💬 NLP

Ontology for Policing: Conceptual Knowledge Learning for Semantic Understanding and Reasoning in Law Enforcement Reports

This paper proposes a symbolic framework that converts unstructured law enforcement narratives into evidence-linked facts and temporal graphs by integrating semantic parsing, ontology mapping, and reasoning, demonstrating high accuracy in extracting key incident details from property crime reports.

Original authors: Anita Srbinovska, Jansen Orfan, Adrian Martin, Ernest Fokoué

Published 2026-05-18

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Anita Srbinovska, Jansen Orfan, Adrian Martin, Ernest Fokoué

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a police report as a two-part story. The first part is like a checklist: it has boxes for the date, the type of crime (like "Burglary"), and the case number. This is easy for computers to read because it's neat and organized.

The second part is the narrative: a paragraph written by the officer describing exactly what happened. This is where the real story lives—details about how a window was broken, who was involved, and the order of events. But because this is written in plain English, it's like a messy pile of puzzle pieces that a computer can't easily sort through without help.

This paper presents a new way to organize those messy puzzle pieces using a "symbolic" system (think of it as a very strict, rule-based translator) rather than a "guessing" system (like modern AI that predicts the next word).

Here is how their system works, broken down into simple steps:

1. The Privacy Filter (Redaction)

Before the computer reads the story, it puts on a pair of "privacy goggles." It scans the text and covers up sensitive names, addresses, and phone numbers, replacing them with generic labels like [PERSON_1] or [SUSPECT]. This ensures the system learns from the events, not the specific identities of the people involved.

2. The Translator (Semantic Parsing)

Once the text is safe, the system acts like a translator that doesn't just read words, but understands the structure of the action. It breaks sentences down into a "Who did What to Whom" format.

The Metaphor: Imagine the sentence "The suspect broke the window" is a sentence in a foreign language. The system translates it into a universal code: Action: Break, Doer: Suspect, Object: Window.
It uses a massive dictionary (called WordNet) and a grammar book (VerbNet) to make sure it understands that "stealing" is a type of "taking," and a "window" is a part of a "building."

3. The Rulebook (The Ontology)

The system then takes these translated facts and tries to fit them into a pre-built Rulebook (called an Ontology). Think of this rulebook as a strict filing cabinet with specific drawers labeled "Theft," "Forced Entry," or "Vehicle Damage."

If the system sees a "taking" action without permission, it tries to file it under "Theft."
If it sees a "breaking" action, it tries to file it under "Forced Entry."
The system checks its own work: "Does this fit the rules for a theft? Yes. Does it have a stolen item? Yes. Okay, I'm confident this is a theft."

4. The Timeline Builder (Temporal Reasoning)

Finally, the system looks for time clues like "then," "before," or "after" to build a timeline.

The Metaphor: It's like arranging a series of dominoes. If the text says, "The suspect broke the window before entering," the system places the "Breaking" domino before the "Entering" domino. It also uses common sense rules (axioms), like knowing that "breaking in" usually happens before "stealing."

What Did They Find?

The researchers tested this system on 450 real police reports about property crimes (like burglaries and stolen cars).

Success Rate: About 54% of the events the system identified were very confident (scored 0.80 or higher out of 1.0).
Accuracy: When human reviewers checked the system's work, they agreed 100% on three key things:
1. How the incident started.
2. What items were stolen.
3. The specific time clues used in the text.
The Struggle: The system had a harder time agreeing with humans on whether an entry was "forced" (like breaking a window vs. an open door). This was the most confusing part, with only 40% agreement.

The Bottom Line

The paper argues that by using strict rules and dictionaries instead of "guessing," this system can turn messy police stories into clear, organized facts that can be traced back to the original sentence. It's like turning a handwritten, messy diary into a structured database, helping investigators see patterns and timelines without having to read every single word manually.

Important Note: The authors are clear that this tool is for analysis and organization. It is not designed to make final decisions on investigations or to judge people; it simply helps organize the evidence so humans can review it better.

Technical Summary: Ontology for Policing – Conceptual Knowledge Learning for Semantic Understanding and Reasoning in Law Enforcement Reports

1. Problem Statement

Law enforcement incident reports consist of structured metadata (e.g., case numbers, offense codes) and unstructured natural language narratives. While structured data supports basic filtering and counting, critical investigative details—such as the chronological sequence of events, specific participant roles, methods of entry, and stolen items—are often embedded exclusively within the narrative text. Accessing these details systematically is difficult due to the volume of reports, linguistic variation, and the necessity of redacting Personally Identifiable Information (PII) for privacy. Current methods relying on shallow string matching or manual review create bottlenecks for investigators and analysts. The core challenge is to convert these redacted, unstructured narratives into "evidence-linked facts" that are machine-readable, auditable, and capable of supporting logical reasoning without losing traceability to the source text.

2. Methodology

The authors propose a symbolic framework that transforms police narratives into a formal ontology using a multi-stage pipeline. The approach avoids machine learning for the final reasoning layer, relying instead on symbolic natural language understanding (NLU) and knowledge representation.

A. Preprocessing and Redaction

Extraction: PDF reports are processed via Optical Character Recognition (OCR) to extract the narrative section, converting text to sentence case and removing artifacts.
Redaction: A rule-based system using spaCy named entity recognition, regular expressions, and metadata redacts PII (names, addresses, dates, vehicle details).
Pseudonymization: Entities are replaced with standardized placeholders (e.g., [PERSON_1], Victim, Suspect) while maintaining a mapping log to ensure traceability.

B. Semantic Parsing (The Semantic Layer)

The redacted text is parsed into Abstract Meaning Representation (AMR) graphs using the AMRlib parser (a BART-large model).

Graph Structure: AMR represents sentence meaning as a graph of predicates (events) and their arguments (participants).
Role Grounding: Pseudonyms are mapped to specific semantic types (e.g., Person, Vehicle) and roles (e.g., Agent, Patient) within the graph.
Lexical Resources: The system utilizes a chain of lexical databases to normalize predicates and arguments:
- PropBank: Provides verb senses and argument roles.
- VerbNet (via SemLink): Maps PropBank senses to semantic verb classes.
- WordNet: Provides hypernym hierarchies for "is-a" typing of arguments (e.g., distinguishing a "window" as a structural part).

C. Ontology Construction and Mapping

Extracted AMR graphs are mapped to an OWL 2 ontology built in Protégé.

Structure: The ontology includes a TBox (class schema), RBox (property schema), and ABox (instance facts). Key classes include Event, Entity, and ParticipationRole.
Mapping Rules: Predicate-argument structures (e.g., steal-01 :ARG0 x :ARG1 y) are converted into ontology assertions (e.g., TheftEvent(e), hasAgent(e,x)).
Description Logic (DL) Constraints: The system enforces structural invariants, such as ensuring every participation is linked to an event and that specific event types (e.g., TheftEvent) require specific participants (e.g., a stolen Item).

D. Event Typing and Confidence Scoring

The system assigns event types (e.g., ForcedEntryEvent, TheftEvent) and calculates a confidence score $c(e) \in [0,1]$ .

Two-Stage Typing:
1. Predicate-Family Stage: Generates coarse candidate types based on the verb family.
2. Argument-Sensitive Stage: Refines the type based on the semantic types of arguments (e.g., strengthening a "theft" interpretation if the object is a property item).
Confidence Calculation: The score is a heuristic sum of:
- Base Score: Determined by the event category (e.g., core incident events start at 0.55).
- Lexical Grounding: Bonus for full PropBank→VerbNet→WordNet mapping.
- Structural Support: Bonus for matching expected patterns (e.g., presence of a stolen item).
- Penalties: Deductions for ambiguity, negation, or uncertain language.
- Final Score: A weighted blend of the rule prior (70%) and the evidence-based score (30%).

E. Temporal Reasoning

The system constructs temporal graphs where nodes are events and edges are precedence relations.

Cue-Based Edges: Derived from temporal markers in the text (e.g., "then," "before," "after").
Domain Axioms: Logical rules enforce precedence based on police procedure (e.g., ForcedEntryEvent typically precedes TheftEvent).

3. Key Contributions

Evidence-Linked Fact Extraction: A pipeline that converts redacted narratives into structured facts traceable to specific sentences, addressing the bottleneck of manual review.
Symbolic NLU for Policing: A demonstration that symbolic methods (AMR + Ontology) can handle the linguistic variation and privacy constraints of police reports without relying on black-box machine learning models for the final reasoning.
Temporal Graph Construction: The integration of linguistic cues and domain axioms to reconstruct event timelines from unstructured text.
Auditable Confidence: A scoring mechanism that quantifies the reliability of extracted events, allowing analysts to filter results based on confidence thresholds.

4. Results

The framework was evaluated on a corpus of 450 property crime reports (Burglary, Larceny, Motor Vehicle Theft, Robbery, Stolen Property) from the Rochester Police Department.

Extraction Performance:
- The system extracted 6,686 events and 6,652 frames.
- 54.1% of extracted events achieved a confidence score $\ge 0.80$ .
- 93.7% of events were successfully mapped through the full PropBank→VerbNet→WordNet semantic path.
- ARG1 (patient/object) coverage was high (78.8%), while ARG0 (agent) coverage was lower (29.8%), reflecting the narrative tendency to explicitly name objects but obscure actor identities.
Frame Slot Filling:
- Entry points were filled in 41.8% of cases.
- Entry methods in 33.4%.
- Stolen items in 22.8%.
Temporal Ordering:
- 102 temporal edges were extracted across the corpus.
- 78.4% of these edges were derived from domain axioms, while only 21.6% came from explicit textual cues.
Human Review:
- A study with 6 reviewers on 5 redacted narratives showed 100% agreement between the system and human majority for: Incident initiation, Stolen items, Time cues, and Participant roles.
- Agreement was 80% for Vehicle involvement, Theft stated, and Overall confidence.
- Agreement was 40% for Forced entry, indicating this remains a challenging interpretation task.
- Precision/Recall/F1 scores for "Stolen items" and "Time cues" were perfect (100%).

5. Significance and Claims

The paper claims that symbolic NLU techniques can effectively recover incident details from redacted narratives that are missing from structured metadata. The significance lies in the ability to:

Scale Analysis: Automate the extraction of evidence-linked facts from large volumes of reports.
Maintain Privacy: Process redacted text without compromising the ability to reason about event structures.
Ensure Traceability: Every extracted fact is linked back to the source sentence and the specific semantic path used, allowing for auditability.
Handle Uncertainty: The confidence scoring system explicitly flags uncertain extractions (e.g., forced entry interpretations) rather than forcing a definitive classification, which is crucial for legal and investigative contexts.

The authors explicitly state that the system is designed to support investigation and review, not to replace human judgment or make conclusions about individuals without human oversight. The work extends the OpenBWC initiative by adding a symbolic NLU pipeline for internal police use.