Abstract
Gun violence in the United States is a critical public health crisis, yet research is often hindered by a lack of high-resolution, structured data. This paper introduces the Gun Violence Database (GVDB), a massive dataset designed to transform unstructured local news reports into a queryable database. By framing gun violence research as a multi-stage Natural Language Processing (NLP) task, we provide a path to bypass federal data restrictions and offer public health researchers the granular data needed for effective policy design.
Introduction
Research into the causes and prevention of gun violence is currently stymied by a "data vacuum." Federal funding restrictions have left national databases fragmented, often missing the demographic and locational details essential for epidemiological study. However, relevant data is abundant in the thousands of local news reports published daily. The challenge lies in the scale: manual cataloging is too slow and costly. This work proposes that NLP can bridge this gap by automating the extraction of incident details from unstructured text, turning local reporting into a powerful tool for social good.
The GVDB Pipeline
The authors developed a crowdsourced annotation pipeline to build the GVDB, consisting of four primary stages:
- Crawling: Automated daily scrapes of local news and TV websites across the U.S.
- Filtering: A high-recall text classifier identifies potential gun violence stories, which are then manually vetted to remove false positives.
- Annotation: Crowdsourced workers use a custom interface to highlight text spans corresponding to a database schema, including shooters, victims, locations, and weapons.
- Baseline Evaluation: Off-the-shelf NLP tools (e.g., Stanford CoreNLP) were tested against human "gold standard" annotations to establish a performance benchmark.
Framing the NLP Task
Extracting information from gun violence reports is a complex task that requires more than simple entity recognition. It necessitates a combination of Semantic Role Labeling (to distinguish shooters from victims), Coreference Resolution (to link different mentions of the same person), and Temporal Resolution.
The GVDB Dataset
At the time of publication, the GVDB included approximately 60,000 articles, growing by roughly 1,000 per day. A subset of 7,366 articles has been fully annotated for Information Extraction (IE) tasks. The dataset covers 1,512 U.S. cities and includes detailed fields such as participant age, race, gender, specific weapon types, and precise temporal information.
Experimental Results
Evaluation of an off-the-shelf IE system on the GVDB revealed a significant performance gap, particularly in identifying incident participants and weapons. While the system performed reasonably well on Dates and Times (approx. 69% precision), it struggled with the nuances of crime reporting.
As shown in the results, identifying Shooters and Weapons proved the most difficult, with strict precision falling as low as 5.8% and 2.1%, respectively. These results highlight that general-purpose NLP tools are currently insufficient for the specific demands of crime reporting, making the GVDB a necessary and challenging benchmark for the field.
Discussion and Future Work
The complexity of news reporting—where multiple articles may cover one event or reference past incidents—presents ongoing challenges. Future research directions include:
- Event Coreference: Recognizing when different articles describe the same incident to prevent double-counting.
- Cross-document Coreference: Linking individuals across different stories to track patterns of violence over time.
- Multimodal Integration: Using visual data from reports (like dashcam footage) to supplement missing textual details such as age or race.
By applying NLP to high-stakes policy issues, researchers can provide the data-driven foundation necessary to address gun violence as a public health priority.
Summary updated