INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN SOCIAL SCIENCE (IJRISS)  
ISSN No. 2454-6186 | DOI: 10.47772/IJRISS | Volume IX Issue XI November 2025  
Legal Text Summarization using Bart and Explainable AI Techniques  
Teh Xiao Thong1, Halizah Basiron1*,Abdul Syukor Mohamad Jaya1, Fitrah Rumaisa2  
1Fakulti Kecerdasan Buatan dan Keselamatan Siber, Universiti Teknikal Malaysia Melaka, Hang Tuah  
Jaya, Durian Tunggal, 76100, Melaka, Malaysia  
2Universitas Widyatama, Jl. Cikutra 204 A, Bandung, 40125, Indonesia  
*Corresponding Author  
DOI: https://dx.doi.org/10.47772/IJRISS.2025.91100607  
Received: 11 December 2025; Accepted: 18 December 2025; Published: 26 December 2025  
ABSTRACT  
Summarizing lengthy documents, especially in the legal domain, posed significant challenges for humans and  
automated systems. Human efforts entailed considerable time and effort, while automated systems sometimes  
faltered in decision-making, leading to ambiguity in the generated summaries. This research explored the use  
of text summarization in legal documentation coupled with an explainability feature. It addressed the  
challenges of condensing lengthy legal texts and improving automated summarization systems' transparency.  
The research involved gathering legal documents, developing a Bidirectional and Auto-Regressive  
Transformers (BART) summarization model, and integrating explainability within the system, visualizing the  
attention mechanism. The system performance, which included BERT Score, cosine similarity, and Recall-  
Oriented Understudy for Gisting Evaluation (ROUGE) score between human-generated and system-generated  
summaries, and evaluation by target users, led to several engaging insights on legal summarization. The model  
demonstrated moderate performance, where user feedback indicated satisfaction with its functionality but  
highlighted the need for user interface improvements. Future improvements were suggested, including refining  
model training, enhancing the user interface, and adding features like adjustable summary lengths and  
language translation.  
Keywords: Text summarization, Natural language processing, Explainable artificial intelligence, Legal Field,  
Bidirectional and auto-regressive transformers  
1. INTRODUCTION  
Reviewing legal documents such as supreme court case documents often require specialized knowledge, and  
reading through the entire document to capture the critical information is time-consuming. As the volume of  
legal documents increases, extracting essential details without delving into the whole content becomes crucial.  
Hence, summarization provides a solution by providing flexibility and convenience to readers. In addition,  
explainable artificial intelligence (XAI) can ensure the system produces concise summaries and provides  
transparent justifications for the decisions made, enhancing trust and comprehension for legal professionals.  
Legal professionals often struggle with obstacles when driving through documentation, which includes time  
and effort. Reading and comprehending pages of documents can be a cumbersome process, which might lead  
to potential oversights or missed critical details. Moreover, clear explanations are necessary and crucial behind  
automated summarization, where transparency and accountability are paramount. This research aims to  
develop a system that can summarize legal documentation with suitable explanations. The system must ensure  
the generated summaries accurately capture the critical points of the original document and provide a clear  
rationale for including specific information.  
The domain of this research falls within the intersection of several fields: the legal domain, natural language  
processing (NLP), machine learning, and explainable artificial intelligence (XAI). This research operates  
within the legal fields, explicitly dealing with legal case documents, and involves NLP techniques for  
Page 7793  
www.rsisinternational.org  
INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN SOCIAL SCIENCE (IJRISS)  
ISSN No. 2454-6186 | DOI: 10.47772/IJRISS | Volume IX Issue XI November 2025  
processing and analyzing textual data. Incorporating explainability features suggests using machine learning  
and XAI algorithms for summarisation and providing insights or explanations about the summarized content.  
This research also delves into interpretability and transparency aspects to ensure the generated summaries and  
explanations are understandable and trustworthy to stakeholders. The related work has been separated into two  
major categories: existing systems and techniques, and the kinds of literature are reviewed respectively.  
1.1.  
Existing Systems  
A hybrid method for automatic text summarization of legal cases using k-means clustering techniques and term  
frequency-inverse document frequency (TFIDF) word vectorizer is proposed by Varun Pandya [1]. The process  
involves data preprocessing to clean the document, clustering similar sentences using k-means, and extracting  
sentences to form a summary. The k-means algorithm groups sentences, which are then vectorized with TF-  
IDF. Clustering minimizes intra-cluster distances and maximizes inter-cluster distances, with optimal clusters  
determined. Sentences are ranked based on TF-IDF score and title similarity, with top-ranked sentences  
selected for the final summary. The dataset comprises Australian legal cases from Auslii. Evaluation is done by  
using Recall-Oriented Understudy for Gisting Evaluation (ROUGE) metrics to compare results with three  
automated tools. Pandya's method showed promising results, where the proposed method performs favourably  
well against other existing methods, as detailed in a comparative table.  
Anand and Wagh have also proposed simple generic techniques using neural network architecture, which are  
feed-forward neural networks (FFNN) based summary and long short-term memory (LSTM) based summary  
[2]. Their approaches require no manual features or domain knowledge and can be applied across various  
domains. The process involves generating labelled data using summary information from court judgment  
headnotes and utilizing this data to extract essential sentences for summarization. Different similarity  
techniques are employed to compute sentence labels, with sentence embeddings (SSE) performing best. FFNN  
transforms sentences into vectors, calculates probabilities, and selects the top-ranked sentences for the  
summary. LSTM, combined with convolutional neural networks (CNN), selects sentences with the highest  
importance likelihood based on LSTM output scores. Evaluation using ROUGE scores on Supreme Court of  
India judgment documents demonstrates the effectiveness of both methods. The result table shows that LSTM  
performs better in many cases.  
Research on the comparison of extractive and abstractive legal case document summarization was done by  
Shukla and his team [3]. This research aims to analyze the performance of various summarization methods on  
legal case judgment documents and explore effective evaluation techniques. Extensive experiments with  
several abstractive and extractive summarizations, including supervised and unsupervised approaches, have  
been carried out over three legal summarization datasets. Some examples of the methods are Luhn,  
Pacsum_bert, Maximal Marginal Relevance (MMR), Bidirectional and Auto-Regressive Transformers  
(BART), Bidirectional Encoder Representations from Transformers - Bidirectional and Auto-Regressive  
Transformers (BERT-BART), and Legal-Pegasus etc. The datasets, Indian-Abstractive, Indian Extractive, and  
UK-Abstractive datasets, are developed from case documents from the Indian and United Kingdom Supreme  
Courts. The analyses, including ROUGE, BERTScrore, and evaluations by legal practitioners, aim to provide  
insights into legal summarization and long document summarization in general, contributing to advancements  
in this field.  
Shifting the focus to another system, the Neural Networks for Text Summarization, with a Keras  
implementation of an attention-based sequence-to-sequence (seq2seq) model, is explored, emphasizing the  
success of the attention mechanism in the context [4]. Like other systems, data preprocessing is done as the  
first implementation step. A model with encoder-decoder architecture, which has global attention, is built, and  
an embedding layer to convert words into appropriate vector representations is used, learning along with the  
seq2seq model. Attention mechanisms in encoder-decoder neural networks enable the generation of a context  
vector at each timestep by considering the decoder's current hidden state and a subset of the encoder's hidden  
states. The dataset used in this study is the Amazon Fine Food dataset found on Kaggle. Since the original and  
generated summaries are short, the performance evaluation is done by comparing both.  
Page 7794  
www.rsisinternational.org  
INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN SOCIAL SCIENCE (IJRISS)  
ISSN No. 2454-6186 | DOI: 10.47772/IJRISS | Volume IX Issue XI November 2025  
An automatic abstractive text summarization model based on a hybrid attention mechanism has been  
introduced by Zhe Wang, where it incorporates a sentence-level attention mechanism to guide word-level  
attention distribution, adjusting the weight of sentence-level attention to mitigate high variance issues in word-  
level attention for shorter documents [5]. The methodology of this study introduces a hybrid-attentional model  
using encoder-decoder networks with recurrent neural networks (RNN). It incorporates attention mechanisms  
to improve decoder focus and a pointer-generator network for word generation or copying. Additionally, a  
dynamic hybrid attention mechanism adjusts attention values at both word and sentence levels to enhance  
summary quality based on document length. Evaluation of the approach using the ROUGE score on a large  
scale Chinese short text summarization (LCSTS) dataset demonstrates the effectiveness of the proposed  
method in capturing critical information and generating concise summaries. Other research on automatic  
abstractive text summarization using deep learning can also be refer to [6] and [7].  
1.2.  
Techniques  
Text summarization creates a short, accurate, and fluent summary of a longer text document [8]. This process  
is crucial for managing the vast volume of online text data, facilitating more efficient discovery and  
consumption of relevant information. There are two main forms of text summarization methods: abstractive  
and extractive. Extractive summarization combines existing sentences without any alterations to create a  
summary, while abstractive summarization involves text generation where the machine writes its own  
sentences [9],[10]. Extractive summarization is more rigid due to directly copying sentences from the source  
text, potentially resulting in awkward reading. Conversely, text generation in abstractive summarization  
initiates a better human writing style, enhancing coherence and readability with concise and coherent output.  
There are several prominent examples of both methods: Luhn, Latent Semantic Analysis (LSA), Text Rank,  
Lex Rank, Position Rank, and Topic Rank for extractive summarization. In contrast, abstractive summarization  
includes BART and pretraining with extracted gap sentences for abstractive summarization (PEGASUS) [11].  
BART, a denoising autoencoder for the pretraining sequence-to-sequence model, is introduced by Mike Lewis  
and his team [12]. BART is trained to reconstruct original text from corrupted versions using a Transformer  
based architecture, which can be seen as a generalization of models like BERT and generative pre-trained  
transformer (GPT). The study evaluates various text corruption methods and demonstrates BART's  
effectiveness in text generation, comprehension, abstractive dialogue, question-answering, summarization, and  
machine translation. Additionally, ablation experiments within the BART framework are conducted to assess  
factors influencing end-task performance. On the summarization task, BART shows an outperformance over  
two datasets (CNN/DailyMail and XSum) surpassing other existing methods. The resulting summaries are  
fluent and grammatically correct, indicating that BART's pretraining has effectively learnt a robust blend of  
natural language comprehension and generation.  
Erkan and Radev have presented a stochastic graph-based method for determining the relative importance of  
textual units, particularly in text summarization [13]. The technique is named Lex Rank. It computes sentence  
importance based on eigenvector centrality in a graph representation of sentences, using intra-sentence cosine  
similarity. In this study, Lex Rank is implemented into the MEAD summarization system [14]. The dataset  
used in the experiments consists of DUC 2003 and 2004 data sets, which involve generic summarization of  
news document clusters. For evaluation, the ROUGE metric, specifically ROUGE-1, which represents the  
unigram based ROUGE score, was used as it aligns closely with human judgements.  
A study by Kamya Singh and his team investigates using BERT based techniques for summarization and  
sentence similarity checks to enhance important question-answering systems [15]. The proposed approach  
combines BERT-based summarization with semantic similarity checking to extract key information and predict  
crucial questions. Experiments on benchmark datasets have been done, showing that this method surpasses  
traditional machine learning and deep learning techniques, achieving state-of-the-art performance. The  
approach was also effective in real-world applications like medical diagnosis, legal case analysis, and financial  
forecasting.  
There are several methods to evaluate the performance of a text summarization system, and one of the  
approaches is the ROUGE score. ROUGE stands for Recall-Oriented Understudy for Gisting Evaluation score  
Page 7795  
www.rsisinternational.org  
INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN SOCIAL SCIENCE (IJRISS)  
ISSN No. 2454-6186 | DOI: 10.47772/IJRISS | Volume IX Issue XI November 2025  
[16],[17]. It is a set of metrics commonly used for text summarization tasks to generate a concise summary of a  
longer text automatically. It was designed to evaluate the quality of machine-generated summaries by  
comparing them to human reference summaries. ROUGE has variants like ROUGE-N focusing on n-gram  
overlap, ROUGE-L on the longest common subsequence (LCS), and ROUGE-S on skip-bigram overlap. The  
ROUGE score ranges from 0 to 1, with higher values indicating better summary quality. It's widely used for  
objectivity but may not fully capture semantic meaning or coherence.  
Another method used to evaluate the quality of text summarization is BERTScore [18],[19]. This method  
measures the similarity between the summary and the original text. It addresses issues encountered by n-  
grambased metrics using contextualized token embeddings from models like BERT to compute similarity. The  
process involves representing sentences with contextual embeddings, measuring cosine similarity, token  
matching for precision and recall, considering word importance using IDF, and rescaling values for readability.  
For a basic level BERTScore calculation, the output will be precision, recall, and F1 score [20]. BERTScore  
enhances text similarity measurement, making it more accurate and balanced, with potential applications in  
various domains of natural language processing. However, this method has its pros and cons. For example,  
BERTScore can handle different types of texts, but it can be biased towards models that are more similar to its  
underlying model.  
In essence, ensuring the transparency and interpretability of the summaries is crucial, where explainable  
artificial intelligence (XAI) plays an important role. Some examples of XAI methods in NLP include  
visualizing attention mechanisms in neural networks, generating textual explanations for model predictions,  
and interpreting the reasoning behind the models' decision-making process [21]. Vaswani introduced the  
attention mechanism in 2017 [22]. In traditional Deep Learning models like LSTMs and RNNs, longer inputs  
pose challenges for retaining relevant information, prompting the need for attention mechanisms to signal the  
model about focus areas [23]. However, transformer models, utilizing self-attention across all encoder and  
decoder layers, circumvent this issue [23]. Attention mechanisms are widely used in text summarization across  
diverse domains like news, reviews, scientific papers, legal documents, and social media posts, where models  
such as the Pointer-generator network, Transformer, and BART exemplify this trend [24].  
Research on an open-source tool for visualizing attention mechanisms in transformer-based language models is  
proposed by Jesse Vig [25]. The tool offers three levels of granularity: attention head, model, and neuron  
views. Its application has been demonstrated on BERT and GPT-2 models. The tool aids in interpreting model  
decisions and identifying patterns, such as model bias detection, recurring pattern identification, and neurons to  
model behaviour linkage. This allows a comprehensive understanding of how the model attends to different  
input parts and how individual neurons contribute to attention computation. It enhances model interpretability,  
enables targeted improvements through user manipulation, and offers versatility for various analysis tasks and  
model types.  
On the other hand, a theoretical analysis of local interpretable model-agnostic explanations (LIME) has been  
done by Garreau and Luxburg [26]. This explainer is commonly used to provide interpretability to machine  
learning models. The study derives closed-form expressions for the coefficients of the interpretable model  
when the function to explain is linear, demonstrating that LIME can uncover meaningful features proportional  
to the function's gradient. It aids in understanding model decisions, improving trust, and facilitating  
compliance with regulations. However, it also highlights potential limitations of LIME where poor parameter  
choices may cause the algorithm to overlook important features.  
2. METHODS  
The system proposed in this research is the text summarization system with the explainability feature using  
NLP and machine learning techniques. Figure 1 shows the architecture of the proposed system. It involved  
backend and frontend parts connected by the Flask framework. In the frontend part, the user interacts with the  
system through a website. The user can upload a legal document to the system. Once the document is sent to  
the backend part, the system will undergo preprocessing, such as chunking and tokenizing. Then, the  
preprocessed document will be passed to the trained BART model to generate a summary. Additionally, the  
attention weight of the tokens in the document is visualized to produce a highlighted original document,  
Page 7796  
www.rsisinternational.org  
INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN SOCIAL SCIENCE (IJRISS)  
ISSN No. 2454-6186 | DOI: 10.47772/IJRISS | Volume IX Issue XI November 2025  
showing important sections or terms. The higher the weight, the deeper the highlight colour, the more  
important the word token is. Once the outputs are generated, both will be sent back to the frontend via Flask,  
and the user can see the result on the website. The user can also download the outputs to keep the results.  
Figure 1. The architecture of the proposed system  
2.1.  
Dataset  
The dataset used is Indian Supreme Court case documents and their abstractive summaries. The system's text  
summarisation model requires two datasets for training and testing purposes. There are a total of 7130  
documents in the original dataset. One hundred (100) documents from the dataset are randomly chosen as  
testing datasets; the remaining will be the training dataset. All the data is in .txt format. The dataset was  
downloaded from Zenodo, a research data repository [27].  
2.2.  
Training Process  
Figure 2 shows the training process of the proposed text summarization model. The necessary libraries, such as  
NumPy, NLTK, Pandas, TQDM and BART_utilities, are imported. BART_utilities is the file containing  
libraries that utilize PyTorch Lightning. The paths used in the following process are defined, including the  
dataset, root path, and output path. The training data, including judgment and summary, are read from specified  
directories and stored in lists. A total of 7030 data (a pair of judgments and summary considered as 1 data) is  
used in the training process.  
A pre-trained sentence transformer is initialized and loaded, specifying that it should run on CUDA (Compute  
Unified Device Architecture). This is because CUDA significantly boosts training and running a model. Next,  
the model will check both the judgement and summary lists. If there is a document in the lists, the model will  
split paragraphs into individual sentences and store them in lists (for example, l1 for judgment sentences and l2  
for summary sentences). The cosine similarity between two lists of sentences is calculated. Based on the result,  
chunks of text from judgment and their corresponding summaries will be generated. Once no more documents  
exist in the judgment and summary lists, the training chunks and summaries are generated and stored in an  
Excel file. The file is then read, and the columns are renamed into source and target data.  
The environment for using the BART model is set up to continue the training process, including loading a  
pretrained BART model and tokenizer. Special tokens are added to the tokenizer, and the token embeddings in  
the BART model are resized. The data module and lightning model with specified parameters (BART model is  
Page 7797  
www.rsisinternational.org  
INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN SOCIAL SCIENCE (IJRISS)  
ISSN No. 2454-6186 | DOI: 10.47772/IJRISS | Volume IX Issue XI November 2025  
used) are initialized. A PyTorch Lightning trainer is set up with GPU (Graphics Processing Unit) acceleration  
and other training parameters. After all settings are done, the training starts until it reaches the maximum  
number of epochs. The training process ends with saving the model weight into the checkpoint.  
Figure 2. The flowchart of the model training process  
2.3.  
Testing Process  
Figure 3 illustrates the testing process of the proposed text summarization model. Similar to the training  
process, it starts with importing libraries, setting up paths, and reading documents. The environment for using  
the BART model is set up by loading the pre-trained BART model and tokenizer. Special tokens are added to  
the tokenizer, and the token embeddings in the BART model are resized. The lightning model with specified  
parameters is initialized, and the saved model weight is loaded (trained BART model).  
The required summary length is set to 15% of the length of the original document. This is because the  
summary length should usually be 10% to 15% of the original text length or even shorter than the range [28].  
The model next retrieves the document's name and content. The word count of the document and the required  
summary length are calculated. The model splits the document into nested chunks of sentences with a  
maximum chunk length of 1024 words. The required summary length per chunk and percentage of document  
length for summary is calculated.  
The testing process continues by generating summaries for each chunk using the BART model on GPU. The  
generated summaries are concatenated into a single string. If the length is more than the required length,  
truncation needs to be done; otherwise, the final summary will be written in the specified output file, which is  
the end of the testing process.  
Page 7798  
www.rsisinternational.org  
INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN SOCIAL SCIENCE (IJRISS)  
ISSN No. 2454-6186 | DOI: 10.47772/IJRISS | Volume IX Issue XI November 2025  
Figure 3. The flowchart of the model testing process  
2.4.  
Explainability Feature  
Figure 4 displays the process of how the explainability feature is done. The process is similar to the testing  
process until the read document step. These steps can be skipped if continued after the summary generation.  
The attention weights are extracted from the model's output, where the last layer's attention is targeted.  
The original document is split into individual tokens to visualize the attention weights. The important tokens  
are highlighted based on the attention weights by setting highlight colours proportional to the attention  
weights. The higher the attention weights, the more important the token and the more obvious the highlight  
colour is. A PDF (Portable Document Format) file containing the original document text with highlighted  
tokens based on attention weights is created. The generated PDF file is saved to the specified output directory.  
Page 7799  
www.rsisinternational.org  
INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN SOCIAL SCIENCE (IJRISS)  
ISSN No. 2454-6186 | DOI: 10.47772/IJRISS | Volume IX Issue XI November 2025  
Figure 4. The flowchart of the explainability feature process  
3. RESULTS  
The methods used to evaluate the model performance in this research, which is the BART text summarization  
model, are BERTScore (including precision, recall, and F1- score), cosine similarity, ROUGE score, and user  
feedback. The first three evaluations are done by comparing the generated summaries and the reference  
summaries, while the last method is the feedback received from the target users.  
3.1.  
BertScore  
In the context of BERTScore, precision reflects how many of the tokens generated by the model are similar to  
the tokens in the reference summary, recall reflects how well the generated summary covers the tokens from  
the reference summary, and F1-score combines both to give an overall sense of how similar the generated text  
is to the reference text in terms of token similarity. High precision but low recall indicates that the model  
generates very accurate tokens, but it might miss out on important content. High recall but low precision  
indicates that the model includes the most relevant content and adds unnecessary or irrelevant tokens. A high  
F1 score suggests a good balance between capturing the important content and avoiding irrelevant content in  
the generated summaries. For this method, the BERTScore of every single testing data and the average are  
calculated. The overall precision, recall, and F1-score are 0.6241, 0.5976, and 0.6091, respectively. From the  
score calculated, the model can be proved to generate entirely accurate tokens without missing some important  
content. However, the performance can still be improved to get a higher score.  
3.2.  
Cosine similarity  
Cosine similarity is a metric used to measure how similar the documents are, irrespective of their size [29].  
This metric is advantageous as the two similar documents could still have smaller angles even if they are far  
apart by the Euclidean distance because of the size. The smaller the angle, the higher the similarity. The cosine  
similarity value is allocated between 0 and 1; the larger the value, the more similar the two documents are. The  
cosine similarity of the reference summaries and summaries generated by the proposed system and the overall  
accuracy based on cosine similarity is calculated. A threshold of 0.8 is defined to calculate the accuracy,  
meaning only summaries with a cosine similarity of more than 0.8 will be considered. From the given result,  
the accuracy is 0.94, showing 94 documents from 100 testing data with cosine similarity higher than 0.8.  
3.3.  
Rouge score  
ROUGE-1, ROUGE-2, and ROUGE-L are used to evaluate the proposed application. ROUGE-N measures the  
overlap of n-grams between the reference summaries and system-generated summaries, where ROUGE-1  
refers to the overlap of unigrams (every single word), and ROUGE-2 refers to the overlap of bigrams (two  
consecutive words). At the same time, ROUGE-L is based on the length of the longest common subsequence.  
The average ROUGE score has been calculated, which ROUGE-1 is 0.4911, ROUGE-2 is 0.2434, and  
ROUGE-L is 0.2449. The scores are considered moderate except ROUGE L, which is lower than 0.3 [30],  
[31]. However, it cannot be concluded that the overall performance of the application is poor because  
ROUGE score relies on reference summaries. The ROUGE score primarily focuses on the proportion of  
relevant information preserved in a summary, which may not always be the most crucial aspect in evaluating  
the system. Researchers may sometimes prioritize how accurately key details are captured, or fluency, which  
assesses the coherence and naturalness of the generated summary.  
3.4.  
Users' feedback  
A survey has been prepared for the users to rate the system's functionality. The target users are those who study  
or work in legal fields. Ten (10) target users are chosen and will continue with the questions about the system's  
functionality. A simple demographic is prepared in the survey to give an essential impression of the target  
users, especially in the years they are involved in the legal field, the frequency they deal with legal documents  
and their experience using text summarization tools. This can help answer the functionality questions better  
Page 7800  
www.rsisinternational.org  
INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN SOCIAL SCIENCE (IJRISS)  
ISSN No. 2454-6186 | DOI: 10.47772/IJRISS | Volume IX Issue XI November 2025  
because different experiences can have different opinions on the system. Figure 5 is the bar chart showing how  
accurately the users find the summaries generated by the system. The rating started from 1 (very inaccurate) to  
5 (very accurate). Most of the target users (6 respondents, 60%) find the summaries generated to be accurate,  
and one of them (10%) said they are very accurate. In comparison, three respondents (30%) found the  
summaries generated to be only moderate.  
Figure 6 displays the bar chart of target users' opinions on the importance of the highlighted words, rating from  
1 (very unimportant) to 5 (very important). Half of the target users (5 respondents, 50%) think the highlighted  
words are important, followed by three respondents (30%) who feel the importance of highlighted words is  
average. One respondent (10%) voted for each for the range of unimportant and very important. Figure 7  
shows the bar chart about the target users' understanding of the legal documents by only reading the generated  
summaries. The rating is between 1 (very poorly) to 5 (very well). Six respondents (60%) from the target users  
can understand well, and one respondent (10%) can understand very well. 30% (3 respondents) of the target  
users only moderately understand the legal documents if they only read the generated summaries.  
Open-ended questions are also prepared for the respondents for extra comments or feedback. Most respondents  
think the system is already good at present, but some also provide their opinions on future improvement. The  
most significant progress to be made in enhancing the user interface and content layout is increasing the line  
spacing, being more creative in colours and design, and ensuring the layout is more organized for easy  
viewing. Users also request extra functions, such as language translator features and an adjustable summary for  
users to select the summary length. Overall, the feedback highlights the users' satisfaction with the current  
system but also points out some valuable suggestions for enhancing the user interface (UI), layout, and  
functionality to improve the user experience (UX) further.  
Figure 5. The bar chart of how accurately the users find the summaries generated by the system  
Figure 6. The bar chart of the importance of highlighted words  
Page 7801  
www.rsisinternational.org  
INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN SOCIAL SCIENCE (IJRISS)  
ISSN No. 2454-6186 | DOI: 10.47772/IJRISS | Volume IX Issue XI November 2025  
Figure 7. The bar chart of how accurately the users find the summaries generated by the system  
4. CONCLUSION  
According to the results, the research successfully meets the set objectives by delivering a functional and  
explainable AI-driven text summarization system. The system demonstrates significant potential, especially  
with the integration of XAI. It also reveals some areas for improvement, such as performance optimization and  
user interface enhancement. The proposed improvements, including adjusting training parameters, upgrading  
the user interface, and adding new features like adjustable summary length and a language translator, will  
further enhance the system's functionality and user satisfaction. To conclude, the research contributes valuable  
insights and tools to the AI and legal field and creates a basis for further development. With the recommended  
improvements, the system has a high potential to become a leading tool in its domain, offering users a more  
powerful, customizable, and accessible text summarization experience.  
ACKNOWLEDGEMENTS  
The authors would like to thank the Applied Intelligent Computing (APIC) research group, the Center of  
Advanced Computing Technology (C-ACT), and Fakulti Kecerdasan Buatan dan Keselamatan Siber,  
Universiti Teknikal Malaysia Melaka (UTeM) for their incredible support in this research.  
REFERENCES  
1. V. Pandya, "Automatic Text Summarization of Legal Cases: A Hybrid Approach," in 5th International  
Conference on Advances in Computer Science and Information Technology, 2019, doi:  
10.5121/csit.2019.91004.  
2. D. Anand, and R. Wagh, "Effective deep learning approaches for summarization of legal texts,"  
Journal of King Saud University - Computer and Information Sciences, vol. 34, no. 5, pp. 2141–2150,  
May 2011, doi: 10.1016/j.jksuci.2019.11.015.  
3. A. Shukla, P. Bhattacharya, S. Poddar, R. Mukherjee, K. Ghosh, P. Goyal, and S. Ghosh, "Legal Case  
Document Summarization: Extractive and Abstractive Methods and their Evaluation." in Proceedings  
of the conference of the Asia-Pacific chapter of the Association for Computational Linguistics and the  
international  
https://arxiv.org/abs/2210.07544.  
4. Adarsh, "Text Summarization with Attention based Networks"  
joint  
conference  
on  
natural  
language  
processing,  
Oct  
2022,  
Available:  
https://medium.com/@iit2018056/text-summarizationwith-attention-based-networks-8492e9277c0  
(accessed Apr, 29, 2024).  
5. Z. Wang, "An Automatic Abstractive Text Summarization Model based on Hybrid Attention  
Mechanism," Journal of Physics: Conference Series, vol. 1848, no. 1, 012057, April 2021, doi:  
10.1088/1742-6596/1848/1/012057.  
6. N, Alipour and S. Aydin, “Abstractive summarization using multilingual text-to-text transfer  
transformer for the Turkish text”, International Journal of Artificial Intelligence, vol. 14, no. 2, April  
2025, pp. 1587-1596.  
Page 7802  
www.rsisinternational.org  
INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN SOCIAL SCIENCE (IJRISS)  
ISSN No. 2454-6186 | DOI: 10.47772/IJRISS | Volume IX Issue XI November 2025  
7. J. K. Adeniyi, S. A. Ajagbe, A. E. Adeniyi, H. O. Aworinde, P. B. Falola, and M. O. Adigun,  
“EASESUM: an online abstractive and extractive text summarizer using deep learning technique”,  
International Journal of Artificial Intelligence, vol. 13, no. 2, June 2024, pp. 1888-1899.  
8. S. Dutta, A. K. Das, S. Ghosh, and D. Samanta, "Graph-based clustering technique for microblog  
clustering," in Data Analytics for Social Microblogging Platforms. Academic Press, 2023, pp. 165-  
192. Accessed: Apr, 29, 2024. [Online]. Available: https://books.google.com.my/.  
9. N. A. Ranggianto, D. Purwitasari, C. Fatichah, R. W. Sholikah, Abstractive and Extractive  
Approaches for Summarizing Multi-document Travel Reviews. Jurnal RESTI (Rekayasa Sistem dan  
Teknologi Informasi). 2023 Dec 30;7(6), pp. 1464-75.  
10. A. M. Zakariae, B. Frikh, and B. Ouhbi. "EXABSUM: a new text summarization approach for  
generating extractive and abstractive summaries." Journal of Big Data 10, no. 1, 2023, pp. 163.  
11. N. Giarelis, C. Mastrokostas, and N. Karacapilidis, "Abstractive vs. Extractive Summarization:  
12. An Experimental Review," Applied Sciences, vol. 13, no. 13, pp. 7620, 2023, doi:  
10.3390/app13137620.  
13. M. Lewis, Y. Liu, N. Goyal, M. Ghazvininejad, A. Mohamed, O. Levy, V. Stoyanov, and L.  
Zettlemoyer, "BART: Denoising Sequence-to-Sequence Pretraining for Natural Language Generation,  
Translation, and Comprehension," in Proceedings of the 58th Annual Meeting of the Association for  
Computational Linguistics, 2020, pp. 7871-7880 Available: https://arxiv.org/abs/1910.13461.  
14. G. Erkan, and D. Radev, "LexRank: Graph-based Lexical Centrality as Salience in Text  
Summarization," Journal of Artificial Intelligence Research, vol. 22, pp. 457-479, 2004, doi:  
10.1613/jair.1523.  
15. D. R. Radev, and Z. Zhang, (2001, October 5). "Experiments in Single and MultiDocument  
Summarization Using MEAD. ResearchGate," in First document understanding conference, 2001, pp.  
1-7.  
16. K. Sharma, K. Singh, K. Sharma, and J. Gupta, "Question Summation and Sentence Similarity using  
BERT for Key Information Extraction." International Journal for Research in Applied Science and  
Engineering Technology, vol. 11, no. 4, pp. 1636–1639, 2023 doi: 10.22214/ijraset.2023.50087.  
17. P. Watanangura, S. Vanichrudee, O. Minteer, T. Sringamdee, N. Thanngam, and T.  
Siriborvornratanakul. "A comparative survey of text summarization techniques." SN Computer  
Science, vol. 5, no. 1, 2023, pp 47.  
18. R.D. Lins, H. Oliveira, and S.J. Simske, “Assessing the Reliability and Validity of the Measures for  
Automatic Text Summarization.” In Proceedings of the ACM Symposium on Document Engineering  
2024, pp. 1-4.  
19. B. Zhao, and Y.M. Lui, “Towards A Reliable Text Summarization Evaluation Metric Using Predictive  
Models”. International Journal of Pattern Recognition and Artificial Intelligence, vol. 36, no. 10, pp.  
2251011, 2022.  
20. A.M. Ibrahim, A. Marco and M. Aref. "A Systematic Review On Text Summarization Of Medical  
Research Articles." International Journal of Intelligent Computing and Information Sciences vol. 23,  
no. 2, pp. 50-61, 2023.  
21. H.  
Özbolat,  
"Text Summarization:  
How  
to  
Calculate  
BertScore,"  
https://haticeozbolat17.medium.com/textsummarization-how-to-calculate-bertscore-771a51022964  
(accessed Apr, 29, 2024).  
22. A.  
Mulkar,  
"Explainable  
AI  
(xAI)  
in  
Natural  
Language  
Processing  
(NLP),"  
https://ankushmulkar.medium.com/explainable-ai-xai-innatural-language-processing-nlp-  
d75d5be216e3 (accessed Apr, 29, 2024).  
23. A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin,  
"Attention Is All You Need," in Advances in Neural Information Processing Systems, vol. 30, 2017  
Available: https://arxiv.org/abs/1706.03762  
24. N. H. A. M. Norkute, "Explainable AI for Text Summarization of Legal Documents,"  
https://hyperight.com/explainable-ai-for-textsummarization-of-legal-documents/ (accessed Apr, 29,  
2024).  
25. "What are the pros and cons of using attention mechanisms in text summarization with RNNs?"  
www.linkedin.com.  
https://www.linkedin.com/advice/3/what-pros-cons-using-attentionmechanisms-  
textsummarization (accessed Apr, 28, 2024).  
26. J. Vig, "Visualizing Attention in Transformer-Based Language Representation Models," 2019,  
Available: https://arxiv.org/abs/1904.02679.  
Page 7803  
www.rsisinternational.org  
INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN SOCIAL SCIENCE (IJRISS)  
ISSN No. 2454-6186 | DOI: 10.47772/IJRISS | Volume IX Issue XI November 2025  
27. D. Garreau, and U. Luxburg, "Explaining the Explainer: A First Theoretical Analysis of LIME," in  
Proceedings of the 23rd International Conference on Artificial Intelligence and Statistics, 26–28  
August 2020, vol 108, pp. 1287–1296 https://proceedings.mlr.press/v108/garreau20a.html  
28. A. Shukla, P. Bhattacharya, S. Poddar, R. Mukherjee, K. Ghosh, P. Goyal, and S. Ghosh, "Legal Case  
Document Summarization: Extractive and Abstractive Methods and their Evaluation," in Proceedings  
of the conference of the Asia-Pacific chapter of the Association for Computational Linguistics and the  
international joint conference on natural language processing, 2022. doi: 10.5281/zenodo.7152317  
29. C. Burnell, J. Wood, M. Babin, S. Pesznecker, and N. Rosevear, "Writing Summaries," Pressbooks.  
https://openoregon.pressbooks.pub/wrd/chapter/writing-summaries/ (accessed May, 12, 2024).  
30. M. H. Asif, and A. U. Yaseen, “Comparative Evaluation of Text Similarity Matrices for Enhanced  
Abstractive Summarization on CNN/Dailymail Corpus.” Journal of Computing & Biomedical  
Informatics, vol. 6, no. 01, 2023, pp. 208-215.  
31. G. Sharma, and D. Sharma, “Automatic text summarization methods: A comprehensive review.” SN  
Computer Science, vol. 4, no. 1, 33, 2022.  
32. T. Chellatamilan, S. K. Narayanasamy, L. G. K. Srinivasan, and S. MN. Islam. "Ensemble Text  
Summarization Model for COVID‐19‐Associated Datasets." International Journal of Intelligent  
Systems 2023, no. 1, 2023, pp. 3106631.  
Page 7804  
www.rsisinternational.org