Forensic investigation relies on reconstructing how criminal events unfold in time, space, and human interaction. This task is inherently complex due to the involvement of multiple sources of evidence; however, it could benefit from information reported in judicial judgments of previous criminal cases, as such documents often contain detailed reconstructions of events that can support investigative reasoning and hypothesis generation. Unfortunately, judicial judgments embed the narrative of criminal events within legally dense and procedurally driven text, which hinders their straightforward reuse for forensic reasoning purposes. In this context, Artificial Intelligence (AI) techniques could facilitate the generation of clear and structured crime narratives from judicial records, making them suitable for use by law enforcement oficers and, in the future, by AI-assisted investigative systems. Based on these considerations, the goal of this paper is to present an architectural proposal in which judicial records can be responsibly repurposed into a forensic crime narrative corpus through a human-centered pipeline that prioritizes transparency, traceability, and selective expert oversight. In detail, this framework combines lexicon-guided recall, transformer-based functional classification, uncertainty-driven active learning, and lightweight event and timeline structuring. By generating a publicly available corpus that curates crime narratives from judicial records, this framework establishes a solid foundation for future research on AI-assisted investigative reasoning and mixed-reality crime-scene support systems.
Crime Scene Investigation (CSI) is a cognitively demanding process in which investigators synthesize
fragmented observations, witness accounts, and physical evidence to form coherent hypotheses about
past events [
Based on this consideration, publicly available judicial judgments could be a valuable and
underused source of structured narrative information for forensic investigation and AI-assisted forensic
reasoning. Indeed, many criminal judgments contain factual accounts that describe the conduct of
the ofense, victim-ofender interactions, and investigative observations of previous criminal events
that could support forensic reasoning and hypothesis generation. However, such narratives are
typically interwoven with legal reasoning, precedent discussion, and procedural commentary. Extracting
investigator-relevant narratives from these documents therefore requires careful linguistic separation
and systematic validation, based on insights from rhetorical role labeling and legal discourse analysis
[
In order to bridge this gap, the goal of this work is to propose a human-centered path toward a forensic
crime narrative corpus derived from judicial records, together with a hybrid Human-in-the-Loop (HITL)
modeling workflow that can isolate, validate, and structure narrative content with high precision and
feasibility. The core claim is not that the process should be fully automated, but that it should be
designed around selective human oversight to ensure reliability, accountability, and practical utility
[
The intended outcome is a dataset that supports functional annotations at the sentence-level that
distinguish narrative events, contextual descriptions, and procedural or legal discourse. In addition,
it is expected to include event templates that capture actors, actions, targets, locations, and temporal
ordering, drawing on principles from event extraction and narrative schema learning [19, 20]. Each
case is also expected to support a canonical narrative reconstruction with explicit links to the source
text in order to preserve provenance and auditability [
The proposed architecture follows an incremental and auditable pipeline designed to balance narrative
coverage, annotation eficiency, and robustness across heterogeneous judicial writing styles. The core
design principle is to maximize the recall of crime narrative content while limiting human efort through
selective verification and iterative refinement. Human oversight is incorporated where automated
confidence is insuficient, supporting scalability without sacrificing traceability [
In the following, all components of our framework are discussed in detail.
The source material consists of publicly available criminal judgments in which the conduct of the ofense and the interactions between the victims are described with suficient detail to support the narrative reconstruction. These documents vary substantially in length, rhetorical structure, and format. Some judgments provide explicit sections such as Facts or The Ofending, whereas others distribute narrative information across multiple paragraphs without clear demarcation. In this work, extraction is restricted
Judicial Case Records (Criminal Judgments & Appeals)
Document Segmentation & Windowing (Paragraphs/Section Cues/Sliding Windows)
Lexicon-Guided Recall Filter
(Behavioural & Event Triggers) Transformers Based Sentence Classifier
Model vt (NE/CS/PL/OT)
Prediction confidence
High Confidence Lightweight Event Extraction and Schema
Generation (Actor/Action/Target/Location/Time) Released Forensic Crime-Narrative dataset
Lexicon Dictionary L0 Lexicon Expansion
L ∪ ∆ t
Human-in-the-Loop Review
(Expert Validation)
High Uncertainaty
to factual descriptions of the crime incident and its immediate context. Legal principles, sentencing
rationales, and procedural arguments are treated as contextual metadata and excluded from narrative
modeling. This separation reduces the risk of conflating factual accounts with judicial interpretation
and aligns with established practices in the labeling of the rhetorical role of legal documents [
Each judicial judgment is segmented into paragraphs. Let a document be represented as an ordered sequence of paragraphs (1, 2, . . . , ). To preserve local narrative coherence while limiting contextual drift, overlapping paragraph windows of fixed length are constructed:
() = { = (, +1, . . . , +−1 ) | 1 ≤ ≤ − + 1 }.
Where available, section headings such as Facts, The Ofending, or Background are treated as weak structural signals to prioritize narrative-dense regions. When headings are unreliable or absent, sliding windows preserve coverage throughout the document.
An initial domain-aware lexicon 0 is constructed by forensic domain inspection. The lexicon includes
behavioral verbs, motion indicators, references to force or instruments, and expressions commonly
associated with crime narratives. The lexicon is used as a recall-oriented prior rather than as a strict
rule-based filter, which aligns with previous work on weak supervision and lexicon-guided annotation
[
A candidate window is retained if it contains at least one lexical trigger:
(; ) = ⊮[∃ℓ ∈ : ℓ ∈ ].
() = { ∈ () | (; ) = 1 }.
To reduce selection bias, windows that lack lexical triggers may still be sampled under the active learning regime so that previously unseen narrative expressions can be identified and incorporated.
Sentences within retained windows are classified according to their functional role in the document,
following rhetorical role labeling traditions [
A transformer-based encoder produces a contextual representation ℎ = (), which is passed to a linear classification head:
( = | ) = softmax( ℎ + ).
The model is trained using cross-entropy loss over manually verified samples, consistent with standard
ifne-tuning approaches in legal NLP [
Human review is focused on sentences with high predictive uncertainty, estimated using entropy: () = − ∑︁ ( = | ) log ( = | ).
Sentences exceeding a predefined uncertainty threshold are routed for expert annotation. Validated narrative expressions are incorporated into the lexicon through incremental updates: +1 = ∪ ∆ .
This iterative design follows established HITL and active learning practices that aim to improve
annotation eficiency while maintaining quality [
Sentences classified as Narrative Event are converted into structured event templates: = ( , , , ℓ , ), where denotes the agent, the action, the target, ℓ the location reference, and the temporal index. This representation is based on the principles of semantic role labeling and temporal event extraction [20, 17, 18].
Events are assembled into a case-level timeline: = (1, 2, . . . , ), 1 ≤ 2 ≤ · · · ≤ .
The representation is intentionally lightweight to ensure robustness across diverse judicial writing styles and to support scalable validation.
For each judicial case, the proposed framework returns a crime narrative dataset that includes structured metadata, sentence-level functional labels with confidence and verification status, event templates with temporal ordering, and narrative reconstructions linked to the source text. The release is intended to include annotation guidelines, inter-annotator agreement statistics [22], ofense-type coverage analysis, and baseline benchmarks.
In addition to the annotated data, a frozen sentence-level functional classification model trained on the ifnal verified corpus is released as a reference artifact. This model is provided to support reproducibility and comparative evaluation, and is not intended to represent an optimized or deployment-oriented system.
The evaluation is expected to cover functional classification performance, event extraction quality,
and annotation eficiency. Classification can be assessed using macro-averaged precision, recall, and
F1 score. Event extraction quality can be measured via slot-level accuracy for event arguments, and
eficiency can be quantified as a reduction in human annotation efort relative to non-HITL baselines
[
The resulting corpus is intended to support research on investigative decision support, precedentbased crime pattern comparison, and knowledge grounding for mixed-reality systems that assist investigators in reconstructing event sequences. These applications should be treated as advisory rather than as determinative. The framework is designed to preserve provenance links to enable inspection and to discourage ungrounded inference.
Judicial narratives are post hoc reconstructions and may reflect legal priorities rather than
investigative completeness. They may also encode biases embedded in judicial processes. These constraints
require careful interpretation and appropriate safeguards in downstream use [
The goal of this paper is to propose a human-centered approach to constructing a forensic crime narrative corpus from judicial records. By combining lexicon-guided recall, transformer-based functional classification, and uncertainty-driven selective human oversight, the proposed framework aims to address data scarcity while supporting transparency and auditability. The design is intended to provide a foundation for future empirical work on forensic narrative modeling and AI-assisted investigative reasoning.
Bilal Ahmed is the main contributor to this work and led the conceptualization of the study, the design of the proposed methodology, and the writing of the manuscript. Giovanni Acampora and Autilia Vitiello provided supervision, methodological guidance and critical feedback, with particular support on research formulation, machine learning design, and evaluation plan. All authors reviewed the manuscript.
[17] O. Kolomiyets, M.-F. Moens, Extracting narrative timelines as temporal dependency structures, in:
Proceedings of ACL, 2012. [18] W. Yao, B. Haghighi, H. Poon, S. Riedel, Temporal event knowledge acquisition via identifying before/after relations, in: Proceedings of ACL, 2018. [19] N. Chambers, D. Jurafsky, Unsupervised learning of narrative event chains, in: Proceedings of
EMNLP, 2008. [20] Z. Zhang, E. Strubell, E. Hovy, Transfer learning from semantic role labeling to event argument extraction, in: Proceedings of EMNLP, 2022. [21] P. Kalamkar, U. Sreejith, B. Nayak, K. Shrivastava, Corpus for automatic structuring of legal documents, in: Proceedings of the Language Resources and Evaluation Conference (LREC), 2022. [22] K. Krippendorf, Reliability in content analysis, Human Communication Research 30 (2004) 298–310. [23] S. Barocas, A. D. Selbst, Big data’s disparate impact, California Law Review 104 (2016) 671–732.