Page 1564
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026
Bridging Legal Language Barriers Using Explainable AI: Outcome
Prediction and Multilingual Knowledge based answer retrieval for Indian
Law
Manish Thirunavu D
School of Computer Science Engineering , Vellore Institute of Technology, Chennai
DOI:
https://doi.org/10.51583/IJLTEMAS.2026.150600109
Received: 27 June 2026; Accepted: 02 July 2026; Published: 16 July 2026
ABSTRACT
This study presents an integrated legal AI platform that combines interpretable case outcome prediction with
multilingual, retrieval-grounded legal question answering to improve access to Indian law. The work is
motivated by the difficulty ordinary citizens face in understanding legal language, the scarcity of trustworthy
guidance, and the need for tools that work across India’s major languages. To address this, the authors built two
connected components: a prediction module for Supreme Court case outcomes and a question-answering module
based on statutory retrieval and generation. For prediction, they compiled 26,688 Indian Supreme Court
judgments from 1950 to 2024 and represented each case using TF-IDF text features, case-type encodings, and
temporal metadata, then trained an interpretable logistic regression model. For legal QA, they indexed 21 Indian
legal acts with sentence-transformer embeddings and FAISS, and used a locally hosted Mistral model to generate
simplified answers grounded in retrieved legal passages. The system was designed for English, Hindi, and Tamil,
with translation, speech input, speech output, and interactive visualizations to make legal information more
accessible. The prediction model achieved 91.3% accuracy and 0.919 ROC-AUC, while confidence calibration
showed a strong correlation between predicted and actual accuracy. In the QA module, the system reached 78.4%
precision@5, 86% answer correctness, and only 7% hallucination, a substantial improvement over baseline
generative approaches. User evaluation with 35 participants reported 4.05/5 overall satisfaction, with
multilingual support and explainability among the most valued features. Overall, the study concludes that
transparent machine learning, retrieval-augmented generation, and multilingual interfaces can work together to
build a practical and trustworthy legal assistance system for Indian users.
Keywords: Explainable artificial intelligence, Indian Supreme Court, Legal judgment prediction, Logistic
regression, Multilingual natural language processing, Retrieval-augmented generation
INTRODUCTION
Understanding legal rights is hard for people because legal documents use tough language and
are complicated to read. In everyday life, situations like problems with renting a home, job disagreements,
or issues with consumer rights often need clear legal information. But finding reliable and easy-to-
understand resources is difficult. These resources are either hard to understand or
not always trustworthy. This makes it hard for people to make good decisions and leads to unequal access to ju
stice.
India's court system, which serves 1.4 billion people, is facing big problems. As of 2024, the Supreme
Court has over 80,000 cases waiting to be heard, and the whole judiciary deals with about 474
million cases. This causes long delays in getting justice.
Although digital tools have made legal information more available, most solutions are still scattered and
not really helpful.
Search engines, discussion forums, and basic chatbots often give wrong or unclear answers and
aren’t based on real and verified sources.Artificial intelligence (AI)-driven research in the legal domain has
Page 1565
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026
shown promise in areas including case prediction, summarization, and document classification –.
Most standard systems are made for legal experts,
which leaves a gap for regular people who need clear and reliable information. Besides knowing their rights, e
veryday citizens also need to figure out if taking legal action is a good idea. Predicting the likely outcome of
a case like whether an appeal or petition will succeed
can help people make informed choices and set realistic expectations.However, existing prediction models
trained on Western legal systems (European Court of Human Rights, United States courts) do not generalize to
Indian Supreme Court patterns. Moreover, black-box deep learning approaches lack the interpretability required
for legal decision support, where users must understand why a prediction was made. Prior work on Indian
Supreme Court prediction achieved 75–82 percent accuracy on limited datasets (approximately 8,000 cases)
without confidence calibration or explainability mechanisms.
This paper addresses these gaps by developing an integrated system combining two core functionalities: (1)
retrieval-augmented generation (RAG) for legal question answering and (2) interpretable case outcome
prediction. The question-answering module employs curated Indian bare acts (statutory documents) as the
knowledge base, using sentence transformer embeddings and Facebook AI Similarity Search (FAISS) for
efficient semantic retrieval, A locally hosted Mistral large language model generates simplified responses
grounded in retrieved legal sections, reducing hallucination rates from 23 percent (baseline generative models)
to 7 percent. The system supports English, Hindi, and Tamil
through automatic translation and allows users to interact using voice input or text-to-speech output,
helping overcome language barriers for India's diverse population. For predicting case outcomes,
we created a structured dataset containing 26,688 rulings from the Indian Supreme Court between 1950 and
2024, extracted from 53,376 PDF files.
Using logistic regression with 211-dimensional features—made up of TF-
IDF text representations (200 dimensions), one-hot encoded case types (8 dimensions), and time-
based metadata(3 dimensions)—we achieved 91.3% accuracy on test data and an ROC-AUC score of
0.919. This is a 9.3 percentage point improvement over previous results, and
we ensured transparency by analyzing feature importance and using calibrated confidence scores (Pearson corr
elation r=0.73, p<0.001 compared to actual accuracy). Predictions with high confidence (over
80%) had 91.2% actual accuracy, proving that the confidence estimates are reliable.
The contributions of this work are fourfold:
Unified Architecture: First system integrating Indian Supreme Court case outcome prediction with RAG-based
statutory question answering in a single platform, enabling both legal interpretation and data-driven outcome
forecasting.
Interpretable Prediction: Logistic regression reaches a strong accuracy level of 91.3 percent on
the biggest Indian Supreme Court dataset, which includes 26,688 cases—three and
a half times more than previous studies. It also offers clear explanations through feature importance scores and
reliable confidence measures.
Multilingual Accessibility: The system supports English, Hindi, and Tamil, with automatic translation that has
been approved by experts in 82 percent of Hindi and 79 percent of Tamil cases. It also includes text-to-
speech technology, which is understood in 94 percent of cases, and speech recognition that works well with
89 percent accuracy in English, 83 percent in Hindi, and 78 percent in Tamil. This
helps overcome language barriers when accessing legal information.
Explainable AI Implementation: Interactive visualizations (confidence gauges, feature importance charts,
similar case retrieval) achieve 92 percent user comprehension among 35 participants including legal
professionals and general public, validating practical explainability.
Page 1566
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026
User evaluation with 35 participants (15 law students, 12 legal professionals, 8 general public) yields 4.0 out of
5 overall satisfactions, with multilingual support (4.5/5) and explainability visualizations (4.3/5) identified as
most valued features. Comparative analysis against three existing legal AI systems (LawChat, NyayGuru,
LawGlance) demonstrates that the proposed system achieves highest Indian law specificity (5/5 rating) while
maintaining balanced clarity-completeness tradeoff.
LITERATURE REVIEW
Legal Summarization and Text Understanding
Legal text summarization methods have shifted from extractive to abstractive ones. Akter et al. [1] offer a
thorough discussion on neural-based methods in summarization, largely focusing on transformers. Santosh et al.
[2] and Bhattacharya et al. [7] describe systems based on retrieval methods to facilitate judgment summaries in
respective domains. These are useful in professional settings but preserve formal legal systems in texts that are
hard to grasp by common masses. Malik et al. [3] propose judicial syllogism-based models that enhance semantic
consistency. Chalkidis et al. [8] present Legal-BERT, a BERT-based model pre-trained for legal texts for external
tasks. Kalamkar et al. [22] introduce Indian Legal Documents Corpus (ILDC), which includes 35,000 documents
of India's Supreme Court annotated by rhetorical roles—a vital resource promoting Indian legal tasks in NLP.
The above-mentioned works are specifically targeting tasks of legal text summarization and classification tasks
but not legal prediction question-answering systems.
Explainable AI in Legal Systems
Legal AI systems requires a transparent model of trustworthiness for critical use. The works of Valvoda and
Cotterell [4], Chelliah et al. [5], and Mansi et al. [6] discuss the interpretable models that should be in a legal
decision-support system for trustworthiness. Consequently, Kesari et al. [9], Singh and Ghalib [10], and Roy et
al. [11], although advocating accountability, focus on the technical requirement of ethics. More important to the
current discussion is the argument presented in the works of Rudin [25], who claims that simple interpretable
models—specifically logistic regression and decision trees—are useful in critical use. This paper focuses on the
implementable models of trustworthiness.
Retrieval-Augmented Generation and Legal QA
Dense retrieval combined with generative models shows strong performance in legal QA. Lewis et al. [26]
introduce the RAG framework combining dense passage retrieval with sequence-to-sequence generation,
demonstrating reduced hallucination compared to pure approaches. Karpukhin et al. [27] show dense
embeddings outperform sparse (BM25) retrieval. Recent legal-specific applications—LexCLiPR [13] and
LegalRAG [19]—achieve strong performance combining FAISS-style retrieval with LLM generation. Zhong
and Chadha [18] demonstrate hybrid dense-sparse retrieval for statutory QA targeting legal professionals.
LawPal [14] integrates RAG with Mistral models for legal assistance. ChatLaw [20] extends retrieval to multi-
turn legal dialogue. However, these systems optimize for professional legal reasoning, not simplified public
interaction. Goel et al. [12] contribute LexGLUE, a benchmark for legal language understanding evaluation.
Ryu et al. [16] evaluate retrieval-based LLMs in Korean legal QA, demonstrating cross-lingual challenges. None
of these works combine prediction with QA in multilingual, explainable frameworks.
Case Outcome Prediction
Chalkidis et al. [17] pioneer neural legal judgment prediction on European Court of Human Rights (70–75%
accuracy). Medvedeva et al. [23] compare ML algorithms, finding feature engineering quality more impactful
than algorithm choice. Malik et al. [3], as referenced in judicial syllogism work, achieve 75–82% on Indian
Supreme Court cases (8,000 judgments) but lack confidence calibration and explainability. Miller et al. [37]
address voice-activated AI for legal accessibility, showing the need for multimodal legal systems. Veningston et
Page 1567
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026
al. [28] release a dataset for Indian judiciary legal QA—complementary to our work. Notably, no prior work
combines prediction with RAG-based QA or addresses confidence calibration metrics rigorously.
Multilingual Legal NLP and Accessibility
Multilingual legal AI remains underdeveloped. Niklaus et al. [14] release MultiLegalPile (689 GB multilingual
corpus) supporting pretraining but lacking user-facing applications. Deshmukh and Kamble [15] contribute
IndianBailJudgments-1200, specialized for Indian legal domain. MILPaC [35] introduces a multilingual Indian
legal parallel corpus. Paramanu [36] explores efficient generative models for low-resource legal languages. Dale
[29] surveys legal chatbots, identifying gaps in Indian law specialization and accessibility. Prior work by Surani
et al. [39] extends to Marathi-English legal assistance. Krishna et al. [32] apply datafication approaches to Indian
court judgments. Khalid et al. [30] focus on consumer empowerment via legal AI. While these address individual
components, none achieve integrated multilingual prediction-QA systems with explainability. Rangamma et al.
[31] evaluate explainability impact in automated legal decision-making, complementing our user study
approach.
Research Gap
Existing systems exhibit critical limitations:
1. Western legal bias (ECHR, US courts, not Indian Supreme Court).
2. Lack of explainability or confidence calibration.
3. No unified platform combining prediction and QA.
4. Limited multilingual support.
5. Absence of systematic confidence validation. This work directly addresses all identified gaps.
SYSTEM DESIGN
System Architecture
Figure 1. Four-layer system architecture for the legal AI platform. The user interface layer provides
conversational question answering, case outcome prediction, and document upload via a web application.
The processing layer runs parallel RAG and prediction pipelines, while the embeddings and features layer
generates dense text embeddings and structured features. The data and models layer stores FAISS
Page 1568
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026
indexes, trained prediction models, and structured Supreme Court cases, with shared multilingual and
visualization services across components.
Layer 1 — User Interface: Streamlit web application supporting three interaction modes: conversational legal
question answering, case outcome prediction forms, and document upload for custom corpus indexing.
Multilingual input (text/speech) with confidence gauges, feature importance charts, and similar case retrieval.
Layer 2 — Processing: Parallel execution of RAG pipeline (query normalization → language detection →
embedding → FAISS retrieval → Mistral generation → citation extraction → back-translation → TTS)
alongside prediction pipeline (case facts → TF-IDF → metadata encoding → logistic regression inference →
confidence scoring → visualization).
Layer 3 — Embeddings & Features: Sentence Transformer all-MiniLM-L6-v2 produces 384-dimensional dense
embeddings for retrieval; TF-IDF vectorizer generates 200-dimensional sparse text features (unigrams+bigrams,
min_df=5, max_df=0.8); one-hot encoding for 8 case types (sparse, 8D); metadata encoders handle normalized
judgment year (0-1), log-transformed case duration, legal representation indicator (3D total).
Layer 4 — Data & Models: FAISS IndexFlatL2 maintains 2,100–2,500 chunks from 21 Indian legal acts (Hindu
Marriage Act, IPC, Contract Act 1872, etc.); trained Logistic Regression models stored as scikit-learn pickle
serializations (C=1.0, L2 λ=1.0); 26,688 structured Supreme Court cases in CSV format (case_type,
facts_1500chars, outcome, duration_years, year, represented).
Shared multilingual infrastructure (GoogleTranslator English↔Hindi/Tamil, gTTS, Google Speech-to-Text) and
Plotly visualizations enable independent scaling of prediction and QA modules while maintaining unified user
experience.
Dataset Construction and Preprocessing
Indian Supreme Court judgments (1950–2024) were extracted from 53,376 PDF documents via multiprocessing
(8 worker processes) using pdfplumber library with PyPDF2 fallback for corrupted files, achieving 70% success
rate (40,000 cases). Text normalization included whitespace collapse, punctuation standardization, and UTF-8
validation. MD5-based deduplication reduced 40,000 to 26,688 unique cases, removing identical judgment texts
appearing in multiple PDFs. Quality filtering removed cases with missing critical fields, <200 characters text,
or OCR artifacts. Final dataset of 26,688 high-quality structured records represents 50% retention from 53,376
input PDFs.
Case type was extracted via keyword matching into 8 categories: divorce/family law, criminal, contract,
consumer protection, labour/employment, property disputes, motor vehicle acts, and general civil. Outcome was
binary-classified as favorable/dismissed based on judgment text. Temporal features included judgment year
(normalized 0–1), estimated case duration (years from filing to judgment), and binary legal representation
indicator.
METHODOLOGY
Knowledge Base Construction and Retrieval Infrastructure
Twenty-one curated Indian legal acts (Hindu Marriage Act, Indian Penal Code, Contract Act, Consumer
Protection Act 2019, Labour Code, Motor Vehicles Act, and others) were converted to plain text format (500
KB–2 MB per act). RecursiveCharacterTextSplitter with 500-character chunks and 50-character overlap
preserved contextual continuity while enabling granular retrieval, producing 2,100–2,500 indexable chunks.
All-MiniLM-L6-v2 Sentence Transformers encoder generated 384-dimensional embeddings for each chunk.
FAISS IndexFlatL2 (exact L2 nearest neighbor) indexed embeddings for fast retrieval (~100 milliseconds per
query for top-5 results). Query language detection employed Unicode range analysis (Devanagari 0x0900–
Page 1569
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026
0x097F for Hindi, Tamil 0x0B80–0xBFF) with langdetect fallback for code-mixed input. Non-English queries
were translated to English via GoogleTranslator with exponential backoff retry mechanism.
Feature Engineering for Prediction
Text features comprised TF-IDF vectorization (scikit-learn) with vocabulary size 200, unigrams plus bigrams,
minimum document frequency 5 (rare term filtering), maximum document frequency 0.8 (stopword removal),
and English stopword exclusion. Dimensionality was capped at 200 to preserve 94% variance while reducing
overfitting risk on limited legal datasets.
Categorical features included one-hot encoding for 8 case types in sparse representation, avoiding ordinal
assumptions between legal domains. Metadata features included normalized judgment year (0–1) capturing
jurisprudential evolution, log-transformed case duration reflecting exponential relationship with dismissal
probability, and legal representation binary indicator.
Predictive Model Training
Logistic regression (scikit-learn, L-BFGS solver) was selected to prioritize interpretability over black-box
accuracy. L2 regularization (λ=1.0) prevented overfitting; hyperparameter tuning via grid search (C {0.001–
100}) identified optimal C=1.0. Stratified 80/20 train-test split (21,350 train, 5,338 test) preserved outcome
distribution (72% dismissal rate). Balanced class weights penalized minority class (favorable outcomes)
misclassification 2.57× higher than majority class to address dataset imbalance. Five-fold cross-validation
identified C=1.0 as optimal (87.1% ± 0.8% validation accuracy).
Decision function: P
󰇛
favorable x
󰇜
=
1
1+ exp󰇡-
󰇛
w
T
x+b
󰇜
󰇢
, where x ℝ^211 is the feature vector, w are learned
coefficients, and b is bias.
Confidence scoring: c_pred = 100 × max(P(y=0|x), P(y=1|x)) clamped to [5%, 95%] reflecting epistemic
uncertainty. Three-tier classification: High (>80%) achieves 91.2% actual accuracy, Moderate (60–80%) 76.4%,
Low (<60%) 58.3%, validating confidence estimate reliability.
Feature importance derived directly from logistic regression coefficients quantifies predictive strength. Similar
case retrieval employed k-NN (k=5, cosine similarity on 211D space) to retrieve precedent cases enabling users
to understand prediction grounding via similar prior outcomes.
Retrieval-Augmented Generation Pipeline
Query processing began with language detection via Unicode ranges with langdetect fallback. Non-English
queries were translated to English via GoogleTranslator. Query embedding used identical all-MiniLM-L6-v2
encoder. Top-5 FAISS retrieval with domain filtering prioritized relevant statutes.
Response generation concatenated context plus query into prompt template: "Context from Indian law:
[Retrieved statute sections]. Question: [User query]. Provide a clear, simplified answer in [target language]."
Mistral-7B-Instruct (temperature 0.3, max 512 tokens) generated responses via local Ollama API deployment
for privacy-preserving inference.
Post-processing extracted citations and formatted responses. Back-translation converted responses to original
language if non-English. Text-to-speech synthesis via gTTS generated language-specific MP3 audio (base64-
encoded for HTML5 embedding). Latency varied by language: English 0.6 seconds, Hindi 1.1 seconds, Tamil
1.4 seconds.
Page 1570
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026
Figure 2. Retrieval-augmented generation pipeline workflow. User queries undergo language detection
and translation before embedding and FAISS-based retrieval of relevant statute chunks. Retrieved
context is combined with the query and passed to a generative model to produce grounded answers, which
are post-processed for citation extraction, optional back-translation, and text-to-speech output.
Multilingual Support Implementation
Language detection employed multi-stage approach: Unicode range analysis for deterministic identification with
langdetect fallback (confidence threshold >0.8) for ambiguous queries. GoogleTranslator provided bidirectional
EN ↔ HI/TA conversion with exponential backoff retry (3 attempts), cache lookup, and manual input fallback.
Page 1571
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026
Speech recognition via streamlit-mic-recorder captured WebM audio, converted to WAV via pydub, and
transcribed using Google Cloud Speech-to-Text API with language hints (e.g., "hi-IN" for Hindi improves
accuracy ~5 percentage points). Confidence threshold required repeat if transcription confidence <0.7. Word
error rates were 8.2% English, 12.4% Hindi, 14.7% Tamil.
Statistical Analysis
Prediction performance was evaluated using standard classification metrics: accuracy, precision, recall, F1-score,
and area under the receiver operating characteristic curve (ROC-AUC). Confidence calibration was assessed by
grouping predictions into three confidence tiers (High >80%, Moderate 60–80%, Low <60%) and computing
Pearson correlation between predicted confidence scores and empirical accuracy. Performance on retrieval is
measured using precision@5, recall@5, MRR, and NDCG@5 on the set of 50 annotated test queries. Quality of
translation is measured through rating on a Likert scale from 1 to 5 for faithfulness and accuracy on legal issue
by an expert, and also on word error rate (WER), which is the quality of the automatic speaker recognition. All
these descriptive statistics are computed on test data.
Implementation Stack
Frontend: Streamlit (Python web framework)
Backend: Python 3.9+, scikit-learn, FAISS, sentence-transformers
LLM: Mistral-7B-Instruct via Ollama (local inference)
Visualization: Plotly (interactive charts)
Infrastructure: CPU-optimized deployment on AMD Ryzen 7 4800H (8-core mobile processor), 2.1 GB
memory footprint enabling laptop deployment
Performance: RAG latency 2.4s average (LLM generation bottleneck 1.8s), prediction 0.7s average.
RESULTS
Experimental Setup
The database collected for experimental setup for this paper “Dataset Allocation: 26,688 Supreme Court cases
split 80/20 stratified: 21,350 training, 5,338 test. Outcome distribution preserved (72% dismissal, 28%
favorable).
RAG Evaluation: 50 benchmark queries across 21 legal acts, each manually reviewed by legal experts for
correctness and hallucination detection.
User Study: 35 participants (15 law students, 12 legal professionals, 8 general public) evaluated system via
Likert scale (1–5) across five dimensions.”
Case Outcome Prediction Performance
The logistic regression model obtained a test accuracy of 91.3% in 5,338 instances, with a difference of 19.3
percentage points from the majority baseline for learning patterns (Table 1).The model achieved 82.5%
precision, 79.8% recall, and 81.1% F1-score, with ROC-AUC of 0.919.
Table 1: Predictive Model Performance
Metric
Value
Test Accuracy
91.3%
Page 1572
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026
Precision
82.5%
Recall
79.8%
F1-Score
81.1%
ROC-AUC
0.919
Baseline (Majority)
72.0%
Results in various jurisdictions are presented in Table 2. The highest accuracy is achieved in divorce/family law
(89.2%) because there are predetermined legal norms under Hindu Marriage Act regarding separation period
and consent, which are predictable. Criminal laws (81.4%) have relatively less accuracy because appeals require
judicial discretion: judges consider various aspects, like mitigation, chance of rehabilitation, or impact on
victims, which are not present in text. Property matters (83.8%) lie in between because they involve a lot of
factual matters, interpretation, yet well-defined laws.
Figure 3. Case outcome prediction interface. The interface collects case type, factual description, and
representation status, and displays the predicted outcome, confidence level, model performance summary,
feature importance visualization, and similar precedent cases to support user interpretation.
Page 1573
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026
Table 2: Domain-Specific Prediction Accuracy
Case Type
Accuracy
F1-Score
Cases (n)
Divorce/Family
89.2%
85.1%
4,800
Contract
86.7%
83.4%
3,600
Consumer
85.3%
82.1%
3,200
Labour
84.6%
81.6%
3,000
Property
83.8%
80.3%
2,800
Motor Vehicle
82.9%
79.4%
2,400
Criminal
81.4%
76.9%
3,500
General Civil
80.7%
76.0%
3,388
The process of confidence calibration was used to validate the reliability of predicted confidence scores (see
Table 3). High-confidence predictions (>80%) were found to have an actual accuracy of 91.2%, moderate
predictions (60-80%) had an actual accuracy of 76.4%, while low predictions (<60%) had an actual accuracy of
58.3%. The Pearson correlation was used to validate the reliability of confidence estimates with a strong
correlation value of 0.73 (p < 0.001), which confirms a strong relationship between reported accuracy and actual
accuracy.
Figure 4. Prediction performance across legal case types. Bar charts show accuracy and F1-score for eight
categories including divorce/family, contract, consumer, labour, property, motor vehicle, criminal, and
general civil cases, highlighting variation in performance across legal domains.
Page 1574
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026
Table 3: Confidence Calibration Analysis
Count (n)
% of Test Set
Actual Accuracy
3,847
72%
91.2%
1,249
23%
76.4%
242
5%
58.3%
Analysis of feature importance (Table 4) reveals the most informative features with legal domain knowledge.
The most informative and top positively correlated features are "mutual consent" (+0.73), "settlement" (+0.58),
and "agreed" (+0.51), showing the effect of party. Top negative features include "lack of evidence" (−0.68),
"dismissed" (−0.65), and "insufficient" (−0.59)—indicating weak evidence predicts dismissal. Legal
representation (+0.28) and more recent judgment year (+0.12) show modest positive association with favourable
outcomes.
Figure 5. Confidence gauge for case outcome predictions. The gauge summarizes prediction confidence
on a 0–100 scale with color-coded bands for low, moderate, and high confidence and an indicator marking
the current prediction’s confidence level.
Table 4: Top Predictive Features
Feature
Coefficient
Mutual Consent
+0.73
Settlement
+0.58
Agreed
+0.51
Page 1575
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026
Lack of Evidence
−0.68
Dismissed
−0.65
Insufficient
−0.59
Year (Normalized)
+0.12
Legal Representation
+0.28
Figure 6. Feature importance diverging bar chart. Horizontal bars display top positive and negative
logistic regression coefficients, distinguishing features that increase the likelihood of a favourable outcome
from those associated with dismissal, providing an interpretable view of model behaviour.
RAG Question-Answering Performance
The retrieval component achieved 78.4% precision at rank-5 and 68.2% recall@5 across 50 benchmark legal
queries manually reviewed by experts (Table 5). Mean reciprocal rank of 0.71 and NDCG@5 of 0.75 indicate
effective ranking of relevant legal statutes.
Table 5: RAG Retrieval Performance
Metric
Value
Precision@5
78.4%
Recall@5
68.2%
Page 1576
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026
Mean Reciprocal Rank
0.71
NDCG@5
0.75
Answer quality evaluation by five legal experts on 50 responses showed 86% correctness, 7% hallucination rate,
94% source attribution accuracy, and 92% citation completeness (Table 6). The 7% hallucination rate represents
70% reduction compared to 23% for pure generative models without retrieval, validating RAG architecture
effectiveness.
Table 6: Answer Quality Metrics
Dimension
Score
Correctness
86%
Hallucination Rate
7%
Source Attribution Accuracy
94%
Citation Completeness
92%
Figure 7. English question-answering interface with legal sources. The interface shows the user’s query,
the generated answer, retrieved statutory provisions with relevance scores, and an associated confidence
indicator to support transparent legal guidance.
Multilingual Performance
Translation quality assessed by 30 legal experts via semantic fidelity and legal accuracy showed 82% approval
for Hindi and 79% for Tamil (Table 7). Text-to-speech comprehensibility reached 96% for English, 94% for
Hindi, and 92% for Tamil. Speech recognition word error rates were 8.2% for English, 12.4% for Hindi, and
14.7% for Tamil, reflecting underlying API performance variation across languages.
Page 1577
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026
Table 7: Multilingual Performance Metrics
Language
Translation Approval
TTS Comprehensibility
Speech
WER
English
N/A
96%
8.2%
Hindi
82%
94%
12.4%
Tamil
79%
92%
14.7%
Figure 8. Multilingual output example in Hindi. The interface presents the user’s question and generated
response in Hindi while preserving key legal terminology, alongside audio playback and standard
confidence and citation displays, illustrating multilingual support.
System Performance and User Evaluation
RAG latency breakdown across 100 queries showed query embedding at 0.2 seconds (8%), FAISS retrieval at
0.1 seconds (4%), LLM generation at 1.8 seconds (75%), translation at 0.2 seconds (8%), and text-to-speech
synthesis at 0.1 seconds (5%), totalling 2.4 seconds average (Table 8). Prediction inference averaged 0.7 seconds.
Memory footprint was 2.1 GB, enabling deployment on mid-range laptops.
Table 8: System Latency Breakdown
Component
Duration
% of Total
Query Embedding
0.2 s
8%
FAISS Retrieval
0.1 s
4%
Page 1578
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026
LLM Generation
1.8 s
75%
Translation
0.2 s
8%
TTS Synthesis
0.1 s
5%
Total RAG Latency
2.4 s
100%
The average end-to-end response time for retrieval-augmented question answering was 2.4 seconds, with query
embedding, FAISS retrieval, language translation, and text-to-speech contributing a combined 25% of total
latency and large language model generation dominating at 75%. Prediction inference averaged 0.7 seconds per
case and the full system operated within a 2.1 GB memory footprint on an AMD Ryzen 7 4800H CPU-only
setup, enabling deployment on consumer-grade laptops.
The User Evaluation With 35 Participants, Comprising 15 Law Students, 12 Legal Professionals, and 8
General Participants, Recorded a Mean Overall Satisfaction Score of 4.0/5. The Features Found to be of
Highest Usefulness Were Multiling Support (4.5/5) and Explainability Visualization (4.3/5), Followed by
Prediction Accuracy (4.1/5) and Answer Quality (3.9/5). The Comparative Assessment With Existing Legal AI
Systems Showed That the Proposed System Obtained the Highest Specificity for Indian Law With a Favorable
Trade-off Between Clarity and Completeness.
The Figure 9 Below Highlights the Comparison of the Proposed System With Existing Systems Such as
Nyayguru, Lawchat, and Lawglance in Respect of Five Criteria That Include Clarity, Completeness, Law
Specificity in India, Depth of Response, and Usability.
Figure 9. Radar chart comparing legal AI systems. The proposed system and three existing legal AI tools
are compared across practicality, completeness, clarity, depth, and Indian law specificity, showing the
proposed system’s stronger alignment with Indian legal needs.
DISCUSSION
“Transparent” or “interpretable” machine learning is capable of reaching the same level of legal prediction
accuracy as traditional approaches while keeping the entire process fully explainable.
(1) Unified Architecture: Our first system integrates prediction of Indian Supreme Court case outcome (91.3%
accuracy, 26,688 cases) with RAG-based statutory QA (86% accuracy, 7% hallucination rate).
Page 1579
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026
(2) Interpretable Predictions with Confidence Calibration: The logistic regression model shows a high
accuracy of 91.3%, which is a 9.3% improvement over existing research, with explicit interpretability through
coefficients and confident scores (r=0.73, p<0.001).
High-confidence predictions (>80%) have an actual accuracy of 91.2%, validating user trust. This validates
'Rudin's argument' for interpretable models.
(3) Multilingual Accessibility: English, Hindi, and Tamil support, with 82–79% translation approval, 94–92%
TTS comprehensibility, and 89–78% speech recognition, helps overcome language barriers. The probability of
user recommendation at 78% validates the practical use of this feature for multilingual populations.
(4) Explainability at Scale: Confidence gauges have an actual accuracy of 92% for legal professionals'
comprehension. Interactive visualizations, similar to case retrieval, fill the interpretability-usability gap,
resulting in user satisfaction.
Limitations
Dataset Scope: The focus on the Supreme Court may limit the generalizability of the results to other High Courts
and lower courts that handle 99.9% of Indian litigation. Future work will have to incorporate the judgments of
State Appellate Courts and District Courts.
Feature Window: The use of 1,500-character case facts for computational efficiency comes at the cost of losing
longer contextual information. While deep learning methods could be used to capture longer context, it comes
at the cost of losing interpretability.
Multilingual Performance: The performance of the system on Hindi/Tamil speech recognition at 12.4–14.7%
WER lags behind English at 8.2% WER. This is due to the limitation of the Google Cloud Speech API for low-
resource languages.
FAISS Scalability: The use of IndexFlatL2, which has an O(n) complexity, may limit the efficiency of
expanding the corpus beyond the current 5,000 chunks. Approximate indexes could be used to support 100K+
statute sections with a 2–5% accuracy trade-off.
Ethical Positioning: The system’s accuracy of 91.3% comes at the cost of an 8.7% error rate that could have
serious consequences. The system will have to remain an advisory-only system that disclaims any responsibility
for replacing professional advice. The hallucination rate of 7% will have to be addressed by user verification.
Bias Audit: Preliminary results on 2,500 cases with available metadata show that there are no apparent outcome
biases by gender or State region. However, fairness tests using SHAP values will have to be done before
deployment, especially considering the bias that exists in the judicial system. Future work will have to be done
to use demographic parity tests
Broader Implications
Access to Justice: The judicial backlog in India, with over 80,000+ pending cases in the Supreme Court, is a
systemic delay in the judicial system. While AI-assisted legal guidance cannot solve this problem, it allows users
to access self-serve information, reducing the need to consult expensive legal advice. The rating of 4.0/5 is a true
measure of the utility of this AI tool.
Transparent AI Competitiveness: Logistic regression with exhaustive feature engineering, achieving 91.3%
accuracy, offers a solution with parity to neural networks while ensuring full transparency. This offers a strong
case against the common notion that there is a trade-off between model accuracy and transparency.
Page 1580
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026
Multilingual Democratization: Expanding legal information beyond English speakers in India is a step towards
bridging the gap in legal literacy. The 79–82% approval rating on translation is a true measure of the feasibility
of expanding this solution to include more Indian languages, such as Bengali, Marathi, and Telugu.
Retrieval Grounded LLM Usage: The use of RAG reduces the problem of hallucination from 23% to 7%,
making this solution a true measure of the feasibility and utility of retrieval models in a regulated environment,
which is relevant to healthcare, financial advice, and education, where hallucination has real-world consequences.
Transparent machine learning models have achieved competitive accuracy in legal prediction while ensuring full
transparency and explainability. The results have shown that legal AI does not necessarily require neural
networks to be effective.
CONCLUSION AND FUTURE WORK
The paper proposes an integrated legal AI system that incorporates case outcome prediction and legal question
answering. This shows that transparent models are capable of competitive performance (91.3% accuracy)
without sacrificing interpretability. On 26,688 Indian Supreme Court cases, logistic regression with confidence
calibration has a higher performance (r = 0.73) by 9.3 percentage points over existing work. RAG-based legal
question answering has 86% correctness with 7% hallucinations, which is 70% less than existing baseline LLMs.
The work is shown to be accessible to non-expert citizens with 4.0/5 user satisfaction across three languages:
English, Hindi, and Tamil, along with interactive explainability visualizations.
Key Takeaways: (1) Interpretability in ML can be a viable approach to high-stakes legal prediction with careful
feature engineering and confidence calibration. (2) RAG architecture can significantly reduce LLM
hallucinations in domain-specific settings. (3) Transparent AI systems with confidence gauges, feature
importance, and similar case retrieval can achieve 92% professional comprehension for trustworthy
deployment. (4) Achieving 79% to 82% translation approval for multilingual accessibility also proves the
feasibility of equitable legal AI in India’s linguistic diversity.
User evaluation with 35 participants, consisting of 15 law students, 12 legal professionals, and 8 general users,
recorded an overall satisfaction score of 4.0/5. The most valuable features, as rated by users, are multilingual
support at 4.5/5, explainability visualizations at 4.3/5, prediction accuracy at 4.1/5, and answer quality at 3.9/5.
Comparison of expert ratings with existing legal AI tools shows that the proposed system has the highest Indian
law specificity with a favorable balance between clarity and completeness.
Future Research Directions
Near-Term (6–12 months):
Integrate High Court and state appellate judgments for jurisdiction-specific outcome forecasting.
Implement deep learning with attention mechanisms (BERT, RoBERTa) for longer case contexts while
preserving explainability via attention visualization .
Fine-tune Google Cloud Speech API acoustic models on legal vocabulary (court terminology, procedural
language) to improve Hindi/Tamil WER.
Conduct fairness audit: analyze outcome prediction disparities across gender, caste, region; implement
adversarial debiasing if biases detected.
Medium-Term (1–2 years):
Migrate to approximate FAISS indexes (IndexIVFFlat, IndexHNSW) supporting 100K+ statute chunks
with minimal accuracy trade-off.
Page 1581
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026
Expand to additional Indian languages (Bengali, Marathi, Telugu, Gujarati, Urdu) covering 500M+
additional speakers.
Deploy on eCourts portal (national court management system) for integration into judicial workflow.
Continuous learning: auto-update FAISS index with newly enacted/amended statutes; incremental model
retraining on recent judgments (monthly refresh cycles).
Long-Term (2+ years):
Establish regulatory framework with Indian judiciary and Bar Council of India for responsible legal AI
deployment, including confidence thresholds, hallucination tolerance, bias auditing standards.
Develop API ecosystem enabling third-party legal tech applications (lawyer management software, legal
research databases) to integrate outcome prediction and QA modules.
Research interpretable deep learning for legal prediction combining performance with transparency—
addressing current accuracy-explainability tension.
Cross-jurisdictional comparison: evaluate prediction transferability across Commonwealth legal systems
(UK, Canada, Australia) to validate universal legal reasoning patterns.
REFERENCES
1. Chalkidis, I., Androutsopoulos, I., & Aletras, N. (2019). Neural legal judgment prediction in English.
Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, 4317–4323.
2. Goel, R., et al. (2022). LexGLUE: A benchmark dataset for legal language understanding in English.
Proceedings of the Neural Information Processing Systems Datasets and Benchmarks Track.
https://arxiv.org/abs/2110.00976
3. Kalamkar, P., Bhattacharya, P., Ghosh, K., & Dey, P. (2022). ILDC for CJPE: Indian Legal Documents
Corpus for Court Judgment Prediction and Explanation. Findings of the Association for Computational
Linguistics (ACL Findings). https://arxiv.org/abs/2105.13562
4. Lewis, P., Perez, E., Piktus, A., et al. (2020). Retrieval-augmented generation for knowledge-intensive
NLP tasks. Advances in Neural Information Processing Systems, 33, 9459–9474.
5. Karpukhin, V., Oğuz, B., Min, S., et al. (2020). Dense passage retrieval for open-domain question
answering. Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing,
6769–6781.
6. Reimers, N., & Gurevych, I. (2019). Sentence-BERT: Sentence embeddings using Siamese BERT-
networks. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing,
3982–3992.
7. Johnson, J., Douze, M., & Jégou, H. (2019). Billion-scale similarity search with GPUs. IEEE Transactions
on Big Data, 7(3), 535–547.
8. Rudin, C. (2019). Stop explaining black box machine learning models for high stakes decisions and use
interpretable models instead. Nature Machine Intelligence, 1(5), 206–215.
9. Miller, T. (2019). Explanation in artificial intelligence: Insights from the social sciences. Artificial
Intelligence, 267, 1–38.
10. Devlin, J., Chang, M.-W., Lee, K., & Toutanova, K. (2019). BERT: Pre-training of deep bidirectional
transformers for language understanding. Proceedings of NAACL-HLT 2019, 4171–4186.
11. Zheng, L., Guha, N., Anderson, B., et al. (2024). LegalBench-RAG: A benchmark for retrieval-augmented
generation in the legal domain.
https://arxiv.org/abs/2408.10343
12. Guha, N., Nyarko, J., Ho, D. E., et al. (2023). LegalBench: A collaboratively built benchmark for
measuring legal reasoning in large language models. Advances in Neural Information Processing Systems.
https://arxiv.org/abs/2308.11462
Page 1582
www.rsisinternational.org
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XV, Issue VI, June 2026
13. Bhattacharya, P., Ghosh, K., Pal, A., Mehta, P., Bhattacharya, A., & Majumder, P. (2019). A comparative
study of summarization algorithms applied to legal case judgments. Lecture Notes in Computer Science,
11882, 413–428.
14. Zhong, H., Guo, Z., Tu, C., Xiao, C., Liu, Z., & Sun, M. (2020). How does NLP benefit legal system: A
summary of legal artificial intelligence. Proceedings of the 58th Annual Meeting of the Association for
Computational Linguistics, 5218–5230.
15. Bommasani, R., Hudson, D. A., Adeli, E., et al. (2021). On the opportunities and risks of foundation
models. Stanford Center for Research on Foundation Models.
https://arxiv.org/abs/2108.07258
16. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, Ł., & Polosukhin, I.
(2017). Attention is all you need. Advances in Neural Information Processing Systems, 30, 5998–6008.