Public Safety News Analysis

Project link: https://github.com/eric-mc2/quantify-justice-news-geos/

Introduction

According to Pew Research, crime is the second-most watched news topic besides weather. Over three quarters of U.S. adults say get news about crime sometimes or often.¹ How accurate is this information? How does it affect public perceptions on crime? And how, ultimately, does that affect public safety conversations, economic behaviors, and social good.

Decades of social science research have shown the importance of news coverage in shaping our world views. The media helps determine what issues are important and deserve attention, a process called “scheduling”. Within each story, journalists also decide what aspects to focus on, showing readers how to make meaning of this information, and “framing” the story in a particular set of values.²

Perceptions of crime

Decades of opinion research have also shown a consistent disconnect between fear of crime victimization and actual crime rates. For the past 29 out of 31 years, a majority of U.S. adults have said that crime is worse this year than last year.³ This is despite official statistics showing consistent decreases in violent crime rates over time.

Overall, fear of victimization has remained a salient issue, with 46% of adults describing crime in their area as being an extremely to somewhat serious problem. These perceptions cut across types of crime: 69% of households report that they frequenty or occasionally worry about identity theft, 30% worry about muggings, 21% worry about sexual assault.

Experience with crime

Violent crimes are overall much rarer than property crimes. Comparatively, certain white-collar and cyber-crimes are surprisingly ubiquitous. Three percent of households report being victims of physical assault or mugging, and 2% of sexual assault. On the other hand, 13% percent report being victims of vandalization, whereas 24% of households have report being victims of credit card hacking, 15% of financial scams, 14% of identity theft.

Perceptions of crime have immediate effects on behavior, with longer term systemic implications carrying over to other policy realms. For instance, almost half of U.S. adults claim to avoid certain places or neighborhoods due to concern over crime and 32% report having bought a gun for protection.³

Accuracy of media coverage

Survey after survey consistently show a positive correlation between percieved victimization risk and news consumption.⁴ Whether one causes the other, or vice versa, is an empirical question. 50% of adults say the media neither over nor under reports crime. Given this link, it is important to investigate, does the coverage of crime reflect actual crime rates? Are locations and types of crimes accurately represented?

The Chicago Justice Project has previously highlighted disparities in stranger vs intimate partner violence rates compared to the types of cases that are covered.

Literature Review

Previous studies have relied mostly on opinion surveys. Succar et. al. analyze Twitter streams by counting the number of tweets mentioning crime-specific keywords, and performing a sentiment analysis on police/crime-related tweets.⁵

The Quantify Justice News project extends their methods in two ways. First, we analyze the content of news articles, specifically the locations of reported crimes. Second, we train a machine learning model to classify articles as crime-related, rather than relying on exact keyword matches.

Methods

Using a database of daily local news articles covering the Chicagoland area, I trained a machine learning pipeline to analyze:

What locations (community areas) covered in news of crime?

Why community areas? For privacy concerns, crime events are not published with exact addresses. Community areas are stable administrative boundaries designed for long-term policy analysis and planning.

ML Tasks

Following this excellent talk from the folks at Spacy I’ve broken down this research question into simpler sub-tasks that are easier to verify and easier to train a ML model to do.

1. Is the article crime-related?⁶ (classification)
2. Which sentences in the article describe the crime incident? (classification)
3. Which words refer to a location? (named entity recognition)
4. Which locations relate to the crime incident? (dependency parsing)
5. Which neighborhood is this in? (classification)

In graphical form:

    flowchart LR

text@{ shape: docs, label: "article text"}
texts@{ shape: docs, label: "sentences"}
label@{ shape: lin-doc, label: "Community area(s)" }
discardA([discard])
discardB([discard])

style RELA fill:#FFF,stroke-width:4px,stroke:#000,stroke-dasharray:2 2
subgraph RELA[Relevance Model]
    CrimeA@{ shape: diamond, label: "Crime-related article?" }
end

style RELS fill:#FFF,stroke-width:4px,stroke:#000,stroke-dasharray:2 2
subgraph RELS[Relevance Model]
    CrimeB@{ shape: diamond, label: "Crime incident-related sentence?" }
end

style DS fill:#FFF,stroke-width:4px,stroke:#000,stroke-dasharray:2 2
subgraph DS[NLP/Geo Models]
    NER
    CLF
end

text --> CrimeA
CrimeA --> |No| discardA
CrimeA --> |Yes| texts
texts --> CrimeB
CrimeB --> |No| discardB
CrimeB --> |Yes| NER
NER --> CLF
CLF --> label
  

For a more in-depth discussion of this model and implementation, see my next post.

Results

Coming Soon!

• crime
• media analysis
• nlp