DESIGN AND IMPLEMENTATION OF A LEGAL

RESEARCH ASSISTANT USING NLP,

SEMANTIC SEARCH, AND KNOWLEDGE GRAPHS
Sahishnu M MOHITH M CHAKRADHAR K MS. TRUPTHI RAO

AI&ML DEPARTMENT AI&ML DEPARTMENT AI&ML DEPARTMENT ASSISTANT PROFESSOR

DAYANANDA SAGAR UNIVERSITY DAYANANDA SAGAR UNIVERSITY DAYANANDA SAGAR UNIVERSITY AI&ML DEPARTMENT

ENG22AM0127@[Link] ENG22AM0115@[Link] ENG22AM0106@[Link] DAYANANDA SAGAR UNIVERSITY

TRUPTHIRAOAIML@[Link]

NACHIKET R

AI&ML DEPARTMENT

DAYANANDA SAGAR UNIVERSITY

[Link]

This paper presents ML tasks, deep

learning models and transformer-based
methods used in legal text processing. It
Abstract introduces feature extraction approaches
The rapid Growth of Electronic legal (e.g.,TF-IDF-based document
corpora and case law document repositories representations, word embedding and
has created a huge challenge for legal contextual language models) and similarity
system in trying to crawl through the huge measuring methods (e.g.,cosine similarity).
amount of legal text to find the precedent It also provides legal datasets, evaluation
and prediction of case result in a timely measures and experimental settings in
fashion. Nevertheless with the power of related works.
recent advances in Artificial Intelligence
and Natural Language Processing there is a
powerful possibility of automating legal Based on the survey, the paper identifying
document understanding (both on a per- several important research gaps including
document basis and on document matching) challenges for long legal document
and case judgment prediction. This survey modeling, domain-specific language
paper is an attempt to give an overview of diversity problem, interpretability
existing work on applying AI methods to challenge, data sparseness issue etc.
legal case analysis with particular focus on Following the research gaps, the paper
legal judgment prediction legal semantic proposing the conceptual framework of
similarity measurement and intelligent Law Match AI enabled legal analytics
legal retrieval systems. systems for legal research and decision
making combining the semantic legal case
similarity and verdict prediction. The
survey is ended with open challenges and
potential future research directions for AI Ambiguities or the subtlety of legal
enabled legal analytics systems. language and reasoning, background
context of fact-providing, use of jargon
1. Introduction specific to jurisdiction all these represent
additional problems. Keyword based
retrieval systems however have been
dominant for many decades, but these often
Within the legal domain, there is a vast level fails to recognize other more complex
of textual data (laws, legal arguments, old relationships among cases. Prediction
cases and old judgment). With the digital models may also succumb to information
availability of such data, legal professionals loss weak interpretations and poor
are now able to have access to a vast generalization within a legal theme.
number of judgments and case data. But the
preservation and accessibility of such large-
scale data brings forth new challenges in
To overcome these problems researches
legal text retrieval. Legal searches, case
build and use various types of classical
comparison and precedence searches still
classifiers and modern deep learning neural
remain highly time-consuming tasks, since
networks from different angles such as
domain-specific vocabularies and jargon
observed features based classifier, neural
needs to be processed within the open-text
network model with attention mechanism,
of the data. The process of identifying citing
contextualized embedding layer,
cases and the possible result of a new case
hierarchical the architectures and etc. So
(judgment) takes considerable amounts of
far, though modern deep learning models
human effort.
have certain comprehension for law and
prediction accuracy of laws judgment,
many research challenges still need to be
Recent developments in AIML and NLP overcome, such as use of long text, transfer
has provided encouraging insights into the learning model interpretation and over big
automation of complex tasks in text demand for training capacity.
processing. Document classification,
learning language semantics model,
learning language representations, among
This survey paper sufficiently constitutes a
others can be used and so far on various
general survey of existing approaches on
application domains including medicine,
applying AI techniques in legal case
finance, education, etc. In the context of
analysis with three prominent themes
law such techniques has also given rise to
including judgment prediction, semantic
new research directions including though
similarity-based approach, and intelligent
not limited to Legal Judgment Prediction
legal information system. It examines the
(LJP), legal document retrieval, AI-
previous studies on the topic in terms of
supported legal research and analysis 1.
commonality of used datasets, evaluation
strategy, and tendencies of frameworks,
also identifies common research
Despite the advances in legal text weaknesses, from which it then introduces
processing, it faces the other NLP Law Match AI as a conceptual AI platform
applications with particular difficulties. implemented on legal case similarity
This is because legal documents are often judgment and outcome prediction and
long complex and specialist language.
supports the related research discussion and competitive predictive performance. This
future projects of AI-assisted legal suggests that LLMs are suitable for the
analytics. latter two more complex semantics and
contexts of the law particularly where
2. Literature Survey nuanced doctrinal reasoning can be
leveraged.
Certainly highlight was the emerging trend
of recent Legal AI advances which showed
a move from single task predictive models
Another promising research direction is
towards end-to-end multi-tasks intelligent
similar case matching (SCM), which is an
judicial systems. For example, Wei et al.
integral part of most legal decision systems.
proposed an end-to-end intelligent trial
To breaks the deadlock of the accuracy
system which successfully demonstrated
dropoff of modern SCM, Xiong and Qiu
various of AI tasks’ collaborative
proposed a deep text understanding model
performances in a intelligent court samples
that exploits the combined strength of the
including information extraction, evidence
accurate fine tuning BERT representations
grouping and classification, dialogue
and the general-purpose global- and self-
summarization, judicial outcome
attention. And then, their dual channels
prediction, and judicial outcome
matching structure can lead to better long-
generation. Although moving from bottom-
text representation and needn’ t as much
up predicting of various of single task
computation. And their results suggest that
toward end-to-end support system in
the general use of semantic relations to
judiciary, the endeavor of Wei et al. System
aggregate the broad context dependencies
was mainly confined to procedural support
would be a necessary condition of any
especially the search of automatic
robust precedent retrieval system. This
interrogation and judgment related issue-
work illustrate that SCM is not merely an
focused summarization. Incorporation of
task of information retrieval, but a semantic
law-specific knowledge embedded into
reasoning task beyond the fine judgment
deep neural networks based on First-order-
prediction systems.
logic enhances the explainability and
logical reasoning of the system.
On top of the text analysis, the researchers
have also recently contributed to the
Alongside these system-level techniques
increasing line of evidentiary challenges
the second major category of work to
(eg. Deepfake detection). Hydara et al.
emerge is transformer-based legal
Explored solutions to AI-based judicial
assistants. For example, al-Shareef
facial evidence verification by developing a
proposed CHRExpert, an AI-human rights
test detection architecture made up of (1)
litigation assistant, which was a fine-tuned
frame selection, (2) confidence thresholds,
generative transformer trained on data from
(3) prediction time-stamps, and (4) heat-
the European Court of Human Rights
map features. Empirical result emphasized
(ECHR). The research demonstrates the
the significance of forensic confidence and
ability of LLMs to perform provision
user trust measures in judicial AI systems
analysis, statutory interpretation, and
advances of Legal AI not solely for
analogical reasoning while reducing legal
prediction, but for evidence validation and
preparation time by orders of magnitude
and without substantially sacrificing
the communication of legal sufficiency with multi-layered architecture for
standards. knowledge infusion, graph-based case-law
retrieval, multi-hop reasoning and case-
level judicial judgment synthesis. All
Legal J udgment Prediction (LJP) task still variables and data in legal model are
remain one of the highly pursued tasks and precisely encoded by with the interactions
hybrid architecture seems to be the way of legal rules, case law and fact. It provides
forward. Sukanya and Priyadarshini has the overcoming solution of many prediction
identified that an augmented hybrid deep models which solely operates on data. Their
learning architecture based on contextual study confirm the importance of
embeddings (ELMo) Dimensionality knowledge-driven reasoning framework.
reduction (IPCA), Bi-GRU cell networks,
modified attention based sequence
modeling was one of the best feature Going even further various other novel
combination sequence modeling approach neural approaches set the stage for future
for Indian criminal cases. They showed that exploration within LJP. To be specific, a
combining semantic based representation hybrid dilated CNN-LSTM structure
as features along with attention based utilizing autocorrelation analysis was
sequence modeling can provide high introduced by Abbasi et al., showing that
accuracy in prediction. This paper has also multi-resolution convolutional
reinforced the popular notion that LJP can representations can be an effective way to
be helped with hybrid feature encode underlying textual signals. This
representation and need not always study offers a new view of architecture-
necessaril be end to end transformer. conscious feature extraction driven by the
statistical properties of legal text.
Extraction of relevant factual information is
Similarly, Abdel Samee et al. Developed still a major sub-component of most Legal
JusticeAI, a collaborative multi-modal AI workflows and Legal NER in particular
system based on large language models is a necessary step. Liu et al. Presented a
(LLMs) and ensemble learning. Their novel state-of-the-art fewshot learning
system leveraged pre-trained embedding scheme that utilizes BERT word vessel,
combined with deep-learning methods to data augmentation, NNshot classifier, and
co- predict court decisions and maintain Viterbi decoding. This work effectively
interpretability by delivering a very tackles the looming problem of limited
transparent architecture. It exemplifies a labeled legal documents and reveals how
general trend in cross-domain and multi- resource-conserving-learning schemes can
modal judicial intelligence systems due to enable large-scale legal knowledge
judicial reasoning often being characterized management.
by heterogeneous data and knowledge
sources.
Finally, Turan and Kucuksille talked about
the use cases for hybrid models of NLP and
Conclusion. Organizedly reasoning and explainable AI approaches applied to legal
knowledge fusion has been established as document analysis. Their work
the mainsteam topic to a significant extend. incorporated extractive-abstractive
Wang and others suggest LegalReasoner summarization, classification models, and
interpretable methods based on SHAP Key
which evidenced the influence of Approac Key
Study Limitatio
summarization quality on classification h Strength
n
predictions accuracy. The manuscript
reveals how important the explainability in Hydara et AI-based
Enhances Limited
legal decisions support systems is and al. – forensic
evidence to media
emphasizes the transparency in resolution Deepfake verificati
reliability forensics
of cases. Detection on

Sukanya
Hybrid
& High
DL for Model
Priyadars predictio
judgment complexit
hini – n
predictio y
HDLMS accuracy
n
F

Multi-
stage
Wang et
LLM + Explicit Error
al. –
knowledg reasoning propagati
LegalRea
e pipeline on risk
soner
integratio
n

Key Liu et al. Works

Approac Key
Study Limitatio – Few- Few-shot with Stability
h Strength
n Shot learning limited challenge
Legal + BERT annotatio s
Multi- End-to- NER ns
High
task end
Wei et al. system
intelligen judicial Turan &
– FITS complexit
t trial assistanc Küçüksill Summari Summari
y Improved
system e e – zation + zation
interpreta
Hybrid predictio sensitivit
Transfor bility
Al- Strong NLP + n + SHAP y
mer- Domain XAI
Shareef – contextua
based dependen
CHRExp l
legal cy
ert reasoning
assistant

BERT +
Xiong & Improved Data-
attention
Qiu – long-text intensive
for case
SCM similarity training
matching
 discover higher order similarities between
legal data and thus miss relevant cases
because the user ‘terms don‘t match the
judge‘s ‘terms.

Most popular legal research systems still

use Boolean search, rule-based indexing
over hand-designed documents, and user-
definable metadata. Though these tools are
powerful, they rely on a user having explicit
and domain and preferably law firm
specific vocabulary at hand to compose
good queries. Existing keyword-based
retrieval of course does not accommodate
paraphrased arguments, widely-accepted
contextual links or covert case analysis
leading practitioners to often sift through a
manual document review of massive result
sets.

3. Existing Systems To overcome their respective limitations,

the keyword-based systems then proposed a
diverse set of machine learning-based legal
analytics models again with focus on LJP.
The first stage uses conventional ML
models, e.g., Support Vector Machine
(SVM), Logistic regression, trained on a
corpus of stylized features. Textual features
are represented as vectorized
representations by applying a wide range of
feature extraction techniques, such as TF-
IDF, Bag-ofWords. Although these models
achieved some prediction power, they could
not capture long-distance dependencies and
The development of legal digital databases other subtleties of legal text.
has led to the development of dozens of
information retrieval tools for legal
research. Most of these are simply indexing Compared with the traditional language
and searching systems to find information modeling models, deep learning models
such as case law, statutes and judicial have been applied more heavily in recent
decisions that match the query with a years in particular supervised neural
similar lexicon to the query. While such network language models. Various neural
systems may not be perfect, their worst architectures (for example CNN, RNN,
characteristic is their dependence on LSTM Hierarchical Attention Network
matching by identifier, which cannot
(HAN)) have been built and tested to learn with the verdict prediction system. How to
the sequential relation and hierarchy resolve the law matching issue, namely the
information in legal texts. An attention approach to transform the plain law texts
mechanism focused on the key legal terms into semantic safe vectors on computer, is
can empower the neural models to better the closely focus of this system. The
understand the legal texts. Then the complete flow of this system consist of
transformer language models, BERT and some constituent parts including data
legal BERT have been proposed to catch the collection text preprocessing feature
learned context semantics. extraction similarity measure classification
and decision support.

Nevertheless, intelligent legal system

nowadays is still relatively poor. Several 4.1 Data Collection and Representation
systems only master the knowledge for the
The built system uses a freely available
distribution on verdict prediction without
legal data set, Indian legal case records with
case similar searching and evidence
outcome labels. This data set contains raw
returning. Other computer aided model is
text report of legal case with outcome labels
difficult to be understood by legal specialist
in a structured format. This data set is
so that, sacrifices the representative power
suitable for applications including
of the prediction. Due to the small size
supervised learning and semantic analysis,
unbalanced class distribution and
where patterns of language use and decision
jurisdiction bias of legal data set, legal
can be inferred from previous judgments,
model is of low prediction accuracy.
are extracted.
Computational consuming is another
concern for legal text mining, especially for
evidence extraction with extremely long
legal document. 4.2 Text Preprocessing
The textual data contained in legal
documents can be both lengthy and noisy as
In conclusion, the existing systems are well as really specific. Case documents raw
either only for retrieving similar legal texts are preprocessed by applying various
documents and predict a certain result or operation before they are fed to the learning
not. An integrated system capable of algorithm: such operations include text
matching both purposes is still not normalization punctuation and symbol
obviously demanded, which makes Law removal stopwords removal lemmatization
Match AI popular and interesting because it and tokenisation to break the documents
is designed to assist in legal research. into small chunks from which to extract
features.
4. Proposed
Methodology
4.3 Feature Extraction
The preliminary Law Match AI are
designed to relax the two critical problems If we want to use machine learning models,
existing in the prediction system of current we need a way of converting our cases
judgment prediction system by integrating descriptions into some numeric vector. In
the semantic similarity detection technique our framework, we have the possibility of
plugging different approaches to feature accuracy in the classify can be adopted. The
extraction, as: trainer trainer trains the classifier on data
set based on the trainer selects the
Method applied in making the object to be
parameters and the evaluators evaluates the
represented by statistical representation,
classifer.
e.g. Selection of significant and trivial
words determined by means of Term 4.6 Decision Support Output
Frequency–Inverse Document Frequency
Finally the result of the similarity analysis
(TF-IDF).
and prediction is combined to produce
Semantic Embeddings: Word and document comprehensible decision support. Instead
embeddings are dense vectors that can of just trying to predict on the data:
encode contextual and semantic relations.
Estimated verdict type
Contextual Language Models: Instead of
For either confidence or probability
using embedding by using models based on
estimates (if asked),
transformers is also possible to use in
capturing deeper language dependencies. Top semantically similar historical cases:
What we may not be able to do with these This enables to compare the retrieved cases
representations is to measure differences. with the prediction we intend to make.
But thanks to these representations, we can Besides the blended result increases the
do just that and we can train models to transparency.
predict accordingly.
4.7 Methodological Advantages
This proposed method makes some
4.4 Case Similarity Analysis enhancements over stand-alone retrieval or
prediction systems:
Legally relevant precedents is a further
Subsection of the Law Match AI. We Merge prediction with semantics
measure the degree of relevance of a query
case to a collection of cases by calculating Less dependence on searching key words
a cosine similarity, which measures the among matching variables.
difference of angles between the features. Enhances interpretability by allowing case
As a result, we can also compare a similar retrieval
meaning irrespective of the specific
keywords. Our system then shows K most The supports extensibility with
similar cases sorted by the score. sophisticated NLP representations is:
The general law match AI strategy broad
functions into improvement of legal
4.5 Verdict Prediction Model research, so as to afford a timely solution of
The supervised classifier is employed for research work. It automates the work of
reasoning out the potential case results. It case citation tracking and case retrieval,
learns a mapping between the extracted finds out the out the primary effect of a
features and the verdicts labels. Other case, and smartly predicts the verdict.
classifiers such as Logistic Regression, The proposed Law Match AI framework is
Support Vector Machines Neural Nets, etc. designed to address key limitations
which fits the data set and achieves observed in existing legal information and
judgment prediction systems by integrating linguistic form, context-dependent
semantic similarity analysis with verdict information, implicit reasoning may all be
prediction. The methodology focuses on beyond the scope of statistical or
transforming unstructured legal text into embedding representations. So the system
meaningful computational representations, may ignore a legal argument or an
enabling both accurate case comparison and interpretation of the context.
predictive modeling. The overall workflow
consists of multiple stages, including data
acquisition, text preprocessing, feature Our introduced similarity analysis system is
extraction, similarity computation, suited for similar semantic table lookup in
classification, and decision support text, but has constraints on its
generation. representation power. TF-IDF or simple
embeddings don’t encode certain important
legal semantics or document structure
deeply. More sophisticated transformers
could address the problem but at a large cost
of compute.
5. Limitations
Despite the potential based on its
Another limitation related to
capability, there are some limitations of
interpretability and trustworthiness.
AI-based legal analytics systems like Law
Decision making in law is based on
Match AI. These limitations have emerged
explainability but certain ML/DL models
because of certain data issues, assumptions
are considered as black-box systems.
in the method and nature of legal
Despite of some achieved high prediction
reasoning.
accuracy the absence of clear rationale
perhaps would reduce user beliefs in the
forecast when the result is applied in
Another major challenge is the availability
matters of the law.
and the quality of legal datasets. The
majority of publicly available legal datasets
are small in size haphazardly annotated and
Further more since the verdict prediction
present class imbalance problems. In
systems are machine learning algorithms
addition, legal judgment differs
they also rely heavily on the past patterns
significantly by jurisdiction court and class
of the data. The accumulated biases of the
of case type, consequently trained models
past decisions in the database might affect
could be less applicable to cases outside the
the results obtained by these systems. The
training data, its prediction accuracy may
bias and fairness issues should be
deteriorate when faced with the cases out of
consciously addressed in all stages of data
the training data.
collection, bias assessment and system
validation.
Another challenge comes from the nature
of legal language. Legal language is
Finally we should bear in mind that AI’s
complex and diverse. Legal texts are
are meant to be used as a tool of support
usually lengthy repetitive and use complex
rather than replacing legal interpretation
legal terms. Small variations in the
and the balancing of human decision- high value of recall identified the stronger
making. Final decisions should always be ability of model to identify the relevant
made through using human ethics and cases in each class to correctly predict the
reasoning which an AI cannot completely class of case; particularly the higher values
do. The predictions made by Law Match AI of recognition and precision in Accepted
should not solely be used as definitive. and Rejected classes demonstrated the
latent ability of model to identify the pattern
from cases and correctly depict the
In general, Law Match AI adds a further contextual similarity, while mentioned
advancement towards efficiency in the system weak in classifying the cases under
work of legal search and analysis. How ever “Other” class.
issues such as data quality, language and
combinatorial explosion, interpretability,
bias and domain generalization should still
be considered before deploying or applying
these model to the real world.

6. Result and Analysis

The results, discussion of proposed legal
case prediction and analysis system
revealed the robust ability of proposed
system to predict, categorize and analyze
legal cases using combination of NLP
technique, Machine learning techniques
and similarity based retrieval approaches
(using MATLAB 2). The system was
trained on legal case dataset with total
29,000 legal cases in case categorization
task segregated into Accepted, Rejected and
Other classes with 80:20 order between
training and testing data (with 80% of data
used for training and 20% for testing in
order to achieve the generalization and
avoid over-fitting of data). The
performance of model has been evaluated
with precision, recall and F1-Score (used
the widely accepted metrics to evaluate An important aspect of the system‘s success
classification performance of model) to is the fact that the system can effectively
gain the comprehensive understanding of capitalize on and take the advantage of
performance for classification task. The robust text processing.
high value of precision of model indicated
fewer false positive predictions and more
reliable prediction of case, similarly the
 7. Conclusion Building upon the research gaps identified
above, this paper further introduced the idea
The employment of Machine Learning, AI and the scope of Law Match AI, an expert
and NLP methods in legal domain has been decision assisting instrument trying to
steadily increasing with growth of digital integrate case sentence similarity analysis
legal records and requirements for with verdict prediction into a single
automated legal research. Traditional legal autonomous system. Conceptually this
information retrieval systems based on hybrid derivative of datum-oriented
keyword search and manual classification precedents retrieving and predictive models
lack granularity, richness and depth of legal aims at strengthening legislature and the
texts. Research indicates that machine accessibility of legal research. Instead of
learning and deep learning methods can confronting human experts, such systems
assist significantly in tasks like predictive should strive to.
modeling of legal decisions, document
classification and detection of semantic
similarity. In conclusion the AI Legal Analytics is one
of the fastest growing research arena which
could deem to be a revolution in doing law
The paper surveyed a number of machine research and decision support. Progress of
learning techniques used in different the language model explainability and
researches in the field of AI based legal constructing legal datasets is likely to
analytics, such as traditional machine overcome existing weaknesses of smart
learning algorithms, neural network legal systems and increase the credibility
models, attention mechanisms and and acceptance of the legal information
transformer models. The survey further system. Future direction shall be towards
presented the evolution from formal the strength, transparency, equity and
linguistic pattern representations into domain flexibility of the system.
contextualised embeddings, thereby
enabling more nuanced linguistic analysis
and discussed representative legal data sets,
evaluation metrics and experimental 8. References
designs used in the field.

[1] B. Wei, K. Kuang, C. Sun, J. Feng, Y.

Despite much progress, there are still other
Zhang, X. Zhu, J. Zhou, Y. Zhai, and F. Wu,
remaining challenges including difficulties
“A full-process intelligent trial system for
to process long legal documents, open data
smart court,” Frontiers of Information
sets, domain specific language,
Technology & Electronic Engineering, vol.
transparency and bias caused by the
23, no. 2, pp. 186–206, 2022, doi:
historical legal data. However, these
10.1631/FITEE.2100041.
remaining challenges indicate the necessity
of a better integrated, explainable and [2] Y. Al-Shareef, “CHRExpert: An AI-
domain adaptive AI solutions. Driven Court of Human Rights Expert
Assistant for Legal Practitioners Utilizing
Transformer Models,” IEEE Access, vol.
13, 2025, doi: a Hybrid Natural Language Processing
10.1109/ACCESS.2025.3547763. Method,” IEEE Access, vol. 13, 2025, doi:
10.1109/ACCESS.2025.3556725.
[3] J. Xiong and Y. Qiu, “Deep Text
Understanding Model for Similar Case [10] M. S. Abbasi, B. Munir, V. Jayaram,
Matching,” IEEE Access, vol. 12, 2024, and P. Rivas, “Leveraging Autocorrelation
doi: 10.1109/ACCESS.2024.3439775. in a Dilated CNN-LSTM Framework for
Predicting the US Supreme Court
[4] E. Hydara, M. Kikuchi, and T. Ozono,
Decisions,” IEEE Access, vol. 13, 2025,
“Empirical Assessment of Deepfake
doi: 10.1109/ACCESS.2025.3609282.
Detection: Advancing Judicial Evidence
Verification Through Artificial
Intelligence,” IEEE Access, vol. 12, 2024,
doi: 10.1109/ACCESS.2024.3480320.
[5] G. Sukanya and J. Priyadarshini,
“Enhanced Hybrid Deep Learning Model
With Improved Self-Attention Mechanism
for Legal Judgment Prediction,” IEEE
Access, vol. 13, 2025, doi:
10.1109/ACCESS.2025.3596180.
[6] N. Abdel Samee, M. Alabdulhafith, S.
M. A. H. Hassan Shah, and A. Rizwan,
“JusticeAI: A Large Language Models
Inspired Collaborative and Cross-Domain
Multimodal System for Automatic Judicial
Rulings in Smart Courts,” IEEE Access,
vol. 12, 2024, doi:
10.1109/ACCESS.2024.3491775.
[7] 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, 2024, doi:
10.1109/ACCESS.2024.3496666.
[8] Q. Liu, J. Wang, J. Fu, J. Liu, and B.
Chen, “Named Entity Recognition Method
of Legal Instruments Based on Improved
Few-Shot Learning,” IEEE Access, vol. 12,
2024, doi:
10.1109/ACCESS.2024.3484765.
[9] T. Turan and E. U. Küçüksille,
“Summarization, Prediction, and Analysis
of Turkish Constitutional Court Decisions
With Explainable Artificial Intelligence and