INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,  
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)  
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XIV, Issue XI, November 2025  
eSummarizer AI Service- Document Summarization Model Using  
BART  
Mriganka Mohan Bora#, Rongdeep Pathak*, Nelson R Varte#  
#Department of Computer Applications, Assam Engineering College, India  
*Department of Computer Applications, Jorhat Engineering College, India  
DOI : https://doi.org/10.51583/IJLTEMAS.2025.1411000093  
Received: 06 December 2025; Accepted: 12 December 2025; Published: 19 December 2025  
ABSTRACT  
The exponential growth of governmental documentation in India has created an urgent need for efficient  
automated summarization systems. This research presents eSummarizer AI Service, a custom Transformer-based  
machine learning model designed specifically for summarizing Indian government documents such as policy  
papers, circulars, legislative texts, and departmental reports. The study employs advanced Natural Language  
Processing (NLP) techniques, including BART (Bidirectional and Auto-Regressive Transformer), reinforcement  
learning, and a custom dataset curated from official government portals. Evaluation using ROUGE metrics  
demonstrates significant improvements over existing baseline models, achieving high coherence, contextual  
relevance, and factual consistency. The proposed system has practical applications for policymakers, researchers,  
and citizens, enhancing the accessibility and comprehension of complex governmental information.  
Keywords: NLP, ROUGE, BART, Streamlit  
INTRODUCTION  
The Government of India regularly disseminates a large volume of documents—policy frameworks, legislative  
bills, departmental circulars, and technical reports. While these documents are critical for transparency and  
informed decision-making, their increasing volume and complexity hinder efficient public access. Manual  
summarization is challenging, subjective, and time-consuming, creating a need for an automated summarization  
model  
capable  
of  
producing  
concise,  
accurate,  
and  
context-aware  
summaries.  
Given the diversity and structural heterogeneity of Indian government documents, a domain-specific  
summarization model becomes essential.  
Objectives  
The research aims to develop a custom summarization model that can:  
Handle Diverse Document Types (policies, bills, reports, circulars).  
Maintain Contextual Relevance to preserve critical information.  
Improve Accessibility by producing short, understandable summaries.  
Ensure Accuracy by minimizing redundancy and factual inconsistency.  
LITERATURE REVIEW  
The text summarization process using NLP was first introduced in 1958. Text summarization was performed by  
computing the value for each statement in a given input. Traditionally, the analytical approaches were used to  
compute a value of each statement, and then the sentences with the highest values opted. To compute this value  
www.ijltemas.in  
Page 965  
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,  
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)  
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XIV, Issue XI, November 2025  
other approaches were used such as TFIDF [2], Bayesian models [3], etc. While computing the summary, crucial  
phrase extraction was used, which could trim the original text. These limitations led to the use of ML strategies  
for summarization like Bayesian learning model, as explained in the article [4].  
The rapid expansion of governmental digital archives in India has increased the demand for automated systems  
capable of summarizing large and complex documents. According to Vaswani et al. [8], Transformer-based  
architectures have revolutionized natural language processing due to their ability to model long-range  
dependencies effectively. Additionally, Lewis et al. [17] demonstrated that BART, a denoising sequence-to-  
sequence Transformer, achieves state-of-the-art results on several summarization benchmarks, making it suitable  
for tasks involving long, structured government texts. Lin [14], established ROUGE as a reliable evaluation  
metric for summarization. Given the lack of domain-specific summarization models for Indian government  
documentation, this study builds upon the foundations laid by these researchers to create a specialized, high-  
accuracy summarization system tailored to public sector documents.  
Recent advancements in NLP highlight the effectiveness of Transformer-based models for tasks requiring  
contextual understanding. Studies indicate that models such as BERT, GPT, T5, and BART outperform  
traditional RNN-based summarizers. BART, in particular, has demonstrated excellence in abstractive  
summarization due to its bidirectional encoder and autoregressive decoder. Evaluation metrics like ROUGE  
remain standard for assessing summarization performance across research. However, limited work has been  
undertaken in developing summarization systems tailored specifically to Indian government documents, which  
involve specialized vocabulary, legal terminology, and structured policy language.  
METHODOLOGY  
System Workflow  
The methodology of the eSummarizer AI Service involves a structured sequence of processes that include dataset  
development, text preprocessing, model design, training, evaluation, and deployment. The overall workflow is  
illustrated in Figure 1.  
Figure 1  
www.ijltemas.in  
Page 966  
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,  
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)  
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XIV, Issue XI, November 2025  
Dataset Collection  
The dataset for this model was meticulously collected from various Indian government websites to ensure a  
comprehensive and representative sample of government documents. The documents include policy papers,  
legislative texts, reports, and circulars. The goal was to gather a diverse range of document types to develop a  
robust summarization model capable of handling different formats and content styles.  
The Sources are:  
Government Portals: Key government portals such as india.gov.in, mygov.in, and various ministry websites were  
primary sources. These portals host a wide range of official documents and publications.  
Ministry and Department Websites: Specific websites of ministries (e.g., Ministry of Finance, 0Ministry of  
Health and Family Welfare) and departments were targeted to collect documents related to their respective  
domains.  
Legislative Bodies: Websites of legislative bodies like the Lok Sabha (parliamentofindia.nic.in) and Rajya Sabha  
were used to gather legislative texts and parliamentary reports.  
Public Sector Undertakings (PSUs): Websites of PSUs and government agencies provide additional sources of  
annual reports, project documents, and circulars.  
Preprocessing  
Text Extraction: Extracting text from PDFs and DOC files involved handling various challenges such as  
embedded images, tables, and inconsistent formatting. Scripts were developed to clean and normalize the  
extracted text, ensuring it was ready for further processing.  
Metadata Extraction: Metadata such as document title, publication date, source URL, and document type were  
extracted and stored in a structured format. This metadata is crucial for organizing the dataset and facilitating  
subsequent analysis.  
Data Anonymization: Sensitive information, if any, was anonymized to ensure compliance with privacy and  
ethical guidelines  
Model Architecture  
The core model uses BART, a sequence-to-sequence Transformer, fine-tuned on a curated dataset. Figure 2  
depicts the BART-based summarization model used in this project.  
Figure 2  
www.ijltemas.in  
Page 967  
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,  
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)  
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XIV, Issue XI, November 2025  
The model was designed using a custom dataset collected from various Indian government websites, ensuring it  
was tailored to handle the unique characteristics of these documents. This dataset included diverse types of texts  
such as policy papers, legislative texts, reports, and circulars, providing a comprehensive training ground for the  
model. Leveraging advanced natural language processing techniques, we employed a Transformer-based  
architecture, specifically fine-tuning models like BART capture the intricate semantic nuances of government  
documents. The training process involved preprocessing steps to clean and normalize the data, and reinforcement  
learning was incorporated to optimize the summarization quality by defining reward functions that prioritize  
coherence, relevance, and factual accuracy. This customized approach allowed the model to adeptly summarize  
complex and diverse government documents, maintaining contextual integrity and providing concise, accurate  
summaries.  
Figure 3  
The figure above is the custom dataset prepared to train the custom model where documents were split in the  
80:10:10 ratio. 80% of the data being for the training part, 10% being for the testing part and 10% being for  
validation.  
Training and Testing  
The training of the summarization model was conducted over 4 epochs with a batch size of 8, striking a balance  
between computational efficiency and model performance. During each epoch, the model iteratively learned  
from the custom dataset, processing small batches of documents at a time. This batch size was chosen to ensure  
that the model could handle the complexity and variability of government documents while maintaining  
manageable memory usage. The relatively low number of epochs was sufficient due to the high-quality, domain-  
specific data and the advanced pre-trained Transformer architecture used as a base. Throughout the training  
process, we monitored key metrics such as loss and validation accuracy, making necessary adjustments to  
prevent overfitting and to fine-tune the learning rate. By the end of the training period, the model had effectively  
learned to generate coherent and concise summaries, demonstrating significant improvements in capturing the  
essential information from diverse government texts.  
The model was tested on a validation set extracted from the custom dataset of Indian government documents,  
ensuring it was evaluated on data representative of its intended application. The results showed that our model  
achieved high ROUGE scores, indicating that the summaries were both comprehensive and relevant, effectively  
capturing the key points of the original documents. These scores validated the model's ability to produce accurate  
and concise summaries, demonstrating its potential for practical use in making government documents more  
accessible and understandable.  
www.ijltemas.in  
Page 968  
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,  
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)  
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XIV, Issue XI, November 2025  
Evaluation  
The model was evaluated using the ROUGE (Recall-Oriented Understudy for Gisting Evaluation) metrics, which  
are standard for assessing the quality of text summarization. The specific metrics used were:  
ROUGE-1: Measures the overlap of unigrams (single words) between the generated summary and the reference  
summary.  
ROUGE-2: Measures the overlap of bigrams (two consecutive words) between the generated summary and the  
reference summary.  
ROUGE-L: Measures the longest common subsequence between the generated summary and the reference  
summary, focusing on fluency and coherence.  
The model was tested on a validation set extracted from the custom dataset of Indian government documents.  
The results indicated strong performance across the ROUGE metrics:  
ROUGE-1: 72.5%  
ROUGE-2: 56.8%  
ROUGE-L: 69.3%  
These scores demonstrate that the model's summaries had a high degree of overlap with the reference summaries,  
indicating that the model effectively captured the essential information from the original documents.  
Deployment  
A simple, interactive interface was developed using Streamlit:  
Allows users to upload government documents  
Generates summaries instantly  
Provides adjustable parameters  
Supports multiple languages  
Offers clean visual analytics  
Streamlit's lightweight architecture ensures quick deployment and easy usability.  
CONCLUSION  
This research demonstrates the development of a domain-specific summarization model for Indian government  
documents. Using BART and custom preprocessing, the system achieves high ROUGE scores, generates fluent  
summaries, and resolves the problem of information overload. While the model performs effectively,  
enhancements are needed for handling highly technical legal texts, improving redundancy control, and  
integrating domain-specific vocabularies.  
REFERENCES  
1. Babar S, Tech-Cse M, Rit, ”Text Summarization: An Overview”.2013  
2. Christian H, Agus M, Suhartono D,”Single Document Automatic Text Summarization using Term  
Frequency-Inverse Document Frequency (TF-IDF)”, ComTech: Computer, Mathematics and  
EngineeringApplications 2016.  
3. Nomoto T ”Bayesian Learning in Text Summarization Models” ,2005  
4. Graves A ,”Generating Sequences With Recurrent Neural Net works”, CoRR abs/1308.0850:2013  
www.ijltemas.in  
Page 969  
INTERNATIONAL JOURNAL OF LATEST TECHNOLOGY IN ENGINEERING,  
MANAGEMENT & APPLIED SCIENCE (IJLTEMAS)  
ISSN 2278-2540 | DOI: 10.51583/IJLTEMAS | Volume XIV, Issue XI, November 2025  
5. Nallapati R, Xiang B, Zhou B , ”Sequence-to-Sequence RNNs for Text Summarization”, CoRR  
abs/1602.06023: ,2016  
6. Hochreiter S, Schmidhuber J , ”Long Short-Term Memory”. Neural Comput 9:1735–1780,  
https://doi.org/10.1162/neco.1997.9.8.1735, 1997  
7. Shi T, Keneshloo Y, Ramakrishnan N, Reddy CK , ”Neural Abstractive Text Summarization with  
Sequence-to-Sequence Models”, CoRR abs/1812.02303:, 2018  
8. Vaswani A, Shazeer N, Parmar N, Uszkoreit J, Jones L, Gomez AN, Kaiser L, Polosukhin I , ”Attention  
is All you Need”. ArXiv abs/1706.03762:,2017  
9. Devlin J, Chang M-W, Lee K, Toutanova K BERT: ”Pre-training of Deep Bidirectional Transformers  
for Language Understanding”, 2019,CoRR abs/1810.04805  
10. Zhang J, Zhao Y, Saleh M, Liu PJ , ”PEGASUS: Pre-training with Extracted Gap-sentences for  
Abstractive Summarization”, 2019, CoRR abs/1912.08777  
11. Dong L, Yang N, Wang W, Wei F, Liu X, Wang Y, Gao J, Zhou M, Hon H-W, ”Unified Language  
ModelPre-training for Natural Language Understanding and Generation”, 2019 CoRRabs/1905.03197  
12. Radford A , ”Improving Language Understanding by Generative Pre-Training”,2018.  
13. Wolf T, Debut L, Sanh V, Chaumond J, Delangue C, Moi A, Cistac P, Rault T, Louf R, Funtowicz M,  
”HuggingFace’s Transformers: State-of-the-art Natural Language Processing”, 2019 CoRR  
abs/1910.03771.  
14. Lin, C. Y. .” ROUGE: A Package for Automatic Evaluation of Summaries.” pages 74–81, 2004  
Association for Computational Linguistics.  
15. Balaji N, Karthik Pai B H, Bhaskar Bhat B, Praveen Barmavatu, ”Data Visualization in Splunk and  
Tableau: A Case Study Demonstration”, in Journal of Physics: Conference Series, 2021.  
16. Lhoest Q, del Moral AV, Jernite Y, Thakur A, von Platen P, Patil S, Chaumond J, Drame M, Plu J,  
Tunstall L, Davison J. “Datasets: A community library for natural language” arXiv arXiv:2109.02846.  
2021  
17. M.Lewis, Y Lin, N Goyal . “BART: Denoising Sequence-to-Sequence Pretraining for Natural Language  
Generation.” ACL, 2020.  
www.ijltemas.in  
Page 970