CONCLUSIONS AND DISCUSSIONS
5.2.2. The QAP Correlation and Partisan Slant Analysis
國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
influencers, and the increasing use of social media contents as news sources in journalism. As the result suggested, the similar distribution of Facebook and news media supported the first reason, the dynamic cross-media information flow, as scholars found the reciprocal information flow between social media and news media (Wang & Guo, 2018; Neuman et al, 2014), which went beyond the traditional ways framing and agenda setting (Wirth et al., 2010), combining the concept of “frame” and “agenda setting” as the term “frame setting” to define the newly intermedia agendas relationship between news media and social media (Wang & Guo, 2018).
However, the similar distruibution also suggested that the frequency volume focused on some specific events, which might reveal the similar salient issues between social media contents and news media outlets. As the results of semantic network analysis answering RQ1, the DPP, KMT political actors and news media suggested the similar top-ten salient concepts.
The similar frequency trend indicating the similar focused events or issues on time series might be one of the reason leading the similar top-ten salient concepts between social media actors and news media in RQ1.
5.2.2. The QAP Correlation and Partisan Slant Analysis
As for the partisan slant analysis in this research, this study firstly used social media as the parameter, using the semantic network correlation analysis (Quadratic Assignment Procedure, QAP) to identify the partisan leaning of online news media. Through conducting the correlation analysis, it could measure the partisan slant through the scores of QAP results correlated to the parameters. This research design could be discussed in two dimensions: the intermedia contents as the parameter, and the applications of QAP analysis to measure partisan
‧
國立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
slant. For the first dimension, this research utilized the intermedia contents as the parameter.
Past studies (e.g., Gentzkow and Shapiro, 2010) used Republican or Democrat congress documents and proposed an index of media to measure political news media’s partisan slant.
This research also decided to apply different media contents (political actor’s Facebook posts) as the parameters to measure news media’s partisan slant. Plus, this research added a non-partisan parameter by fan pages of environmental groups.
Secondly, this research used the QAP correlation score to define the news media’s partisan slant as pro-DPP or pro-KMT. QAP analysis (Krackhardt, 1988) could examine whether two networks were strongly connected, by computing the Pearson’s correlation of two networks and recomputeing the matrices with equal size of nodes (Jiang et al., 2016; Guo, 2015). QAP was a sort of regression analysis to define the correlation between network relationships (Krackhardt, 1988). It was adopted by scholars to compare the agenda issue networks (Vargo, Guo, McCombs & Shaw, 2014; Guo & Vargo, 2015; Guo, 2012) to analyze networks which were made up by “themes” as nodes, or compare two semantic networks which were made up by “words” as nodes (Jiang et al., 2016).
As mentioned aboved, during the analysis process of QAP analysis, this study should keep the same size of networks to conduct semantic network analysis. The past studies using the angenda issue networks could keep the same size of all the comparing matrices, since they considered a fixed number of nodes in every matrix, which using “themes” as nodes after content coding analysis (Guo & Vargo, 2015; Vargo, Guo, McCombs & Shaw, 2014; Guo, 2012). These semantic networks made up of “themes” as nodes were called “the network agenda setting (NAS) model” (Guo & Vargo, 2015; Vargo, Guo, McCombs & Shaw, 2014;
‧
國立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
Guo, 2012). For example, Guo and Vargo (2015) compared associative issues of two United States presidential election nominees (Obama and Romney) through QAP analysis and found that the networks were consisted of 16 issues as nodes such as “economy,” “education,”
“welfare,” and so forth. Different from creating NAS models, this research took reference on Jiang and her colleages’ research (2016), in which networks were made up by “word” as nodes.
This research processed the data cleaning through word segmentation and eliminating the stopwords (such as function words). Therefore, the nodes of matrices and networks were composed by words or terms. As the study processed the word segmentation, every document (post or article) pocessessed unequal number of words. Also, this research should select a number of words in every network unit to keep same node number in each network. This research selected top 100 words with highest document frequency as nodes to construct each network. Through conducting QAP analysis, the results could reveal the correlation between networks. However, it was difficult in delving into the analysis based on comparing networks as they were made up of words. Several reasons are as follows.
First, it was not an easy way to observe the relationship between 100 words. Second, they were highly co-occurred in documents. Third, it would limit the lexicon or word base representing either political party of speech or positions, from two aspects. For one thing, selecting a small number of words as the node size, leading the higher possibility of insufficient to represent the speech of either political party. For another, there is a limit to judge the position or the standpoint of statements on the issue based only on words and the words’ relationship made up of co-occurrence frequency. Besides, in order to highlight the salient concepts of the issue representing any related topic, this research eliminated the stopwords, and the words that
‧
國立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
might influence the position of the issue (e.g. disagree, approve, oppose). That is, it was not that easy to delve into the position of the issue. Plus, as the partisan slant analysis was based on semantic correlation analysis, in which the networks were also based on word co-occurrence frequency, it is also a limit to obtain deeper investigation of the partisan leaning analysis through media contents.
Forth, in order to create the semantic networks, this research conducted word segmentation, and used the words or terms as the nodes to build up networks. It was the process to decontextualize the story of news articles or Facebook posts. Therefore, it was not an easy way to read the thread of the story, or the position of the issue based on segmented words and their co-occurrence relationship in networks. The limited unstanding of the contents might cause the misinterpretation of the story or be restricted to analyze the position of the statements, when this research would like to recontextualize and contruct the story on the basis of the semantic network.
As for using the NAS model, which might be made of several themes as nodes, it would be easier to compare the relationship of nodes and issues between networks. Take the Guo and Vargo’s (2015) research for example again. They compared associative issues of two United States presidential election nominees (Obama and Romney) not only through QAP analysis with correlation scores, but also through the relationship of the nodes referring 16 issues, including “economy,” “education,” “welfare,” and so forth, between the Obama network and Romney network. As this study using QAP analysis to measure the partisan slant of news media, the correlation score still indicated the results of the partisan leaning of each news media brand or news media type. However, based on the limitations mentioned in the previous
‧
國立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
paragraphs, that comparing networks made up by words was difficult in delving into the analysis, for the observation of relationship between 100 words, the high co-occurrence in documents, and the limitation of the lexicon or word base, the future study was suggested to include thematic analysis, to delve into relationship observations in QAP correlation analysis.
Regarding the aspect of analysis process, this research expected the difficulty in getting non-significant semantic network correlation (QAP) scores in the second section of analysis (RQ2) that would directly cast influences on the third section of analysis (RQ3). In RQ2, this study explored the partisan slant of news media. As the results suggested, both media types and each news media brand revealed significant values, which meant that each news media brand was classified into either pro-KMT or pro-DPP group for RQ3.