We used Crimson Hexagon's ForSight platform (now Brandwatch) to retrieve relevant news articles from a list of 25 traditional and emerging media outlets.

Using the keyword combination (gun OR firearm OR nra OR "2nd amendment" OR "second amendment" OR AR15 OR "assault weapon" OR rifle OR "brady act" OR "brady bill" OR "shooting"), we collected a total of 42,917 articles from 2018.

We used the LDA topic modeling to explore prominent topics in these news articles.

We tried 5, 10, and 15 topics, and below is an example of a 5-topic outputs.

Examples of manually-assigned labels on topics

Based on the 5-topic LDA output, we can manually assign labels to these topics:
1) mass shootings
2) police officers
3) school shootings and demonstrations
4) gun rights and gun control
5) the second amendment

Based on the 10-topic LDA output, we can manually assign labels to these topics:
1) gun rights groups and gun sales
2) the second amendment
3) school shootings
4) police officers
5) politics
6) mass shootings
7) mental health
8) children and family
9) student-led demonstrations
10) gun control

Based on the LDA modeling results from Step 1 and the literature review of the media framing of gun violence, we decided the following list of frames:

 1. Gun/2nd Amendment rights
 2. Gun control/regulation
 3. Politics
 4. Mental health
 5. School or public space safety
 6. Race/ethnicity
 7. Public opinion
 8. Society/culture
 9. Economic consequences

Note that some frames are issue-specific frames that are unique to the media coverage of gun violence such as "gun/2nd Amendment rights" and "mental health," other frames are generic frames that apply to all kinds of issues including gun violence such as "economic consequences."

It is also important to note that although the LDA topic modeling results do not have explicit information related to "society/culture," we still include it because it is a media frame discussed in the previous literature about gun violence media coverage. It is possible that we would end up finding that this frame appears very rarely in our data. This would still be an important finding. After all, media framing is not just about inclusion and emphasis but also about exclusion.

We then drew a sample of over 2,000 articles, and recruited two communication students to manually annotate the headlines to determine 1) whether the headline is relevant to gun violence, and 2) the main frame of the headline based on the list of nine frames we decided from Step 2. We used the quantitative content analysis in communication research to annotate the data, which involves creating a codebook and checking intercoder reliability.

Based on the annotated sample and the state-of-the-art multilingual deep learning model XLM-Roberta (Conneau et al., 2019), we trained a new model to classify media frames in headlines. In order to assess the model performance, we implemented the 5-fold cross-validation approach. The model to predict frames in the English news headlines reached 0.83 precision and 0.83 recall, which are considered satisfactory according to the communication research standard.