Journal of Computational Analysis and Applications VOL. 33, NO.

2, 2024

Advancing Multi-Document Text Summarization through

Deep Learning and Personalization Techniques
Veena R1*, Dr. Ramesh2, Dr. Hanumanthappa M3
1ResearchScholar, Sri Siddhartha Academy of higher Education, Tumkur-572105,
Email: [email protected]
2Professor &Head, Department of Master of Computer Applications, Sri Siddhartha Institute of

Technology, Tumkur-572105
3Professor &Chairman, Department of Computer Science and Applications, Bangalore University,

Bangalore 560056.
*Corresponding Author

Received: 08.04.2024 Revised: 15.05.2024 Accepted: 24.05.2024

ABSTRACT
This study examines the impact of customisation on multi-document text summarisation (MDTS) and the
use of deep learning techniques to enhance summary quality. We demonstrate significant improvements
in summarisation tasks across various texts by including pre-trained Long Short-Term Memory (LSTM)
networks and Recurrent Neural Networks (RNNs) into our model. These networks are used to derive
insights from prior experiences. By integrating user selections into the document encoding process, the
model effectively captures complex relationships, yielding summaries that are both valuable and concise.
The results of our empirical study indicate that our customised LSTM and RNN-based approach surpasses
many current benchmarks. This illustrates that customisation effectively generates summaries that are
more relevant to the current context. We advise doing more research into sophisticated graph-based
representations, such as knowledge graphs, to augment the semantic depth of summaries and more
accurately depict the intricate structures of texts. This work not only emphasises the potential for
enhancing MDTS via the integration of deep learning and user-centric customisation but also provides
future pathways for improving the robustness and flexibility of models.

Keywords: LSTM, results, RNN, flexibility, models.

INTRODUCTION
In today's information age, the vast amount of textual data available on the internet presents both
opportunities and challenges. While access to information has never been easier, the sheer volume of data
can overwhelm users seeking to extract relevant insights efficiently. Multi-document text summarization
emerges as a critical solution to this challenge, offering a means to distill large collections of documents
into concise and informative summaries.
The primary objective of multi-document text summarization is to produce condensed representations
that capture the essential information across multiple documents while retaining the key insights and
context present in the original texts. By condensing the content into a more manageable form, summaries
enable users to grasp the core ideas and themes without the need to sift through extensive volumes of
text. Traditional approaches to multi-document summarization have often relied on rule-based or
statistical methods, which may struggle to capture the nuances and complexities of natural language.
However, recent advancements in deep learning techniques have opened new avenues for improving the
quality and effectiveness of text summarization. By leveraging neural networks and large-scale language
models, researchers have achieved remarkable progress in generating summaries that exhibit greater
coherence, relevance, and fluency
In this paper, we introduce an innovative method for multi-document text summarization that extends
the capabilities of deep learning techniques while integrating personalized elements to customize the
summaries according to individual user preferences and requirements. Inspired by recent progress in
personalized recommendation systems and natural language processing, our approach endeavors to
provide summaries that not only encapsulate the essential information from multiple documents but also
align with the user's interests and goals. By integrating deep learning models, personalized profiling, and
content analysis, our methodology aims to advance the frontier of multi-document text summarization,
providing users with a customized and user-friendly approach to traverse the extensive realm of online

871 Veena R et al 871-877

Journal of Computational Analysis and Applications VOL. 33, NO. 2, 2024

textual information. Through the utilization of machine learning and natural language understanding, our
goal is to equip users with the capabilities necessary to comprehend the continuously expanding digital
knowledge ecosystem. In the following sections of this paper, we will explore the specifics of our
proposed methodology, covering aspects such as the architecture of our deep learning models, the
methods utilized for personalized profiling, and the evaluation framework employed to gauge the
effectiveness of our approach. Furthermore, we will provide insights into experimental outcomes and
examine the implications of our findings, shedding light on potential applications and avenues for future
research in personalized multi-document text summarization

2. Related Works
[1] In recent years, multi-document text summarization has garnered attention due to the proliferation of
textual data. Veningston et al. (2023) introduce a personalized approach using deep learning, building on
prior NLP and machine learning research. Extractive and abstractive methods have been explored, with
deep learning, particularly RNNs and transformer models, showing promise. Challenges include
information fusion and coherence, addressed through user feedback and domain-specific knowledge
integration. The field continues to evolve with ongoing research aimed at developing innovative
techniques for summarizing diverse textual data sources.
[2] Recent research in multi-document summarization has focused on developing innovative approaches
to generate concise summaries from multiple documents. Puduppully et al. (2022) propose centroid-
based pretraining, where a model learns document embeddings representing the centroid of a document
set to guide summary generation. This method aims to capture essential information shared across
documents. Challenges in multi-document summarization include effectively capturing the essence of
multiple documents and generating coherent summaries. Prior research has explored various techniques,
including extractive and abstractive methods, as well as the integration of external knowledge sources.
Despite advancements, there are ongoing efforts to improve the summarization of diverse textual
sources.
[3] Recent studies in multi-document summarization have delved into inventive strategies to augment
summarization models. Ketineni and Sheela (2023) introduce a hybrid optimization model, which
integrates metaheuristic algorithms with LSTM networks. Despite advancements, challenges persist in
optimizing models to generate informative and coherent summaries. Their approach utilizes
metaheuristic algorithms such as genetic algorithms to optimize LSTM parameters. This hybrid model
endeavors to enhance LSTM-based summarization by improving information capture. Current research
efforts in the field are concentrated on devising models for concise and informative summaries..
[4] Extractive multi-document summarization remains a formidable challenge within the realm of natural
language processing .Ghadimi and Beigy (2023) present SGCSumm, a novel approach that incorporates
pre-trained language models, sub modularity, graph convolutional neural networks. Their methodology
is geared towards producing informative summaries by harnessing a variety of techniques. Previous
research has explored various methods, but challenges persist in capturing document semantics and
structure. SGCSumm addresses these challenges by combining complementary approaches. This
integration allows SGCSumm to produce high-quality summaries from multiple documents. Ongoing
efforts in the field focus on developing innovative techniques to improve summary quality.
[5] Query-focused multi-document summarization (QMDS) is a significant research area in natural
language processing. Roy and Kundu (2023) conduct a comprehensive review of QMDS techniques,
providing insights through comparative analysis. Previous research has explored diverse approaches,
including retrieval-based and generation-based methods. Challenges persist in summarizing multiple
documents while considering specific user queries. The review categorizes existing techniques and
evaluates their performance using standard metrics. This analysis guides the development of more
effective QMDS approaches. Ongoing efforts aim to address the complexities of QMDS and improve
summary quality.
[6] Automatic multi-document text summarization poses a significant challenge within the field natural
language processing. Abo-Bakr and Mohamed (2023) propose a large-scale sparse multi-objective
optimization algorithm to address this challenge. Their approach aims to generate high-quality
summaries by optimizing multiple objectives simultaneously. Previous research has explored various
techniques, including graph-based methods and neural network architectures, but challenges persist in
efficiently summarizing large volumes of textual data while ensuring quality. The algorithm introduced by
Abo-Bakr and Mohamed (2023) leverages sparse representation techniques and multi-objective
optimization to efficiently optimize summary informativeness and diversity. This approach enables the
algorithm to generate high-quality summaries from multiple documents. In summary, recent research

872 Veena R et al 871-877

Journal of Computational Analysis and Applications VOL. 33, NO. 2, 2024

continues to advance with the development of innovative algorithms for automatic multi-document text
summarization.
[7] Multi-document summarization for learning materials is crucial in educational technology research.
Sakkaravarthy Iyyappan and Balasundaram (2023) propose a novel approach combining concept-based
Inductive Logic Programming (ILP) and clustering methods to address this challenge. Their methodology
aims to generate informative summaries by augmenting document elements with concepts and
relationships extracted from the text. Previous research has explored various techniques, but challenges
persist in effectively summarizing learning materials while maintaining content relevance. The approach
integrates ILP techniques and clustering methods to enhance summarization. This enables the generation
of informative and concise summaries tailored to learning materials. In summary, recent research
continues to advance with the development of innovative approaches for multi-document summarization
in educational contexts

3. Methodology Description
3.1 Research Focus
Advancing machine comprehension presents a critical challenge in artificial intelligence research. Despite
advancements, machines still encounter difficulties in reading and comprehending text with human-like
understanding. Hermann et al. (2015) introduced innovative methods in "Teaching Machines to Read and
Comprehend," yet obstacles persist in grasping context and reasoning. This study endeavors to tackle
these challenges by developing original techniques to augment machine comprehension capabilities.
Through the utilization of neural network architectures and training methodologies, our aim is to expand
the frontiers of machine comprehension. Our goal is to narrow the disparity between human and machine
comprehension of textual information, laying the groundwork for more sophisticated natural language
understanding systems.

3.2. Text Processing

Prior to model training, it is essential to preprocess the raw data to ensure optimal learning outcomes.
Disorganized data can hinder learning despite its volume, necessitating preprocessing techniques to
enhance model performance. Several preprocessing steps are employed to prepare the data before
feeding it into the model.
Whitespace removal: Unnecessary whitespace is removed from the text to ensure uniformity and
facilitate subsequent processing steps.
Stop word removal: Common stop words, which carry little semantic meaning, are eliminated from the
text to reduce noise and focus on relevant content.
Special symbol removal: Unreadable special symbols are filtered out from the text to improve
readability and avoid potential encoding issues.
Contraction Mapping: Contraction mapping is employed to address contractions within the text,
enhancing consistency and facilitating better comprehension in the summary text.
Lowercasing: Documentary informationis converted to lowercase to standardize the text and avoid
redundancy in model training. However, some standard acronyms are preserved in their original case to
maintain their semantic significance.

3.3 Designing a Customized Recursive Neural Network Structure for Summarization.

A new method for multi-document text summarization is introduced, utilizing a Recursive Neural
Network (ReNN) architecture enhanced with Convolutional Neural Networks (CNNs) and Siamese
Networks (SNs). This architecture is designed to tackle the challenge of summarizing extensive textual
data while integrating user preferences to generate personalized summaries. The ReNN initially
processes each sentence individually, extracting Word-to-Vec embeddings using Global Vectors for Word
Representation (GloVe) embedding, and then passing them through a CNN to acquire feature
representations. These representations serve as the foundation of a hierarchical structure, where each
layer captures progressively abstract features of the input text. At each level, a ReNN is constructed to
recursively process the input text and capture both local and global dependencies. The output
representations from each layer of the hierarchical structure are merged to produce the final summary,
which is tailored to the user's preferences using SNs to evaluate the similarity between user preferences
and summary candidates.
Incorporating user preferences into the summarization process is crucial for producing contextually
relevant summaries. To accomplish this, a Siamese Network (SN) is utilized to evaluate the similarity
between user preferences and potential summaries. The SN comprises two identical sub-neural networks
with shared weights, enabling the comparison of feature vectors between user representation and

873 Veena R et al 871-877

Journal of Computational Analysis and Applications VOL. 33, NO. 2, 2024

summary candidates. The proximity or similarity between user preferences and the output summary is
estimated using an objective function, facilitating the generation of personalized and contextually
relevant summaries. In summary, the proposed ReNN architecture, in conjunction with CNNs and SNs,
allows for the integration of user preferences into the summarization process, resulting in summaries
that effectively capture the essential information from multiple documents while closely aligning with the
user's interests and objectives.
L = (Upreference, Csummary, y) (1)
The proximity or similarity between user preferences Up and the output summary Cs is approximated
using Equation (1), where y represents a Binary signal denoting if the two text segments are from the
same category
In this approach, the temporal aspect of user preference is integrated, a feature absent in many related
works. The sequence in which a user reads articles is deemed crucial; thus, the order of article
consumption influences subsequent article choices. The user's article reading history is conceptualized as
a sequential data set, capturing the order in which articles are consumed. The objective is to summarize a
collection of news articles according to the preferences of user U, leveraging the user's historical data. A
ReNN or LSTM-based encoder is employed to encode information from the user's historical data, which is
then combined with the representation of candidate news articles. Subsequently, the representation with
the highest similarity to the user's reading history is selected to generate the summary. The user's
reading history serves as the basis for constructing the user preference model, encompassing articles
read up to time period tUtilizing both traditional LSTMs and Attentional LSTMs, the outputs of the user
model and candidate articles are combined and processed through additional layers or a Multilayered
Perceptron. The resulting summary of multiple news articles is refined through backpropagation,
facilitating weight updates across all layers, from the Multilayered Perceptron to the RNN layers, if the
summary fails to meet expectations [11, 12].

Fig.1: User Preference Integration in Text Summarization.

4. Experimental Assessment
4.1 Data Discription
Evaluation of the Proposed Work Using Two Distinct Datasets
1. Daily mail dataset
SIZE: This dataset contains approximately 200,000 text documents
Description: The documents in this repository are sourced from news articles published by the Daily
Mail, covering a wide range of topics and domains

874 Veena R et al 871-877

Journal of Computational Analysis and Applications VOL. 33, NO. 2, 2024

2. Multinews dataset
Source :The MultiNews dataset is composed of news articles accompanied by human-transcribed precise
summaries
Discription : This dataset comprises news articles collected from a diverse array of over 1,500 news
websites. Each article is accompanied by a human-generated summary that encapsulates the key points of
the article's content

4.2 Baseline model

1 TextRank
In our comparative analysis, the first baseline model we consider is TextRank. Operating on graph-based
principles inspired by algorithms like Hyperlink Induced Topic Search (HITS) and PageRank, TextRank
autonomously extracts keywords and sentences for summarization by assessing the significance of
vertices within the text graph.
2 Standard Convolutional neural network
The second baseline model, Standard Convolutional Neural Network (CNN), presents a deep learning
approach to text summarization. It involves passing document embeddings through convolutional layers
with Rectified Linear Unit (ReLU) activation, followed by max-pooling and dropout layers. Subsequently,
fully connected layers with sigmoid activation are utilized, and error estimation is performed using the
cross-entropy loss function.
3 Hierarchical Network
We examine the Hierarchical Network (HNet) model. HNet combines architectures of Convolutional
Neural Network (CNN) and Recurrent Neural Network (RNN) to process each word in a sentence and
derive Word2Vec embeddings, resulting in a matrix representation for each sentence. These matrices are
treated as images and passed through CNN layers to generate representations for each sentence. These
representations form nodes in a hierarchical tree structure imposed on an RNN, enabling comprehensive
semantic comprehension in text summarization tasks [15].

4.3. Evaluation metrics used

In evaluating the efficacy of our proposed summarization technique, We utilize the 'Recall-Oriented
Understudy for Gisting Evaluation (ROUGE) score’, a widely accepted metric in text summarization
research. We specifically employ multiple variants of the ROUGE metric [13, 14] to ensure a
comprehensive assessment of our deep learning-based text summarization approach.
ROUGE-1 score :This evaluation metric assesses the agreement between the unigrams present in the
reference Summary of the test instance and the summary generated by the model.
ROUGE-2 score:The ROUGE-2 score assesses the agreement between the bigrams in the reference
Summary of the test instance and the summary generated by the model.
ROUGE-L score : This metric quantifies the length of the longest common subsequence present in both
the reference Summary of the test instance and the summary generated by the model. Through the
identification of the longest overlapping sequence of n-grams shared between them, the ROUGE-L score
evaluates the similarity between the reference summary and the generated output.

4.4. Assessment of LSTM and Recursive neural network Model Performance

Over time, numerous foundational models have been developed for text summarization, each aimed at
enhancing summarization quality. In our evaluation, we compare the performance of various
summarization models, including traditional techniques and deep learning approaches. The experimental
findings are consolidated in Table 1 and Table 2, highlighting the effectiveness of our proposed LSTM and
Recursive Neural Network-based technique

Table 1: Performance Metrics on Daily Mail Dataset

Summary Models ROUGE-1 ROUGE-2
LexRank 32.430 5.880
Standard Convolutional neural network 34.190 7.230
Hierarchical network 37.210 8.960
LSTM and Recursive neural network 41.340 9.510

875 Veena R et al 871-877

Journal of Computational Analysis and Applications VOL. 33, NO. 2, 2024

Table 2: Performance Metrics on MultiNews Dataset.

Summary Models ROUGE-1 ROUGE-2
LexRank 35.290 7.540
Standard Convolutional neural network 35.730 8.690
Hierarchical network 39.170 9.610
LSTM and Recursive neural network 43.940 11.290

Based on the data presented in the tables, it becomes clear that deep learning methodologies,
encompassing standard Convolutional Neural Network (CNN), Hierarchical Network, and LSTM-based
summarization techniques, demonstrate superior accuracy as indicated by ROUGE scores in comparison
to the traditional LexRank-based model. Among the array of deep learning models evaluated, our
proposed LSTM and Recursive Neural Network (RNN)-based model consistently exhibits better
performance across both datasets.
These results highlight the efficacy of our proposed approach based on LSTM and Recursive Neural
Network in producing high-quality summaries, underscoring its potential for practical applications in text
summarization tasks.

4.5. Assessment of Personalized Model

Data Collection
To evaluate the effectiveness of personalized multi-document summarization, we collected news articles
from a commonly used search engine, specifically Google. Using the headlines of news articles as search
queries, we retrieved 15 results per query, typically selecting from the top results. We then extracted only
the text content from these news articles to serve as input data for our summarization task. Additionally,
in our evaluation process, we incorporated the search histories of 100 users to provide further validation
for the proposed model. This data collection methodology was chosen to ensure a reasonable degree of
diversity, as there is currently no established benchmark dataset available for evaluating personalized
summarization models."

Assessment by Human Judges:

Alongside objective evaluation, we subjected the proposed approach to subjective human assessment.
Three skilled annotators proficient in English script writing evaluated fifty randomly chosen test
examples from the MultiNews test set. These annotators ranked the summaries based on criteria such as
in formativeness, fluency, and succinctness. Their evaluation focused on determining if the final summary
effectively conveyed important facts from the input articles, maintained fluency and grammatical
correctness, and avoided redundant information.
Reviewers assessed the summaries produced by each technique on a scale ranging from 1 to 5, where 1
represented the highest quality and 5 the lowest. If different techniques demonstrated similar quality,
they could receive the same ranking position. The rating for each technique was calculated as the average
score across all test examples used in the experiment

Table 3. Ranking Results of Summaries According to Human Assessment

Summary models Reviewer 1 Reviewer 2 Reviewer 3 Average rating
LexRank 27.00 25.00 28.00 26.670
Standard Convolutional neural network 31.00 30.00 30.00 30.340
Hierarchical network 35.00 37.00 38.00 36.670
LSTM and Recursive neural network 39.00 41.00 41.00 40.340

These rankings provide insight into the perceived quality of summaries generated by each technique, as
assessed by human annotators. The LSTM and ReNN-based model consistently achieved the highest
average score across all annotators, indicating its effectiveness in producing high-quality summaries
tailored to individual preferences

CONCLUSION
In summary, this paper has examined the importance of personalization in text summarization and
utilized deep learning techniques to improve the quality of summaries. Our approach, which involves
incorporating pre-trained LSTMs and RNNs into our model, has resulted in notable improvements in
multi-document summarization tasks. Integrating user preferences into the encoding process of
documents has enabled us to capture more nuanced relationships, resulting in summaries that are both
informative and concise. Empirical evidence indicates that our model, which utilizes personalized LSTM

876 Veena R et al 871-877

Journal of Computational Analysis and Applications VOL. 33, NO. 2, 2024

and RNN architectures, surpasses several established benchmarks by a significant margin. Moving
forward, future research will explore the integration of additional graph representation models, such as
knowledge graphs, to further enhance the quality of generated summaries.

REFERENCE
[1] Veningston, K., Rao, P. V., & Ronalda, M. (2023). Personalized Multi-document Text Summarization
using Deep Learning Techniques. Procedia Computer Science, 218, 1220-1228.
[2] Puduppully, R., Jain, P., Chen, N. F., & Steedman, M. (2022). Multi-document summarization with
centroid-based pretraining. arXiv preprint arXiv:2208.01006.
[3] Ketineni, S., & Sheela, J. (2023). Metaheuristic Aided Improved LSTM for Multi-document
Summarization: A Hybrid Optimization Model. Journal of Web Engineering, 22(4), 701-730.
[4] Ghadimi, A., & Beigy, H. (2023). SGCSumm: An extractive multi-document summarization method
based on pre-trained language model, submodularity, and graph convolutional neural
networks. Expert Systems with Applications, 215, 119308.
[5] Roy, P., & Kundu, S. (2023). Review on Query-focused Multi-document Summarization (QMDS) with
Comparative Analysis. ACM Computing Surveys, 56(1), 1-38.
[6] Abo-Bakr, H., & Mohamed, S. A. (2023). Automatic multi-documents text summarization by a large-
scale sparse multi-objective optimization algorithm. Complex & Intelligent Systems, 1-16.
[7] Sakkaravarthy Iyyappan, K., & Balasundaram, S. R. (2023). A novel multi document summarization
with document-elements augmentation for learning materials using concept based ILP and
clustering methods. International Journal of Computers and Applications, 1-12.
[8] Christensen, J., Soderland, S., & Etzioni, O. (2013, June). Towards coherent multi-document
summarization. In Proceedings of the 2013 conference of the North American chapter of the
association for computational linguistics: Human language technologies (pp. 1163-1173)
[9] Mascarell, L., Chalumattu, R., & Heitmann, J. (2023). Entropy-based Sampling for Abstractive Multi-
document Summarization in Low-resource Settings. In 16th International Natural Language
Generation Conference (INGL 2023)..
[10] Liu, S., Zhou, M. X., Pan, S., Qian, W., Cai, W., & Lian, X. (2009, November). Interactive, topic-based
visual text summarization and analysis. In Proceedings of the 18th ACM conference on Information
and knowledge management (pp. 543-552).
[11] Chen, K. Y., Liu, S. H., Chen, B., Wang, H. M., Jan, E. E., Hsu, W. L., & Chen, H. H. (2015). Extractive
broadcast news summarization leveraging recurrent neural network language modeling
techniques. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 23(8), 1322-1334.
[12] Lin, C. Y. (2004, July). Rouge: A package for automatic evaluation of summaries. In Text
summarization branches out (pp. 74-81).
[13] Lin, C. Y., & Och, F. J. (2004, July). Automatic evaluation of machine translation quality using longest
common subsequence and skip-bigram statistics. In Proceedings of the 42nd Annual Meeting of the
Association for Computational Linguistics (ACL-04) (pp. 605-612).
[14] Singh, A. K., Varma, V., & Gupta, M. (2018). Neural approaches towards text
summarization. International Institute of Information Technology Hyderabad.

877 Veena R et al 871-877