analyzing the obesity coverage in the Brazilian media as a case study.