Integrating Extractive and Abstractive Summarization: A Hybrid

Approach

P. Yeshwanth Chowdary, K.Vishruth Solomon Kumar, B. Shashi Kiran and S. Aswani

Department of CSE (AI&ML), Institute of Aeronautical Engineering, Dundigal, Hyderabad, Telangana, India

Keywords: Text Summarization, Extractive Summarization, Abstractive Summarization, KL Divergence, BART, Key

Sentences, Hybrid Summarization.

Abstract: This project presents a comprehensive methodology for text summarization that integrates both extractive and

abstractive techniques to enhance the quality of generated summaries. In the extractive summarization phase,

KL divergence is utilized to identify and select key sentences or phrases from the source text, effectively

capturing the most relevant information. These selected segments are then passed as input to an abstractive

summarization model, specifically BART (Bidirectional and Auto-Regressive Transformers), which

processes and refines the extracted information to produce a coherent and fluent summary. By combining the

precision of extractive summarization with the fluency and coherence of abstractive approaches, the proposed

methodology aims to generate high-quality summaries that offer improved coverage of the source text,

enhanced fluency, and a significant reduction in redundancy.

1 INTRODUCTION

In today's data-driven world, the volume of text

generated across various industries, such as

journalism, research, and business, is overwhelming.

The challenge lies in efficiently extracting relevant

information from these large text corpora while

preserving the core meaning and essential details.

Text summarization is a vital tool for addressing this

issue by condensing extensive documents into shorter,

more manageable summaries. However, traditional

methods, such as frequency-based techniques, often

fail to capture the deeper semantic meaning and

context of the text, resulting in suboptimal summaries

that may lack coherence or relevance.

Our project aims to develop a more sophisticated

text summarization system that leverages advanced

techniques like KL-Divergence for extractive

summarization, identifying key sentences based on

their divergence from the overall text distribution.

Additionally, we use BART, a state-of-the-art

transformer model, for abstractive summarization,

which generates fluent, coherent summaries by

rephrasing or creating new sentences based on the

original content. This hybrid approach ensures that the

final summaries are both informative and concise,

making them useful across a wide range of industries

and domains that require efficient text processing and

analysis

2 LITERATURE SURVEY

Numerous studies have explored text summarization

across various domains, applying both traditional and

modern techniques. Machine learning offers various

potential methods, but its effectiveness hinges on

selecting an appropriate algorithm tailored to the

specific domain.

(B Rajesh, K Nimai Chaitanya, P Tejesh

Govardhan, K Krishna Mahesh, & B Sudarshan,

2024) Proposed a systematic approach to extractive

text summarization, focusing on text preprocessing,

sentence scoring, selection, and post-processing.

Utilizing Python libraries like SpaCy and NLTK, they

score sentences based on word frequencies and merge

similar sentences for coherent summaries. Challenges

include potential redundancy in extracted sentences

and limitations in handling diverse text formats.

(J.N. Madhuri & R. Ganesh Kumar, 2019) This

work presents a statistical method for single-

document extractive summarization by ranking

sentences based on assigned weights. Their approach

aims to condense text into concise summaries,

784

Chowdary, P. Y., Kumar, K. V. S., Kiran, B. S. and Aswani, S.

Integrating Extractive and Abstractive Summarization: A Hybrid Approach.

DOI: 10.5220/0013602400004664

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 3rd International Conference on Futuristic Technology (INCOFT 2025) - Volume 2, pages 784-788

ISBN: 978-989-758-763-4

evaluated through weights prioritizing important

sentences. They identify challenges such as context

loss when extracting sentences in isolation and

difficulties in accurate weight assignment.

(Asha Rani Mishra, V.K Panchal, & Pawan

Kumar, 2019) Address the challenge of extracting

insights from large textual datasets using topic

modelling and key phrase extraction. Their

multifaceted approach employs techniques like LSI

and TF-IDF to identify key topics and phrases, while

summarization methods like LSA are applied for

concise outputs. Challenges include combining

techniques effectively and handling diverse text

structures.

(Sanchit Agarwal, Nikhil Kumar Singh, &

Priyanka Meel, 2018) Introduced a method for

extractive summarization that combines K-Means

clustering with sentence embeddings. By clustering

sentences based on semantic similarity, they select the

most relevant sentences for summaries, evaluated

using ROUGE scores on the DUC 2001 dataset.

Challenges include sensitivity to cluster numbers and

the need for high-quality embeddings for accurate

results.

(Wen Xiao & Giuseppe Carenini, 2019) Presented

a novel extractive summarization model that combines

global and local context to identify key information

from long documents. Their methodology evaluates

content using both contexts, achieving superior

ROUGE-1 and ROUGE-2 scores on Pubmed and

arXiv datasets compared to traditional models.

However, the authors note challenges in effectively

integrating these contexts, the risk of redundancy in

summaries, and the need for robust evaluations across

diverse document types.

(L. Lebanoff, K. Song, & F. Liu, 2018) The

authors address the challenge of generating text

abstracts from multiple documents, utilizing a neural

encoder-decoder framework traditionally designed for

single-document summarization. Their approach

incorporates the Maximal Marginal Relevance

(MMR) method to select representative sentences

from various documents, subsequently fusing these

sentences into an abstractive summary. This method

does not require additional training data,

demonstrating its robustness in Multi document

contexts.

(Glorian Yapinus, Alva Erwin, Maulhikmah

Galinium, & Wahyu Muliady, 2014) This study

introduces a hybrid approach to multi-document

summarization specifically designed for Indonesian

documents. The authors aim to effectively condense

information from multiple sources into a coherent

summary by combining WordNet-based text

summarization (abstractive) with title word-based

summarization (extractive). This method is evaluated

against Latent Semantic Analysis (LSA), highlighting

its ability to generate well compressed and readable

summaries.

(A. Ghadimi & H. Beigy, 2022) This research

presents HMSumm, a hybrid approach to multi

document summarization which integrates extractive

and abstractive techniques by utilizing pre-trained

language models. The methodology involves

generating an extractive summary by selecting key

sentences from the documents while employing a

determinantal point process (DPP) to minimize

redundancy. Subsequently, the extractive summary is

passed to BART and T5 models for abstractive

summarization, with the final output chosen based on

sentence diversity. The study highlights the

effectiveness of combining multiple models to

enhance summarization quality.

(Christian, H. Agus, M.P., & Suhartono, 2016)

This study investigates the application of the TF-IDF

algorithm for single-document automatic text

summarization, aiming to enhance information

retrieval amidst the abundance of online content. The

methodology involves ranking sentences based on the

frequency of important terms while minimizing the

impact of common words. The performance of the

TF-IDF summarizer is assessed against other

summarization tools using the F-Measure for

comparison. While effective, the paper notes

limitations such as the algorithm's reliance on term

frequency, which may overlook critical contextual

information.

(G. Di Fabbrizio, A. Stent, & R. Gaizauskas,

2014) This study presents STARLET-H, a hybrid

summarization system designed for synthesizing

reviews of products and services. The methodology

involves a dual approach, utilizing extractive methods

to select key quotes from reviews, which are then

blended into an abstractive summary. However, the

paper highlights challenges such as potential

inconsistencies when merging extracted quotes with

generated text, which can disrupt the narrative flow

and affect overall coherence.”

3 DESIGN AND PRINCIPLE OF

MODEL

3.1 Methodology

In this study, we developed a hybrid document and

text summarization system that integrates both

Integrating Extractive and Abstractive Summarization: A Hybrid Approach

785

extractive and abstractive techniques using the BART

pre-trained model. The primary objective was to

generate comprehensive, informative, and coherent

summaries that effectively convey the main ideas of

the source documents.

3.1.1 Pre-trained Model Selection and Pre-

Processing:

For abstractive summarization, we employed the

BART model. The model's architecture allows it to

generate fluent and contextually rich summaries based

on the input text.

The preprocessing step cleans and normalizes the

input text to improve summarization. It removes

emojis, emails, URLs, phone numbers, and HTML

tags while normalizing hyphenated words, extra

spaces, Unicode characters, quotation marks, and

bullet points. These steps ensure the text is clean and

ready for summarization

3.1.2 Extractive Summarization

The extractive summarization process begins with

the implementation of the KL Divergence algorithm

to identify key sentences from the source text. The

algorithm calculates the divergence between the

probability distributions of words across the entire

text and the candidate summary sentences. Sentences

with the lowest KL Divergence scores are selected,

ensuring that the extractive summary retains the most

relevant and informative parts of the text.

The extractive summarization function processes

the input text by first splitting it into sentences. The

algorithm then computes the importance of each

sentence by analysing word frequencies, ultimately

selecting the top sentences that best represent the

original content.

Figure 1: Extractive Summarization Approach.

3.1.3 Abstractive Summarization

The key sentences identified in the extractive phase

serve as input for the BART model, which generates

an abstractive summary. This step enhances the

relevance and coherence of the final output by

allowing the model to focus on the most critical

information from the extracted sentences.

The abstractive summarization function tokenizes

the input text and generates a summary using the

BART model's capabilities. The model's ability to

understand and rephrase content ensures that the

summaries produced are succinct while retaining the

original meaning.

3.1.4 Hybrid Summarization

The hybrid summarization methodology leverages

the extracted key information to guide the generation

of the abstractive summary. By performing

summarization on the key sentences obtained from

the KL Divergence algorithm, the system combines

the strengths of both extractive and abstractive

approaches. This results in high-quality summaries

that are coherent and rich in content, effectively

addressing the limitations of purely extractive or

purely abstractive methods.

Overall, the proposed system exemplifies a robust

approach to document summarization by effectively

integrating a variety of advanced techniques. This

integration allows the system to produce summaries

that are not only informative and concise but also

contextually relevant to the source material.

4 RESULTS

The performance of the summarization model is

evaluated using the ROUGE metric suite. The

summarization model was evaluated using a dataset

of news articles to measure its performance. The

model's performance was assessed on a

representative subset of a news article dataset, which

provided an initial indication of the model's

capability. The results are summarized through the

average ROUGE scores as follows:

ROUGE-1 Score (Unigram Overlap): The

ROUGE-1 score evaluates the overlap of unigrams.

A recall of 0.6466 indicates that 64.66% of the

unigrams in the reference summary were captured in

the generated summary. The precision of 0.5636

suggests that 56.36% of the unigrams in the generated

summary are relevant to the reference summary. The

F1-score of 0.5783, which balances recall and

precision, highlights a reasonable level of

summarization accuracy.

ROUGE-2 Score (Bigram Overlap): The

ROUGE-2 score measures the overlap of bigrams. A

recall of 0.5359 indicates that 53.59% of the bigrams

INCOFT 2025 - International Conference on Futuristic Technology

786

in the reference summary were matched in the

generated summary. The precision of 0.4316 shows

that 43.16% of the bigrams in the generated summary

are relevant to the reference summary. The F1-score

of 0.4521 reflects a balanced measure of bigram

overlap.

ROUGE-L Score (Longest Common

Subsequence): The ROUGE-L score assesses the

longest common subsequence between the generated

and reference summaries. A recall of 0.6247 suggests

that 62.47% of the longest common subsequence in

the reference summary was captured in the generated

summary. The precision of 0.5438 indicates that

54.38% of the longest common subsequence in the

generated summary is relevant. The F1-score of

0.5585 provides a balanced evaluation of recall and

precision for LCS overlap.

Table 1: ROUGE Scores.

Model Recall Precision F-measure

Rouge-1 0.6466 0.5636 0.5783

Rouge-2 0.5359 0.4316 0.4521

Rou

e-L 0.6247 0.5438 0.5585

Table 1 presents the ROUGE scores, highlighting

the model’s accuracy in generating coherent and

informative summaries.

The performance of the summarization model is

further evaluated using BERT Score, a metric that

leverages pre-trained language models to assess the

semantic similarity between the generated and

reference summaries. BERT Score computes

precision, recall, and F1 scores based on

contextualized embeddings, providing a more

nuanced evaluation compared to traditional overlap-

based metrics.

The summarization model's performance was

assessed on a representative subset of a dataset

containing summaries, providing a robust indication

of its effectiveness in generating semantically relevant

summaries. The results, as summarized by the average

BERT Score metrics, are as follows:

BERT Precision: The precision score of 0.8791

indicates that 87.91% of the words in the generated

summary were relevant, capturing the key

information from the reference summary.

BERT Recall: The recall score of 0.8887 shows

that 88.87% of the key words in the reference

summary were captured in the generated summary.

This high recall demonstrates the model’s ability to

capture a large proportion of the essential information

from the reference summary, ensuring completeness.

BERT F1-Score: The F1-score of 0.8836 balances

both precision and recall, indicating that the

generated summaries are both highly relevant and

comprehensive.

Table 2: BERT Score.

Metric Score

BERT Precision 0.8791

BERT Recall 0.8887

BERT F1-Score 0.8836

Table 2 presents the BERT Score metrics,

highlighting the model’s ability to generate

semantically relevant and accurate summaries based

on contextualized embeddings.

In addition to the ROUGE, BERT Score

evaluations, the text summarization application

developed using Streamlit provides users with an

intuitive interface to generate summaries from

various text inputs. The application consists of a user-

friendly layout where users can upload documents or

paste text directly into a designated input area. Upon

submission, the application processes the input and

displays both the summarized and original texts,

alongside their respective word counts

Key Features and Functionalities.

Input Interface: Users can easily upload

documents in PDF or Word format or input text

directly.

Text Summarization: The application utilizes a

robust BART model for abstractive summarization

and a KL Divergence approach for extractive

summarization.

Figure 2: Web App for Text and Document Summarization.

Integrating Extractive and Abstractive Summarization: A Hybrid Approach

787

5 FUTURE WORK

In future work, several avenues can be explored to

enhance the capabilities of the text summarization

system. First, efforts will focus on improving model

performance by integrating advanced speech-to-text

(STT) technologies, allowing the system to generate

accurate transcripts from video audio, thereby

broadening the range of input sources.

Another significant direction involves extending

the hybrid summarization approach to handle multiple

documents simultaneously. This enhancement would

enable more comprehensive content synthesis by

integrating information from various sources and

identifying common themes or patterns. It would also

facilitate cross-document analysis, enabling users to

draw richer and more insightful conclusions from

diverse datasets, thus broadening the scope and

applicability of the summarization system.

Additionally, deploying the summarization

pipeline in real-time applications represents a

promising opportunity. By adapting the system for

platforms like news aggregators or chatbot interfaces,

users could receive timely and relevant information

summaries, improving the overall user experience.

Lastly, addressing multilingual summarization is

crucial for expanding the system’s reach. By

leveraging transformer models adept at handling

diverse languages, the methodology could support a

wider audience and cater to the global demand for

effective text summarization.

By pursuing these future directions, the project

aims to significantly advance the effectiveness and

applicability of text summarization technologies.

6 CONCLUSION

This research project focused on hybrid text

summarization using KL Divergence and BART,

demonstrating significant potential in generating

concise and informative summaries from textual data.

By integrating both extractive and abstractive

techniques, the project effectively leveraged the

strengths of KL Divergence for content relevance in

sentence selection and BART for producing fluent

and coherent summaries.

The findings highlight the efficacy of this hybrid

approach, showcasing its ability to create more

effective and contextually aware summarization

solutions. As natural language processing technology

evolves, the integration of diverse methods becomes

increasingly essential for addressing the growing

demand for efficient information extraction and

synthesis across various applications.

The project lays a solid foundation for further

advancements in the field of text summarization. The

combination of statistical methods and deep learning

techniques presents a robust framework for

developing innovative solutions. Future

enhancements could involve refining the model

through feature engineering, real-time data

integration, and exploring additional transformer

architectures, ultimately contributing to the ongoing

evolution of text summarization methodologies and

their applications in a wide range of domains.

REFERENCES

A. Ghadimi, & H. Beigy. (2022). Hybrid Multi Document

Summarization Using Pre-Trained Language Models.

Expert Systems with Applications.

Asha Rani Mishra, V.K Panchal, & Pawan Kumar. (2019).

Extractive Text Summarization- An effective approach

to extract infromation from text. International

Conference on Contemporary Computing and

Informatics (IC3I).

B Rajesh, K Nimai Chaitanya, P Tejesh Govardhan, K

Krishna Mahesh, & B Sudarshan. (2024). Text

Summarization Using NLP. International Research

Journal of Engineering and Technology(IRJET).

Christian, H. Agus, M.P., & Suhartono. (2016). Single

Document Automatic Text Summarization Using Term

Frequency Inverse document frequency (TF-IDF).

ComTech: Computer, Mathematics and Engineering

Applications, 285-294.

G. Di Fabbrizio, A. Stent, & R. Gaizauskas. (2014). A

Hybrid Approach to Multi-Document Summarization

of Opinions in Reviews. INLG, 54-63.

Glorian Yapinus, Alva Erwin, Maulhikmah Galinium, &

Wahyu Muliady. (2014). Automatic Multi-Document

Summarization for Indonesian Documents Using

Hybrid Abstractive Extractive Sumarization

Technique. International Conference on Information

Technology and Electrical Engineering (ICITEE).

J.N. Madhuri, & R. Ganesh Kumar. (2019). Extractive Text

Summarization Using Sentence Ranking. International

Conference on Data Science and Communication

(IconDSC).

L. Lebanoff, K. Song, & F. Liu. (2018). Adapting the

Neural Encoder Decoder Framework from Single to

Multi Document Summarization. ACLEMNLP.

Sanchit Agarwal, Nikhil Kumar Singh, & Priyanka Meel.

(2018). Single Document Summarization Using

Sentence Embeddings and K-Means Clustering.

International Conference on Advances in Computing,

Communication Contral and Networking (ICACCCN).

Wen Xiao, & Giuseppe Carenini. (2019). Extractive

Summarization of Long Documents by Combining

Global and Local Context. EMNLPIJCNLP.

INCOFT 2025 - International Conference on Futuristic Technology

788