Editorial
Guide
Is AI Making Your Organization More Productive or More Dependent?
Key Takeaways
- Graph-based narrative analysis exposed a key inconsistency in a Supreme Court case that took years of appeals to uncover.
- The project identified major obstacles to automating legal analysis, including entity, ambiguity, translation differences and domain-specific vocabulary.
- The research found that domain-specific lexicons and manual validation are essential before scaling narrative extraction with AI.
- The long-term goal is to automate graph-based narrative extraction so that professionals can analyze complex, unstructured documents faster and more accurately.
In Part 1 and Part 2, we established the problem — knowledge-intensive domains drowning in unstructured narrative text — and walked through the pipeline architecture behind Visual Judgements (VJ): NLP preprocessing with NLTK and spaCy, graph construction in Neo4j using Cypher queries and the aggregation-disaggregation strategy that made complex judicial narratives visually interpretable.
This final installment examines the concrete results from applying this approach to a Supreme Court of India criminal appeal, catalogs the challenges that surfaced and outlines the path from manual proof of concept to automated extraction at scale.
What the Graphs Revealed
The aggregation-disaggregation strategy produced a set of directed sub-graphs in Neo4j, each capturing a distinct narrative dimension of Criminal Appeal No. 1785 of 2011.
The overview sub-graph mapped every entity in the case:
- At the centre sat the victim, Sanjay Sorle, connected by a "DECLARED" edge to a "Dead" node, which in turn linked to...
- The presiding judge via an edge labeled with the case citation
- Radiating outward, the eleven accused were organized by charge: five connected to a node representing IPC 302/149 and 148, and six connected to Section 3(2)(5) of the SC/ST Act
- Prime witnesses — Amit (who filed the FIR), Savitri Bai and Ganesh Prasad — were linked to a "Prime Witness" node with edges distinguishing their roles
- Dr. Nilay Jain, the medical examiner, was connected through a "POST_MORTEM" edge
Information that required reading dozens of pages was collapsed into a single visual network.
The victim sub-graph traced the chronological sequence: Deepak arriving at the marriage ceremony on a motorcycle, the altercation between Sanjay Sorle and Deepak, Deepak threatening Sanjay while leaving, the return with ten accomplices and the group assault.
Each attacker radiated inward toward the Sanjay Sorle node — Shanker, Suresh, Shiv and "Gudda @ Sudhanshu" connected by "ATTACKED" edges, while the sequence continued:
"AFTER_ATTACK → Injured → FIR_Lodged → Police Station → Taken_to → Hospital → DECLARED_DEAD → Dead."
This visual separation produced the study's most important finding. The two appellants appeared in the graph connected to the victim only through edges representing fist blows. The fatal knife wound to the chest — confirmed by the post-mortem report as the sole cause of death — traced to an entirely different cluster of accused.
The disconnect between the appellants' alleged actions and the documented cause of death was visible in the graph within seconds. In the textual record, this inconsistency had taken over a decade and multiple rounds of appeals — through the trial court, the High Court of Madhya Pradesh at Jabalpur (Criminal Appeal Nos. 315, 316, 333, and 590 of 1999) and finally the Supreme Court — to surface.
Related Article: AI-Assisted vs. Human-Curated Metadata: The Hybrid Approach That Actually Scales
The Domain Challenge Taxonomy
The manual construction process produced a systematic inventory of obstacles that any automated narrative extraction system must overcome. These challenges fall into categories recognizable across domains, but the Indian legal context gives each one a distinctive and instructive character.
Vocabulary Misalignment
Vocabulary misalignment between domain-specific terms and general purpose NLP tools was pervasive. The Brill tagger misclassified "gupti" (a concealed dagger) as a proper noun because the term does not exist in standard English lexicons. Verbs describing culturally specific actions and adjectives describing injury types required manual correction. The implication is that production-grade narrative extraction requires custom lexicons built from within the domain — not as an enhancement, but as a prerequisite.
Entity Ambiguity
Entity ambiguity proved structurally challenging. One accused appeared as "Shiv" and "Shiv Kumar" without clear indication of whether these were the same person. Another was referenced under two different alias combinations — "Chokhu @ Sudhanshu" and "Dablu @ Sudhanshu" — creating genuine uncertainty. The "@" symbol in Indian legal documents denotes an alias, but when aliases are inconsistently applied, even careful human readers struggle. Standard named entity recognition could not resolve these ambiguities.
Translation Variances
Translation-induced variance is endemic in a system operating across 121 languages. The FIR in this case was likely filed in Hindi or a regional dialect, translated for court proceedings and then referenced in an English-language Supreme Court ruling. Each layer introduced interpretive drift — not just in word choice, but in emphasis, framing and factual detail. The Supreme Court itself highlighted discrepancies between the original FIR and subsequent witness statements.
Documentation Inconsistencies
Documentation inconsistency across courts compounded all of the above. The numbering of accused, the assignment of charges and basic factual details varied between the trial court, the High Court and the Supreme Court records. Automated systems must be designed not just to tolerate these inconsistencies but to flag them — since the inconsistency itself is often the finding.
Transferable Design Principles
- Sentence-level granularity is essential. The POS-tagged data frame — with each row representing one sentence and each cell holding a word-tag tuple — proved to be the right unit of analysis. Coarser chunking loses the precision needed to attribute specific actions to specific entities.
- Decomposition before composition. The failed attempt to render the entire judgement as a single graph confirmed that the aggregation-disaggregation architecture is not optional. Complex narratives must be broken into thematic sub-stories, each built independently and linked through shared entity nodes.
- Domain lexicons must be built from the ground up. Standard NLP pipelines trained on news articles or web text do not transfer to Indian legal prose. The "gupti" misclassification was symptomatic of a fundamental vocabulary gap that will recur in any specialized domain.
- Manual first, automate second. The temptation to skip to automation is strong, but premature automation encodes unvalidated assumptions. The manual phase is where you discover what your automated system needs to handle.
The Automation Roadmap
The path from manual proof of concept to automated system is clearly defined, informed by the specific failures and successes of the first phase.
- Lexicon construction at scale is the immediate next step. At least one thousand judgements under IPC sections 300 through 309 — covering culpable homicide, murder and suicide — will be processed to extract and validate a comprehensive legal vocabulary accounting for regional, cultural and temporal variation.
- Information-theoretic sentence ranking draws on Shannon's concept of entropy to score sentences by information density, identifying passages that carry the most evidentiary weight regardless of word frequency. In legal text, the most important facts are often stated once, not repeatedly.
- Machine learning for relationship extraction will be explored through classical supervised classification and LSTM networks — designed for sequential data — to learn the syntactic rules that the manual phase identified. The goal is to encode domain expertise in a form that scales.
- Probabilistic modeling of latent relationships addresses the reality that not every narrative connection is stated explicitly. Bayesian networks and hidden Markov models offer frameworks for inferring unstated connections — moving from visualizing what is explicitly documented to detecting what is
- implicitly present.
- Expert validation through a Delphi process — structured consultation with legal practitioners, NLP researchers and domain specialists — will ensure outputs align with the actual information retrieval needs and decision-making behaviors of working lawyers.
Related Article: The Architecture of the Agentic Enterprise: Semantics, Governance & Safe Autonomy
The Broader Implication
This work began with a single criminal appeal, manually processed, producing sub-graphs in Neo4j that made a decade-old inconsistency visible in seconds. But the underlying insight is domain-agnostic: complex narratives buried in unstructured text are a universal bottleneck in knowledge work, and graph-based computational narratives — built through principled decomposition and domain-specific tooling — are a viable, powerful tool for overcoming it.
From Automation to Intelligence: How Leading Teams Are Rethinking Operations
Register
Ai4 2026
Register
How to Win the War for Agentic Citations: The AEO Playbook You Need Now
Register
Why Some Dealers Are Pulling Ahead With AI
Watch Now
The technology stack — Python, NLTK, spaCy, Neo4j, Cypher — is mature, open-source and accessible. What has been missing is the willingness to invest in the domain-specific groundwork: custom lexicons, entity resolution logic and the manual discovery phase that reveals what automation must handle.
That groundwork has now been laid for the Indian legal domain. The proof of concept validates the approach. The roadmap is defined. The next step is to scale and to demonstrate that the same architecture can serve any field where justice, safety or accuracy depends on seeing a story that the text alone cannot show.
This concludes the three-part series on extracting graphical computational narratives from unstructured text.
Learn how you can join our contributor community.