This is a great question and I agree with you entirely. Ultimately, Framing (thematic vs. episodic) is the most important consideration for these articles. It'

This is a great question and I agree with you entirely. Ultimately, Framing (thematic vs. episodic) is the most important consideration for these articles. It's important for articles to discuss road design, safety initiatives, etc. in order to equip readers with the knowledge necessary to get governments to act. However, Framing is also the most difficult pattern to detect simply because there are so many different thematic elements you can discuss, and many different ways to say each one. Here are a few examples: - This is the 3rd pedestrian fatality this year. - Residents have long complained about people racing down their street. - Washington St. has had a long lasting problem with weather-related crashes. - According to the NHTSA, there are X crashes per year related to prescription drug-induced drowsy driving. - Street width directly correlates with driving speed, which in turn increases the time needed to react and brake. I'm relying on manually annotated articles (700+ so far) and a custom-trained spaCy model to detect these elements, and the accuracy of that model depends on having enough training data. Since thematic elements are so rare in real-life articles, I'm having a difficult time providing enough training data. Hence why the tool emphasizes the other editorial anti-patterns, for now, but this will change once the Framing reporter gets better. Anyone have any suggestions for augmenting my training set for thematic elements?