Dynamics and tipping point of issue attention in newspapers: quantitative and qualitative content analysis at sentence level in a longitudinal study using supervised machine learning and big data

Abstract

This study aims to provide a more sensitive understanding of the dynamics and tipping points of issue attention in news media by combining the strengths of quantitative and qualitative research. The topic of this 25-year longitudinal study is the volume and the content of newspaper articles about the emerging risk of gas drilling in The Netherlands. We applied supervised machine learning (SML) because this allowed us to study changes in the quantitative use of subtopics at the detailed sentence level in a large number of articles. The study shows that the actual risk of drilling-induced seismicity gradually increased and that the volume of newspaper attention for the issue also gradually increased for two decades. The sub-topics extracted from media articles during the low media attention period, covering factual information, can be interpreted as a part of episodic frame patterns about the drilling and its consequences. However, a sudden major shift in newspaper attention can be observed in 2013. This sudden disjointed expansion in the volume of media attention on this large-scale technology occurred after a governmental authority classified the drilling-induced earthquakes as a safety issue. After the disjointed issue expansion, safety and decision making were the main subtopics linked to the thematic frames, responsibility, conflict, human interest, and morality. We conclude that SML is a promising tool for future analysis of the growing number of publicly available digitalized textual big datasets, particularly for longitudinal studies and analysis of tipping points and reframing.

Integrating big data and marketing concepts into tourism, hospitality operations and strategy development

“…With regards to data classification, data collection methods, analytical perspectives and statistical techniques, when different methods or techniques were used in the same study, all of them were included. Therefore, the current measurement is consistent and extended with those of previous studies (e.g., Oh et al 2004;Oleinik et al 2014;Opperhuizen and Schouten 2021) that used content analysis to measure selected articles.…”

The Onset of Habituation Effects

“…A last finding to be discussed here concerns the role of key events (H4; Brosius & Eps, 1995). While our analysis was merely descriptive, it did appear that key events prompted a spike in news use (Downs, 1972), which has also been demonstrated in other studies (Kostkova et al, 2013;Opperhuizen & Schouten, 2020;Vasterman & Ruigrok, 2013). Moving forward, scholars could attempt to account for the role played by key events in more detail than was possible here.…”

“…From the above, it follows that, once the overdose stage is reached, it takes an extraordinary development to prompt "spasmodic recurrences of interest" (Downs, 1972, p. 39)-what Brosius and Eps (1995) called "key events" (p. 391). For instance, one study showed that earthquakes increased the level of attention to fracking (Opperhuizen & Schouten, 2020). Another study revealed that similar mechanisms were at play in the media attention awarded to H1N1, which is more commonly known as the swine flu (Vasterman & Ruigrok, 2013).…”

Framing is Mightier than the Sword: Detection of Episodic and Thematic Framing in News Media

“…Video framing annotation should be performed at more granular levels. Existing literature on framing analysis and annotation focuses primarily on textual sources (Burscher et al, 2014), (Cremisini et al, 2019), and also on smaller granularity such as news headlines (Guo et al, 2021) or sentences (Opperhuizen and Schouten, 2020). Thus, given the high probability that news videos would contain both mentions of episodic and thematic framing, we argue that the identification of framing type in videos should be performed at more granular levels, such as video fragments.…”

“…To discriminate between ten generic new frames in online news items about the Iraq war, Morstatter et al (2018) found that a simple linear regression classifier yielded the best results, irrespective of language, whereas a deep LSTM (Graves and Schmidhuber, 2005) performed poorly. Likewise, a linear regression model which uses NLP-generated features as input gave the best results in (Opperhuizen and Schouten, 2020), where authors use expert-annotated frames in Dutch news articles about gas drilling were used to train a classifier (average F1-score varies between 0.718 to 0.889 among 14 frames of interest). Differently, Saez-Trumper et al (2013) employed an unsupervised method to investigate whether clustering methods can help analyze different types of bias (selection, coverage, and statement bias) in social media news posts from different countries.…”