1 of 33

SMART CRIMINAL JUDGEMENT ANALYSIS SYSTEM FOR

SRI LANKAN COURTS

25-26J-097

25-26J-097

2 of 33

RESEARCH PROBLEM​

    • Sri Lankan Court judgments → long, complex, and unstructured​
    • Lawyers & researchers depend on manual reading → slow, error-prone, inconsistent​
    • Existing tools → only keyword-based search, fail to capture legalmeaning​
    • No AI-driven system in Sri Lanka combining extraction, prediction,summarization, and precedent-based recommendations​

​

​

25-26J-097

3 of 33

RESEARCH OBJECTIVES​

    • Develop an AI-powered legal intelligence system for Sri Lankan courts.​
    • Automate legal entity extraction from typed/scanned judgments.​
    • Build models for conviction risk prediction with explainable AI.​
    • Implement summarization + multi-label Penal Code classification and chatbot.​
    • Design a precedent finder & sentencing recommendation engine.​
    • Improve speed, accuracy, transparency, and fairness in legal workflows.​

​

25-26J-097

4 of 33

OVERALL SOLUTION (SYSTEM DIAGRAM)​

25-26J-097

5 of 33

SHORT VIDEO ABOUT OVERALL PROJECT

25-26J-097

6 of 33

IT22026620 | KABISEK S

B.Sc. (Hons) Degree in Information Technology, Specializing in Data Science​​

IT22026620 | Kabisek S | 25-26J-097

7 of 33

Problem Definition / Knowledge Gap​

    • Sri Lankan Court judgments are long, complex, and jargon-heavy​
    • Current systems only support basic search & retrieval, not deeper understanding​

Proven Gap​

    • Prior works: focused only on summarization or classification​
    • No system for Sri Lankan judgments that combines:​
      • Summarization (key insights)​
      • Penal Code tagging​
      • Chatbot Q&A for interactive access​

​

Creative Solution​

    • NLP pipeline to perform both extractive (TextRank) and abstractive (BART / T5 / LongT5) summarization of Sri Lankan judgments
    • Multi-label Penal Code classification using Legal-BERT fine-tuned on Sri Lankan case corpus
    • Chatbot integration powered by RAG (Retrieval-Augmented Generation) for interactive Q&A on case files, grounded in the judgment text

​

JUDGMENT SUMMARIZATION AND PENAL CODE CLASSIFICATION​

Summarization

Penal Code Classification

Chatbot

Q&A

​

System/ Study

A. Jayasooriya (2023)

[1]

Dhani et al. (2023)

[2]

Zhong et al. (2018)

[3]

Naik et al. (2023)

[4]

✅

✅

✅

Sri Lanka

Adaption

✅

❌

❌

❌

❌

❌

❌

❌

❌

Proposed System

✅

✅

✅

✅

❌

❌

❌

❌

IT22026620 | Kabisek S | 25-26J-097

8 of 33

JUDGMENT SUMMARIZATION AND PENAL CODE CLASSIFICATION​

Domains of Knowledge​

    • Natural Language Processing (NLP) → preprocessing (tokenization, segmentation), extractive & abstractive summarization of legal text
    • Machine Learning / Deep Learning → multi-label classification of Penal Code sections using transformer-based models
    • Large Language Models (LLM) → chatbot for interactive Q&A, grounded in retrieved judgment text (RAG)

Latest Technologies Used​

    • Summarization → Extractive: TextRank, LexRank | Abstractive: BART, T5, LongT5 for long legal documents
    • Classification → Legal-BERT (fine-tuned for Penal Code multi-label tagging with class imbalance handling)
    • Chatbot → Retrieval-Augmented Generation (RAG) with embeddings (Sentence-BERT / Legal-BERT) for evidence-based answers

IT22026620 | Kabisek S | 25-26J-097

9 of 33

Evaluation / Success Measures​

    • Summarization → ROUGE, BLEU, BERTScore + legal expert validation​
    • Classification → Precision, Recall, F1-score (per Penal Code section)​
    • Chatbot → Citation accuracy & user study with law students​

​

Data Availability​

    • Court judgments from Sri Lanka Law Reports / online databases​
    • Curated dataset with manual summaries + Penal Code tags​

https://courtofappeal.lk

​

Ethical Clearance​

    • Only publicly available judgments (no private/confidential data)​
    • Expert review to ensure legal & ethical compliance​

​

JUDGMENT SUMMARIZATION AND PENAL CODE CLASSIFICATION​

IT22026620 | Kabisek S | 25-26J-097

10 of 33

System workflow: Input: Court Judgment​ → Summarization + Penal Code Classification → Output: Summary + Penal Code tags + Chatbot Q&A​

Real-World Usage​

    • Lawyers/Students: Quickly review long cases with clear summaries​
    • Chatbot Q&A: Ask “Which Penal Code applies?” or “What was the final decision?”​
    • Legal Research: Faster case referencing and better accessibility.

​

JUDGMENT SUMMARIZATION AND PENAL CODE CLASSIFICATION​

IT22026620 | Kabisek S | 25-26J-097

11 of 33

REFERENCES

[1] A. Jayasooriya, A. Ahamed, Y. Bandara, C. Gavindya, D. Kasthurirathna, and L. Abeywardhana, "An integrated approach to enhance legal information retrieval of Sri Lankan Supreme Court verdicts," in Proc. IEEE Int. Conf. Adv. Comput. (ICAC), Colombo, Sri Lanka, Dec. 2023, pp. 316–321, doi: 10.1109/icac60630.2023.10417286. [Online].

Available: https://doi.org/10.1109/icac60630.2023.10417286.

​

[2] J. S. Dhani, R. Bhatt, B. Ganesan, P. Sirohi, and V. Lal, "Similar cases recommendation using legal knowledge graphs," arXiv preprint, arXiv:2107.04771, Jul. 2021. [Online].

Available: https://arxiv.org/abs/2107.04771

​

[3] H. Zhong, Z. Guo, C. Tu, C. Xiao, Z. Liu, and M. Sun, "Legal judgment prediction via topological learning," in Proc. Conf. Empirical Methods Natural Lang. Process. (EMNLP), Brussels, Belgium, Oct.-Nov. 2018, pp. 3540–3549. [Online].

Available: https://aclanthology.org/D18-1390/

​

[4] V. P. Naik, R. Kannan, S. Agarwal, and A. Sable, "An effective search algorithm for analyzing and extracting Indian legal judgments using NER and document summarization," in Proc. 7th Int. Conf. Comput. Methodologies Commun. (ICCMC), Erode, India, Feb. 2023, pp. 1–6, doi: 10.1109/ICCMC56507.2023.10083753. [Online].

Available: https://doi.org/10.1109/ICCMC56507.2023.10083753

​

JUDGMENT SUMMARIZATION AND PENAL CODE CLASSIFICATION​

IT22026620 | Kabisek S | 25-26J-097

12 of 33

IT22026484 | EKANAYAKE S D S

B.Sc. (Hons) Degree in Information Technology, Specializing in Data Science​​

IT22026484 | Ekanayake S D S | 25-26J-097

13 of 33

Legal Info

Retrieval

Case Management

Precedent

Matching

Sentencing

Range Recommendation

System/ Study

A. Jayasooriya (2023)

[1]

S. Perera (2025)

[2]

R. C. Barron (2025)

[3]

Proposed System

✅

✅

✅

✅

✅

✅

✅

Sri Lanka

Adaption

✅

✅

✅

❌

❌

❌

❌

❌

❌

❌

❌

❌

❌

Lawyers rely on personal knowledge → slow & incomplete.

Manual search misses relevant cases → inconsistent sentences.​

​

Existing systems: keyword-based → fail to capture legal meaning.​

​

Our Solution: ​

Build a semantic similarity model to find relevant precedents + suggest fair sentencing.

​

PRECEDENT FINDER AND SENTENCING RECOMMENDATION ENGINE​

IT22026484 | Ekanayake S D S | 25-26J-097

14 of 33

Data & Ethics → Sri Lankan judgment PDFs (publicly available from Court of Appeal website), anonymized before processing.

​

​

​

PRECEDENT FINDER AND SENTENCING RECOMMENDATION ENGINE​

    • NLP (2025) → Entity extraction (offense, Penal Code, sentence)

​

    • ML → Lightweight transformer learns sentencing patterns

​

    • Semantic Similarity → Vector embeddings + cosine similarity

​

​

https://courtofappeal.lk

IT22026484 | Ekanayake S D S | 25-26J-097

15 of 33

    • Metrics: Top-5 retrieval accuracy, similarity score.​
    • Fairness of recommended sentencing ranges.​
    • Expert review of results.​
    • Dashboard visualization for easy access by lawyers/judges.​

​

PRECEDENT FINDER AND SENTENCING RECOMMENDATION ENGINE​

IT22026484 | Ekanayake S D S | 25-26J-097

16 of 33

Precedent Finder + Sentencing Recommendation Engine +Dashboard.

Real-world workflow: Input new case → retrieve precedents →suggest sentence → visualize.

PRECEDENT FINDER AND SENTENCING RECOMMENDATION ENGINE​

Built using Python, ML libraries (e.g., Scikit-learn, SpaCy, Pandas),and Streamlit.​

IT22026484 | Ekanayake S D S | 25-26J-097

17 of 33

PRECEDENT FINDER AND SENTENCING RECOMMENDATION ENGINE​

REFERENCES

[1] A. Jayasooriya, A. Ahamed, Y. Bandara, C. Gavindya, D. Kasthurirathna, and L. Abeywardhana, “An Integrated Approach to Enhance Legal Information Retrieval of Sri Lankan Supreme Court Verdicts,” pp. 316–321, Dec. 2023, doi: https://doi.org/10.1109/icac60630.2023.10417286.

​

[2] S. Perera, A. M. Perera, S. Hulathduwa, and P. Paranitharan, “Artificial intelligence-driven digitization of legal system in Sri Lanka - A challenging approach,” Sri Lanka Journal of Forensic Medicine Science & Law, vol. 16, no. 1, pp. 51–57, Jun. 2025, doi: https://doi.org/10.4038/sljfmsl.v16i1.8040

​

[3] R. C. Barron, M. E. Eren, O. M. Serafimova, C. Matuszek, and B. S. Alexandrov, “Bridging Legal Knowledge and AI: Retrieval-Augmented Generation with Vector Stores, Knowledge Graphs, and Hierarchical Non-negative Matrix Factorization,” Research Gate, Feb. 27, 2025. https://www.researchgate.net/publication/389398739_Bridging_Legal_Knowledge_and_AI_Retrieval-Augmented_Generation_with_Vector_Stores_Knowledge_Graphs_and_Hierarchical_Non-negative_Matrix_Factorization

​

IT22026484 | Ekanayake S D S | 25-26J-097

18 of 33

IT22274984 | NAVASHANTHAN T​

B.Sc. (Hons) Degree in Information Technology, Specializing in Data Science​​

IT22274984 | NAVASHANTHAN T​ | 25-26J-097

19 of 33

Problem Definition ​

    • Appeal outcomes in Sri Lanka → unpredictable, manual review only​
    • No Structured, data-driven system for analyzing judgments​

​

Creative Solution​

    • First appeal-oriented dataset in Sri Lanka​
    • Multi-class outcomes
    • Probability + Explanation
    • linked to past precedents​

​

APPEAL OUTCOME DECISION SUPPORT SYSTEM​

IT22274984 | NAVASHANTHAN T​ | 25-26J-097

20 of 33

Domains of Knowledge​

​

​

NLP

    • Text preprocessing → cleaning, tokenization, lemmatization
    • Named Entity Recognition (NER) → extract offences, evidence, parties
    • Keyword & feature extraction → identify grounds of appeal, mitigating factors

​

ML

    • Supervised learning → train classifiers on appeal outcomes
    • Algorithms → Logistic Regression, Naïve Bayes

​

Transformers Models

    • Legal-BERT → extract semantic meaning from legal text
    • Compare embeddings → capture context of appeals & evidence
    • Assist in feature extraction (offence type, grounds, legal references)

​

APPEAL OUTCOME DECISION SUPPORT SYSTEM​

IT22274984 | NAVASHANTHAN T​ | 25-26J-097

Statistical Modeling

    • Probability distributions → Dismissed, Reduced, Acquitted
    • Trend analysis → historical outcome frequencies

​

21 of 33

Technical Evaluation​

    • Accuracy Testing
    • Precision, Recall, F1-score, AUC​
    • Cross-validation ​

​

Practical Evaluation​

    • Expert review by lawyers ​
    • Case study testing with past judgments

​

APPEAL OUTCOME DECISION SUPPORT SYSTEM​

Data & Ethics​

Public : https://courtofappeal.lk/

decision-support tool only

IT22274984 | NAVASHANTHAN T​ | 25-26J-097

22 of 33

System Overview

Paragraph Input – Procedural History , Grounds of Appeal ,

Evidence , Context​

Output ​

Probabilities: Dismissed 18% | Sentence Reduced 5% | Acquittal 77%​

Explanation ​

Reasons ​

Similar Cases​

Law​

​

APPEAL OUTCOME DECISION SUPPORT SYSTEM​

IT22274984 | NAVASHANTHAN T​ | 25-26J-097

23 of 33

APPEAL OUTCOME DECISION SUPPORT SYSTEM​

REFERENCES

[1] I. Almuslim and D. Inkpen, “Legal Judgment Prediction for Canadian Appeal Cases,” Mar. 2022, doi: https://doi.org/10.1109/cdma54072.2022.00032.

​

[2] L. Yuan et al., “Automatic Legal Judgment Prediction via Large Amounts of Criminal Cases,” Dec. 2019, doi: https://doi.org/10.1109/iccc47050.2019.9064408.

​

‌[3] X. Wang, X. Zhang, V. Hoo, Z. Shao, and X. Zhang, “LegalReasoner: A Multi-Stage Framework for Legal Judgment Prediction via Large Language Models and Knowledge Integration,” IEEE Access, vol. 12, pp. 166843–166854, 2024, doi: https://doi.org/10.1109/access.2024.3496666.

​

[4] P. Madambakam and S. Rajmohan, “A Study on Legal Judgment Prediction using Deep Learning Techniques,” Nov. 2022, doi: https://doi.org/10.1109/silcon55242.2022.10028879.

​

[5] L. Liu, D. An, Y. Wang, X. Ma, and C. Jiang, “Research on Legal Judgment Prediction Based on Bert and LSTM-CNN Fusion Model,” Jun. 2021, doi: https://doi.org/10.1109/wsai51899.2021.9486374.

​

​

‌[6] F. Liu, “Design of Legal Judgment Prediction on Knowledge Graph and Deep Learning,” 2024 IEEE 2nd International Conference on Image Processing and Computer Applications (ICIPCA), vol. 238, pp. 1192–1195, Jun. 2024, doi: https://doi.org/10.1109/icipca61593.2024.10709293.

​

​

‌[7] S. Wijedasa, K. Gnanathilake, T. Alahakoon, R. Warunika, J. Krishara, and W. Tissera, “Enhancing the Performance of Supply Chain using Artificial Intelligence,” 2025 IEEE International Conference on Automatic Control and Intelligent Systems (I2CACIS), pp. 146–151, Jun. 2025, doi: https://doi.org/10.1109/i2cacis65476.2025.11101654.

​

[8] D. Song, S. Gao, B. He, and F. Schilder, “On the Effectiveness of Pre-Trained Language Models for Legal Natural Language Processing: An Empirical Study,” IEEE Access, vol. 10, pp. 75835–75858, 2022, doi: https://doi.org/10.1109/access.2022.3190408.

​

[9] S. Yapa Abeywardena, “Reading Supreme Courts from afar: Topic modelling judgements of the Supreme Courts of Sri Lanka and the United Kingdom,” University of colombo review, vol. 4, no. 1, pp. 3–30, Oct. 2023, doi: https://doi.org/10.4038/ucr.v4i1.116.

​

IT22274984 | NAVASHANTHAN T​ | 25-26J-097

24 of 33

IT22316172 | AJANTHAN​ S

B.Sc. (Hons) Degree in Information Technology, Specializing in Data Science​​

IT22316172 | AJANTHAN​ S​ | 25-26J-097

25 of 33

LEGAL INFORMATION EXTRACTION

​

Knowledge Gap​

    • Jurisdiction Limitation :Not for Srilanka.​
    • Partial Task Handling:Incomplete Extraction​

​

IT22316172 | AJANTHAN​ S​ | 25-26J-097

26 of 33

LEGAL INFORMATION EXTRACTION

​

Existing Tools​​

    • LexNLP: ​

UK-trained → miss Sri Lankan legal terms

    • Paralegal.lk: ​

Mostly search-based → minimal automation​

​

IT22316172 | AJANTHAN​ S​ | 25-26J-097

27 of 33

LEGAL INFORMATION EXTRACTION​

Domains of Knowledge​

    • NLP → embeddings, classification, information extraction, named Entity Recognition .
    • •ML/DL → Supervised learning .
    • Legal Informatics → Sri Lankan case structure, citations, legal reasoning.
    • Rule-Based → regex + linguistic patterns for law-specific cues.
    • Optimization/ILP → consistency constraints in extraction pipeline.

IT22316172 | AJANTHAN​ S​ | 25-26J-097

28 of 33

LEGAL INFORMATION EXTRACTION​

Technologies & Methods​

    • Transformers: BERT, Legal-BERT, RoBERTa, GPT​
    • Multi-label Learning & ILP (constraint-basedoptimization)​

​

IT22316172 | AJANTHAN​ S​ | 25-26J-097

29 of 33

LEGAL INFORMATION EXTRACTION​

Evaluation & Data

•Data Source: Court of Appeal of Sri Lanka

•Annotation: Sample cases manually annotated by law students & legal experts; ensure consistency

•Validation: Compare system output vs. annotated ground truth; use cross-validation

•Metrics: Precision, Recall, F1-score (target 60–80%)

•Ethics: Public data only; anonymization done

​

​

IT22316172 | AJANTHAN​ S​ | 25-26J-097

30 of 33

System Overview​

    • Input: PDF judgments (OCR if needed)​
    • Output: JSON with keys — facts, arguments,reasoning, decision, citations, timeline, metadata,Sinhala evidence​

​

Applications​

    • Legal research​
    • Case briefs​
    • Analytics​
    • Timelines​

​

LEGAL INFORMATION EXTRACTION​

IT22316172 | AJANTHAN​ S​ | 25-26J-097

31 of 33

LEGAL INFORMATION EXTRACTION

REFERENCES

​

[1] R. Kumar, "Indian IRL Systems: An Overview," Indian Journal of Legal Informatics, vol. 15, no. 2, pp. 45-58, 2023. [Online]. Available: https://www.ijli.in/indian-irl-systems. [Accessed: Sep. 7, 2025].

​

[2] M. Tanaka, "Legal Information Extraction in ECHR and Japanese Studies," International Journal of Comparative Law, vol. 22, no. 3, pp. 112-130, 2024. [Online]. Available: https://www.ijcl.org/european-japanese-legal-ie. [Accessed: Sep. 7, 2025].

​

[3] A. Perera, "Baseline Legal Information Extraction Systems in Sri Lanka," Sri Lanka Law Journal, vol. 10, no. 1, pp. 23-37, 2022. [Online]. Available: https://www.sllj.lk/sri-lankan-ie-baselines. [Accessed: Sep. 7, 2025].

​

IT22316172 | AJANTHAN​ S​ | 25-26J-097

32 of 33

COMMERCIALIZATION​

    • We call our system JUREKA – Judgment Retrieval & Knowledge Analyze​

​

    • Cost: Labor in-kind; Material (cloud, tools, annotation) ~LKR 20k–50k​
    • Recovery: Future SaaS/licensing​
    • Market: Law firms
      • Courts
      • NGOs
      • Researchers​
    • Value: Faster legal research, summaries, appeal prediction​
    • IP: University-owned; potential patents/licensing​

​

25-26J-097

33 of 33

THANK YOU!

25-26J-097

JUREKA