Building knowledge graphs of homicide investigation chronologies

Abstract

Homicide investigations generate large and diverse data in the form of witness interview transcripts, physical evidence, photographs, DNA, etc. Homicide case chronologies are summaries of these data created by investigators that consist of short text-based entries documenting specific steps taken in the investigation. A chronology tracks the evolution of an investigation, including when and how persons involved and items of evidence became part of a case. In this article we discuss a framework for creating knowledge graphs of case chronologies that may aid investigators in analyzing homicide case data and also allow for post hoc analysis of the key features that determine whether a homicide is ultimately solved. Our method consists of 1) performing named entity recognition to determine witnesses, suspects, and detectives from chronology entries 2) using keyword expansion to identify documentary, physical, and forensic evidence in each entry and 3) linking entities and evidence to construct a homicide investigation knowledge graph. We compare the performance of several choices of methodologies for these sub-tasks using homicide investigation chronologies from Los Angeles, California. We then analyze the association between network statistics of the knowledge graphs and homicide solvability.