Clustering Technique for Policy Categorization and Text Summarization based on Deep Learning Approaches

Abstract

Every citizen has the right to control their own data in this digital society. For data privacy, policies are created by social media applications. The users often skip the entire policies of websites or applications to save energy and time without realizing the main points of these policies. Moreover, most of the hypermedia users won’t read the policies because of the verbose explanations and obscure language. If these privacy policies are summarized and classified into useful data, that will be beneficial for users. For that reason, text summarization models are used, which improve social media privacy by automatically summarizing very long privacy policies for users. Recently, many machine learning-based text summarization models have been introduced by researchers. However, these models suffer from issues like inaccuracies, scalability issues, and difficulty in tuning. To solve this problem, in this work, a hybrid classifier (transfer learning (TL)-based modified long short-term memory (LSTM) + bi-directional gated recurrent unit (Bi-GRU)) based policy categorization and text summarization model is proposed, starting with policy categorization as the initial step. During this phase, the data annotation process takes place using an Improved fuzzy C-means algorithm to provide an accurate policy categorization by dividing the policies into paragraphs. After clustering, a preprocessing phase is applied to each cluster, and features such as smoothed dispersion coefficient (SDC)-based term frequency-inverse document frequency (TF-IDF), Thematic features, and Bag of Words are extracted. These features are then used in the text summarization phase, where a hybrid classifier, combining TL-based modified LSTM and Bi-GRU models, is utilized. The fusion of TL-based modified LSTM and Bi-GRU produces a single model that generates the final summarized text. Analysis shows that the proposed hybrid classifier (TL-based modified LSTM + Bi-GRU) achieved superior accuracy of 95% and 95.6% on both APP-350 and BillSum datasets.

Keywords

text summarization improved FCM improved TF-IDF TL-based modified LSTM Bi-GRU

1 Introduction

Each user has the right to protect and control their personal data while using social media. To preserve the privacy of the user's personal data, the policies are utilized by the social media applications. These privacy policies outline the rules that a specific application promises to follow concerning the data of the users (Abdi et al., 2021; Ertam & Aydin, 2022). Generally, this data includes the kind of data, the way to use the data, and in which situations the data will be shared, who has access to the data, and the security concerning the data storage and a lot more (Jang & Kim, 2021). It is beforehand mentioned all the risks that a user might take via agreeing to the application's policies (Tomer & Kumar, 2020; Xu et al., 2020).

The main problem is caused by the user's inability to understand the policy and comprehend the risks related to it because of some reasons (Bhuyan et al., 2023; Li et al., 2020). Today, almost everyone has a smartphone, which even enables illiterate people to get access to diverse applications due to the technology advancement (Rahul Raj et al., 2020). Since they don’t have any technical knowledge, they agree to the policies without reading and understanding them (Alqaisi et al., 2020; Hou & Lu, 2020). Moreover, some users simply accept these policies without reading them due to the arduous effort required in comprehending the verbose and technical terms used by these privacy policies (Yang et al., 2020). To save the time and effort of the users, these privacy policies should be summarized and categorized into valuable data. Text summarization can be a better solution for that (Wijayanti et al., 2023; You et al., 2020). Text summarization plays a crucial role in today's era of information overload by helping users quickly comprehend large volumes of text or policies through condensed summaries that focus on the most important details (Belwal et al., 2021; Hazra et al., 2021).

Since privacy policies include many attributes and plenty of text, a direct summary might exclude important information. Classifying the policies according to the attributes it covers is therefore necessary to address that issue (Alqaisi et al., 2020; Jindal & Kaur, 2020). Text that is categorized into more than two classes is referred to as a multiclass text classification challenge (Wazery et al., 2023). The current approaches to solving the multiclass classification issue center on applying machine learning (ML) methods such as support vector machines (SVM) and Naïve Bayes (NB) (Alami et al., 2021a; Cheng et al., 2020). An alternative approach to address the issue of multiple categorizations is to approach each problem as a binary classification, solve it one at a time, and then combine the outcomes of these binary classifications (Ghodratnama et al., 2020; Su et al., 2020; Yao et al., 2018). Existing advancements have been made in the use of deep learning (DL) and neural network (NN)-based methods for text classification, document modeling, and other natural language processing (NLP) problems (Alami et al., 2021b; Deotale et al., 2021; Hernández-Castañeda et al., 2022). Convolutional neural network (CNN) and recurrent neural network (RNN)-based frameworks (Ertam & Aydin, 2022) are crucial among these methods for labeling numerous categories in generative models that create text artificially for training and classification (Sanchez-Gomez et al., 2021; Sanchez-Gomez et al., 2021). Capsule networks are used for difficult NLP tasks, including multilabel classification of texts and question answering (QA), whereas long short-term memory (LSTM) models are used for aspect-based sentiment analysis (ABSA) in recent advances in DL (Basheer et al., 2020; Du & Huo, 2020; Khaleghi et al., 2021). Moreover, other conventional summarizing techniques frequently lack contextual awareness, along with inaccuracy, scalability issues, and difficulty in tuning. To tackle such challenges, we have developed a hybrid classifier (HC) (transfer learning (TL)-based modified LSTM + bi-directional gated recurrent unit (Bi-GRU)) for policy categorization and the text summarization process. This model attempts to address these issues by utilizing the advantages of both sequence modeling architectures. To improve the relevance and coherence of the resulting summaries, advanced feature extraction techniques (smoothed dispersion coefficient (SDC)-based term frequency-inverse document frequency (TF-IDF), Thematic Features, BOW (Bag of Words)), and an EDCos membership-based fuzzy C-means (FCM) algorithm for policy classification are incorporated. The work is essentially motivated by the requirement for intelligent, precise, and category-aware text summarization utilizing a deep learning architecture that guarantees both computational efficiency and semantic richness, especially for policy texts. Below are the contributions of the proposed text summarization model:

Introducing an EDCos membership-based FCM algorithm to provide an effective data annotation process, which leads to an accurate and reliable policy categorization.

Extracting a novel SDC-based TF-IDF feature along with conventional thematic and BoW features to offer a better summarization.

Introducing a new text summarization approach that utilizes a hybrid model that merges both TL-based modified LSTM and Bi-GRU to provide an effective summarization by introducing a new activation function within the modified LSTM model. This new activation function includes softmax, modified comb-H-sine, and hard Tanh activation functions.

The review of recent policy categorization and text summarization-based papers is in Section 2. Novel hybrid DL-based policy categorization and text summarization is described in Section 3. Sections 4 and 5 provide the results and conclusion parts.

2 Literature Review

2.1 Related Works

A few recent policy categorization and text summarization-based papers are reviewed below. A unique DL-based technique termed as RDLS was presented by Abdi et al. (2021) in 2021 for the general opinion-based summarization and extraction of multiple documents. Opinion summarizer module (OSM), text summarization embedding spaces (TSS), and sentiment analysis embedding space (SAS) were the components of the approach. RNNs, which were built using LSTM, were used by SAS to benefit from sequential processing, and it prevailed over several problems, including loss of word order and information. TSS enhanced word-level embedding and extracted a suitable group of sentences from numerous documents by utilizing characteristics from several sources, including linguistic and statistical information. The restricted Boltzmann machine (RBM) method was another tool that TSS employed to optimize and refine those characteristics, increasing the accuracy of the output without sacrificing any crucial data. Sentence selection and categorization were the two stages of OSM, which combined to create a better summary.

In 2022, Alahmadi et al. (2022) introduced a unique topic-aware abstractive Arabic summarization model (TAAM) that used an RNN in an effort to improve text summarization accuracy. In TAAM, two experiments, one quantitative and the other qualitative, were carried out. The TAAM model, which relied on a statistical approach with ROUGE matrices, outperformed other baseline models in terms of accuracy by 10.8%. Additionally, the TAAM approach was able to generate an easily readable, cohesive Arabic summary that conveyed the primary concept of the input text, using a qualitative technique that gathered users’ viewpoints.

In 2023, Wijayanti et al. (2023) looked into two bilingual word-encoding techniques for the Indonesian English language, VecMap and BiVec, and assessed how well they performed on tasks including text summarization and bilingual lexicon induction. The produced bilingual embedding was contrasted with the multilingual term constructed using the generative adversarial network approach, known as multilingual unsupervised and supervised embeddings (MUSE). Creating standard word spaces and transferring the private ones across languages also helped to enhance the VecMap.

A unique extractive text summarizing technique that utilizes word embedding and quantitative characteristics of a single document was described by Yaser M. Wazery et al. in 2023 (Wazery et al., 2023). Based on word embeddings and statistical data, a CNN and an feed-forward neural network (FFNN) were used to encode each phrase. An multilayer perceptron (MLP) was used to categorize the text by concatenating the outputs of the CNN and FFNN. To find the most effective hybrid model, the KerasTuner optimization (KTO) approach was also used to tune the parameters of the hybrid model.

A textual entailment was proposed by Bhuyan et al. (2023) in 2023 as a possible metric for assessing abstractive summaries. The findings demonstrated the usefulness of text representation as a potent automated assessment approach for these types of summaries. To provide a reasonable indication of the efficiency of the created summaries, textual entailment scores were computed between the text and the generated summaries, as well as between the reference and forecasted summaries. These values were then combined to provide an overall summarizer score.

In 2021, Deotale et al. (2021) deployed using ML-based models for policy categorization, which would classify policy paragraphs according to suggested qualities such as security, contact, and so on. An artificial neural network (ANN) approach outperformed other ML-based classifier models on a difficult dataset of linguistic privacy regulations, as demonstrated by comparing these models against one another. This demonstrated how ML may assist in condensing the pertinent paragraphs under the different qualities so that the user can quickly grasp the essence of the subject.

In 2021, Jiang et al. (2021) introduced four unique automatic text summarization (ATS) models with a sequence-to-sequence (Seq2Seq) structure and an Attention Bi-LSTM. This work contains improvements to stop cumulative mistakes in created text summaries from spreading, solve the issue of out-of-vocabulary (OOV) terms, suppress repetitive words, and increase the correlation between produced text summaries and the original text.

In 2020, Hernández-Castañeda et al. (2020) suggested a novel approach to the ATS problem that improves keyword recognition by utilizing semantic information. By grouping the sentences to determine the primary subjects in the source material, this suggested technique improved both coverage and precision. Additionally, it finds the key phrases within the clusters.

In 2025, Verma et al. (2025) introduced the generative artificial intelligence-based modified abstractive cross-attention enabled sequence-to-sequence model (GAI-MACSeqRT) for abstractive Hindi text summarization, which incorporated the MAC attention mechanism, an architecture of encoder-decoder, a model of sequence-to-sequence transformer including a generative AI. The accuracy and quality of Hindi text summarization were enhanced by the multi-objective-based graph embedding and pre-processing techniques. The experimental outcomes confirmed that the GAI-MACSeqRT model creates meaningful summaries while maintaining a lower computational cost.

In 2025, Deo et al. (2025) introduced a new multi-document summarization (MDS) technique named CLSTM-MDS. Initially, using the BERT tokenizer, the input data was converted into tokens. Afterwards, features like TF-IDF, BoW, thematic features, along with improved aspect term-based features, were extracted. Finally, using the CLSTM with a pre-eminent layer, the summarization process was conducted. The experimental outcomes proved that CLSTM-MDS provided an accurate and high-quality summary.

In 2025, Phani et al. (2025) deployed a query-based text summarization model named MATSFT. To enhance the overall summarization process, a co-attention mechanism along with a shared encoder-decoder model is employed. The cross-lingual dependencies are captured by the co-attention, which helps the model understand the relationships and nuances among diverse languages. Moreover, an Indian language dataset was created in this work. This model's experimental outcomes were evaluated by language-agnostic target summary evaluation and ROUGE metrics, and it highlighted the MATSFT model's versatility and effectiveness in multilingual summarization tasks.

Features and challenges of some existing text summarization approaches are highlighted in Table 1.

Table 1.
Review of Conventional Methods.

Author Methodology Feature Limitation

Abdi et al., 2021 • RNN-LSTM Greater accuracy was obtained Its inability to differentiate between active and passive sentences is its main flaw.

Alahmadi et al., 2022 • TAAM F1 score, Recall, and Precision values were higher It can be computationally expensive and sluggish, particularly when working with long sequences.

Wijayanti et al., 2023 • Cross-lingual transfer learning (VecMap and Bi-Vec) ROUGE score outperformed classical models This approach's application is limited since data for many language combinations may be hard to come by or prohibitively expensive to acquire.

Wazery et al., 2023 • CNN and FFNN The highest ROUGE score was obtained Statistical characteristics must be added to the input vectors to improve them.

Bhuyan et al., 2023 • Deep learning-based text Summarization It performed better in generating more precise summaries In this work, comparatively simpler and more compact designs were developed on certain epochs due to resource constraints.

Deotale et al., 2021 • POCASUM Maximal Accuracy, F1 score, Precision, and recall values were attained. Still need enhancement in model scalability.

Jiang et al., 2021 • Hybrid ATS model The ROUGE score was maximized Its limited utilization of network tiers is a drawback.

Hernández-Castañeda et al., 2020 • EATS algorithm The highest ROUGE score was obtained When confronted with noisy or ill-structured content, it could find it difficult to correctly identify important sentences.

Verma et al., 2025 • GAI-MACSeqRT model It provides meaningful summaries while maintaining a lower computational cost The performance still needs enhancement since the attained precision and F1 score values are very low.

Deo et al., 2025 • CLSTM-MDS It provided an accurate and high-quality summary This IBi-GRU might increase the model complexity and training time

Phani et al., 2025 • MATSFT More robust in handling various datasets with diverse linguistic complexities. However, the performance can still be optimized via advanced ML models

Author	Methodology	Feature	Limitation
Abdi et al., 2021	• RNN-LSTM	Greater accuracy was obtained	Its inability to differentiate between active and passive sentences is its main flaw.
Alahmadi et al., 2022	• TAAM	F1 score, Recall, and Precision values were higher	It can be computationally expensive and sluggish, particularly when working with long sequences.
Wijayanti et al., 2023	• Cross-lingual transfer learning (VecMap and Bi-Vec)	ROUGE score outperformed classical models	This approach's application is limited since data for many language combinations may be hard to come by or prohibitively expensive to acquire.
Wazery et al., 2023	• CNN and FFNN	The highest ROUGE score was obtained	Statistical characteristics must be added to the input vectors to improve them.
Bhuyan et al., 2023	• Deep learning-based text Summarization	It performed better in generating more precise summaries	In this work, comparatively simpler and more compact designs were developed on certain epochs due to resource constraints.
Deotale et al., 2021	• POCASUM	Maximal Accuracy, F1 score, Precision, and recall values were attained.	Still need enhancement in model scalability.
Jiang et al., 2021	• Hybrid ATS model	The ROUGE score was maximized	Its limited utilization of network tiers is a drawback.
Hernández-Castañeda et al., 2020	• EATS algorithm	The highest ROUGE score was obtained	When confronted with noisy or ill-structured content, it could find it difficult to correctly identify important sentences.
Verma et al., 2025	• GAI-MACSeqRT model	It provides meaningful summaries while maintaining a lower computational cost	The performance still needs enhancement since the attained precision and F1 score values are very low.
Deo et al., 2025	• CLSTM-MDS	It provided an accurate and high-quality summary	This IBi-GRU might increase the model complexity and training time
Phani et al., 2025	• MATSFT	More robust in handling various datasets with diverse linguistic complexities.	However, the performance can still be optimized via advanced ML models

2.2 Research Gaps

The official language of the policies, together with their perplexing semantics, makes it annoying for social media users, who then agree to follow these policies without no understanding. Companies also hold users accountable for any abuse of data if users have consented to the privacy policies. A smaller number of users refer to the policy since certain programs make it obligatory for the user to review it. Although data's immense power might provide insights, its massive numbers can also provide challenges. A policy summarizer that displays the different properties covered by the policy is necessary to help users understand the policies. However, despite the fact that there are currently some methods for summarizing information, it is getting more difficult to produce accurate and timely summaries due to the growing amount of information available online. Although it might make the article difficult to read, the concatenation technique for extraction does little to guarantee that the summaries are cohesive. An additional difficulty is assessing summarizers. If the summary is believed to be a trustworthy alternative to the original source, it is important to make certain that it accurately captures the pertinent information from the source. Recently, AI models have received more attention in text summarization tasks. Mainly, RNN-LSTM (Abdi et al., 2021) was introduced for multi-document summarization. Although this model can offer higher accuracy, this model cannot differentiate between active and passive sentences. Moreover, TAAM (Alahmadi et al., 2022) and CLSTM-MDS (Deo et al., 2025) were introduced to provide an accurate and high-quality summary; however, they can be computationally expensive and sluggish, particularly when working with long sequences. Furthermore, ROUGE score outperformed classical models. Cross-lingual transfer learning (Wijayanti et al., 2023) and Hybrid ATS (Jiang et al., 2021) models have provided a superior ROUGE score. But these approaches’ application is limited, and the utilization of network tiers is limited. Furthermore, with CNN and FFNN, a hybrid model was introduced by Wazery et al. (2023); although it provides the highest ROUGE score, it only utilized three statistical features. Another DL-based text summarization was proposed by Bhuyan et al. (2023), which performed better in providing more accurate summaries. However, comparatively simpler and more compact designs were developed on certain epochs due to resource constraints. Moreover, GAI-MACSeqRT (Verma et al., 2025) and POCASUM (Deotale et al., 2021) models provided precise summaries, but they still need enhancement in scalability and performance. Furthermore, MATSFT (Phani et al., 2025) was more robust in handling various datasets with diverse linguistic complexities. However, the performance can still be optimized via advanced ML models. While many reviewed models offer valuable insights, their limitations, such as scalability, semantic depth, and computational cost, are directly addressed by our proposed HC. To tackle such challenges, a new model needs to be introduced. By combining TL-based modified LSTM and Bi-GRU with advanced feature extraction and clustering, our model offers a comprehensive solution for policy categorization and summarization. As a consequence, a HC-based policy categorization and text summarization model has been developed, which is briefed below.

3. HC-Based Policy Categorization and Summarization Model

To help social media users comprehend the policies via reading short paragraphs or summaries and make them aware of the complete regulations of these privacy policies, we have introduced an HC-based policy categorization and text summarization model, which saves time and effort of the users compared to looking through the full policies. This HC-based policy categorization and text summarization model includes four phases, which are policy categorization, pre-processing, feature extraction, and summarization, that are explained below.

Policy categorization: It is the initial phase, where the policy's each paragraph is labeled as per the categories such as collection, information usage, location, share, and contact using the data annotation process, where an EDCos membership-based FCM algorithm is used for accurate policy categorization.

Preprocessing: After categorization, two processes, namely, tokenization and stemming, are performed in this phase.

Feature extraction: Next to this, the features, including SDC-based TF-IDF, thematic features, and BOW, are extracted from the pre-processed text.

Summarization: Finally, based on these extracted features, the texts are summarized accurately using a HC that combines both TL-based modified LSTM and Bi-GRU models. The overall architecture of the hybrid DL-based policy categorization and text summarization model is given in Figure 1.

Figure 1.

Architecture of HC-based policy categorization and text summarization model.

3.1 Policy Categorization

The first step is the classification of policies. In the policy categorization phase, the data annotation process is conducted in the text using an EDCos membership-based FCM clustering. Initially, the input text is extracted from the dataset. Afterwards, it is subjected to the data annotation process, where EDCos membership-based FCM manually labels each distinct paragraph in the text using five distinct traits or categories. When the data annotation process is conducted on the APP-350 dataset, five distinct traits or categories are utilized. These categories (labels) are:

Collection: The type of personal data that the company collects from the user.

Info usage: How is the data being utilized by the company.

Location: How is the location information being used.

Share: Under what circumstances will the user's data be disclosed or shared.

Contact: Do they provide any contact information about the company.

Similarly, in the case of the BillSum dataset, the same five categories: collection, information usage, location, share, and contact, are utilized.

3.1.1. Data Annotation Using EDCos Membership-based FCM

The data annotation process will take place via the EDCos membership-based FCM algorithm to categorize the policy accurately.

Generally, a clustering technique called FCM (Zhu et al., 2019) permits a single data point to be a part of multiple clusters. The FCM is predicated on minimizing Equation (1).

\begin{aligned} J_{N} = \sum_{i = 1}^{M} \sum_{j = 1}^{D} {(U_{i j})}^{N} \cdot {‖ X_{i} - d_{j} ‖}^{2} \end{aligned}

(1)

where, N refers to real integers above 1,

U_{i j}

refers to the membership degree of

X_{i}

j^{t h}

cluster,

X_{i}

refers to

i^{t h}

D i

-dimension data and

i = 1, 2, \dots ., M

. Also,

d_{j}

refers to the cluster center, and

‖ * ‖

refers to any norm articulating the resemblance between any center and the computed data.

Fuzzy partitioning is accomplished by iteratively optimizing the above-mentioned objective and updating the membership $U_{i j}$ and $d_{j}$ as indicated by Equation (2).

\begin{aligned} U_{i j} & = \frac{1}{\sum_{K = 1}^{D} {(‖ X_{i} - d_{j} ‖ / ‖ X_{i} - d_{K} ‖)}^{2 / (N - 1)}} \end{aligned}

(2)

\begin{aligned} d_{j} & = \sum_{i = 1}^{M} (U_{i, j}^{N} . X_{i}) / \sum_{i = 1}^{M} U_{i, j}^{N} \end{aligned}

(3)

The iteration ends if $max_{i j} {U_{i j} (K + 1) - U_{i j} (K)} < ε$ , wherein $ε$ refers to the termination criterion between 0 and 1, while K refers to the iteration stages. This process converges to a saddle point or local minimum of $J_{N}$ .

Even though existing FCM outcomes are better in clustering, the Euclidean distance considers only the degree of variation among the vectors and ignores the directional association, which leads to an inaccurate categorization. Moreover, Euclidean distance turns out to be less meaningful in high-dimensional space due to the data sparsity. To overcome these issues, we have introduced a new EDCos membership-based FCM clustering, which involves a modification in the computation of the membership degree.

As per EDCos membership-based FCM clustering theory, the conventional degree of membership is enhanced using a new function as shown in Equation (4). This proposed degree of membership is computed based on the Euclidean distance $(E D)$ , weight factor $(W)$ , and cosine distance $(C D)$ . The Euclidean distance $(E D)$ and cosine distance $(C D)$ are computed based on $X_{i}$ and $d_{j}$ as shown in Equations (5) and (6).

\begin{aligned} E D Cos p U_{i j} & = \frac{1}{\sum_{K = 1}^{D} {[0.5 * W * (E D + C D) / ‖ X_{i} - d_{K} ‖]}^{2 / (N - 1)}} \end{aligned}

(4)

(5)

\begin{aligned} C D & = 1 - \frac{X . d_{j}}{{‖ X ‖}_{2} {‖ d_{j} ‖}_{2}} \end{aligned}

(6)

In Equation (5),

w g

denotes the weights for every value in X and d array.

The weight factor $(W)$ is evaluated based on a random integer as shown in Equation (7); here, $β = 0.7$ and $w_{1}$ is a random number between 0 and 1.

\begin{aligned} W = \sin (β π / w_{1}) \end{aligned}

(7)

Thus, based on the EDCos membership $(E D C o s p U_{i j})$ , the objective function is modelled as given in Equation (8).

\begin{aligned} J_{N} = \sum_{i = 1}^{M} \sum_{j = 1}^{D} {(E D Cos p U_{i j})}^{N} \cdot {‖ X_{i} - d_{j} ‖}^{2} \end{aligned}

(8)

This EDCos membership-based FCM clustering approach utilizes the enhanced membership degree, which captures the angular similarities among vectors using cosine distance, thus focusing on the directions along with the magnitude. Thus, considering both Euclidean distance and cosine distance, the information about directions as well as magnitude can be captured. Moreover, the EDCos membership-based FCM clustering approach is found to be more effective for diverse datasets. We have compared the membership function based on Euclidean distance and cosine distance with our EDCos membership function, which is shown in Figure 2. For different data points, a stable distance is attained by the EDCos Membership function, which proves the effectiveness of EDCos membership-based FCM clustering.

Figure 2.

Comparison of membership function for diverse data points.

After applying Improved FCM, five clusters will be formed, each containing five attributes or categories on collection, info usage, location, share, and contact. Each individual cluster is then subjected to a preprocessing phase.

3.2. Pre-processing

After categorization, the labeled text data will undergo preprocessing. Preprocessing is done using diverse processes so as to make the dataset understandable by an ML or DL approach.

Here, tokenization and stemming processes are adopted to carry out pre-processing.

Tokenization (Mohamed, 2011): It is the method of reducing the provided text to the lowest sentence-level unit in NLP, known as a token.

Stemming (Mohamed, 2011): By eliminating the word's affixes, stemming creates base words from derived terms. The preprocessed texts are denoted as $T x^{P}$ .

3.3. Feature Extraction

The feature extraction algorithms are deployed to find the key sentences in the text $T x^{P}$ . Finding significant features or characteristics in the data is a process known as feature extraction. Among a vast collection of texts, this approach may be used to recognize patterns and discover common topics. Here, we extract features like:

SDC-based TF-IDF

BoW

Thematic features

3.3.1. SDC-based TF-IDF features

A statistical technique called TF-IDF is used to evaluate a word's significance within a particular list or corpus (Roul et al., 2017). Here, the pre-processed text $T x^{P}$ is provided as input to extract TF-IDF features.

TF: It shows how frequently a phrase occurs in a text. The documents differ in length; lengthier documents often include more word occurrences than those that are shorter. The word frequency is further split by the text's length to standardize this. Equation (9) expresses the TF mathematical representation, where T denotes the term.

\begin{aligned} T F_{T, d m n t} = \frac{O c c u r e n c e c o u n t o f T \in d m n t}{E n t i r e l e n g t h o f d m n t} \end{aligned}

(9)

IDF: It functions as an evaluation of a term's rarity in a particular document. IDF assigns a word's importance according to how uncommon it is; words that are less common are given higher IDF ratings. Equation (10) expresses the IDF mathematical representation. Here, p denotes the entire corpus.

\begin{aligned} I D F_{T} = \log_{10} (\frac{D o c u m e n t c o u n t \in p}{D o c u m e n t c o u n t o f p c o n t a i n i n g T}) \end{aligned}

(10)

Consequently, using both TF and IDF, the TF-IDF is computed and is shown in Equation (11), where $T F_{t, d}^{t}$ refers to the frequency of T within the document $d m n t$ .

\begin{aligned} F^{T F - I D F_{T, d m n t}} = T F_{T, d m n t} * I D F_{T} \end{aligned}

(11)

Although TF-IDF computes the significance of terms T within the document $d m n t$ , it suffers from certain drawbacks.

The term weight is determined by using the TF. It doesn’t consider a term's meaning, location within the text, or overlap with other words in a document. Although the document similarities are calculated explicitly in the vector space, a big corpus may cause it to run slowly. Also, it deceives the classifier during classification since it lacks the capacity to disperse across classes. Also, it could decrease the IDF's ability to discriminate.

Thereby, we developed an SDC-based TF-IDF model, which depends upon SDC and $T F_{max} (d m n t)$ as shown in Equation (12). $T F_{max} (d m n t)$ refers to the maximal TF in the document $d m n t$

\begin{aligned} F^{S D C - T F - I D F} = \frac{T F_{T, d m n t} * I D F_{T}}{T F_{max} (d m n t) + 1} * S D C \end{aligned}

(12)

The SDC term in Equation (12) is computed based on $I D$ and $S I D F$ as in Equation (13),

\begin{aligned} S D C = I D * S I D F \end{aligned}

(13)

The $I D$ parameter in Equation (20) is evaluated as in Equation (14), where $f (T, j)$ refers to the count of $d m n t$ having the term T. Here, m is the number of classes

\begin{aligned} I D & = \frac{1}{m} \sum_{j = 0}^{m} {[f (T, j) - a v g (f (T, j))]}^{2} \end{aligned}

(14)

\begin{aligned} S I D F & = \log (\frac{1 + E n t i r e d m n t i n c o r p u s p}{1 + c o u n t o f d m n t c o n t a i n i n g T}) \end{aligned}

(15)

On deploying the SDC-based TF-IDF, the TF-IDF scores can be normalized. This accounts for the length of the document and prevents long documents from suffering undue influence. It offers a balanced representation of features and ensures the term significance is not skewed by the length of the document. Moreover, the smoothing factor also avoids extreme IDF values by modifying the computation. This leads to more reliable and stable TF-IDF scores. To prove the SDC-based TF-IDF model's reliability, we have compared its TF-IDF score with the conventional TF-IDF model's score, and the outcome is shown in Figure 3. The conventional TF-IDF model attained a maximum TF-IDF score of 0.5, while the SDC-based TF-IDF model attained a score of 0.44.

Figure 3.

Comparison of SDC-based TF-IDF scores with conventional TF-IDF scores for different terms.

3.3.2 BoW features

A method for extracting attributes from pre-processed text $T x^{P}$ is called BoW (Rupapara et al., 2021). Due to its simplicity, it is the best option for text categorization and language modeling as it is simple to use and comprehend. BoW's efficacy at extracting features from textual data makes it extremely valuable and appropriate in text summarization. This method involves turning words into numerical numbers to represent them within a document. Word frequencies are quantified by BoW as textual attributes, which are then combined into a “bag.” Word locations, structure, and syntax are not crucial for BoW; instead, it uses a lexicon containing identified terms and extracts features based on their occurrences. Every word has equal weight and functions as a feature in BoW. In BoW, numbers 1 and 0, respectively, denote the existence and lack of a feature. The BoW features are represented as, $F^{B o W}$ .

3.3.3. Thematic features

A term that appears commonly in a document and is most likely connected to the topic is referred to as a thematic word (Dixit & Apte, 2012). A theme word is one that has the highest degree of relativism. The occurrence frequency of each word in $T x^{P}$ was calculated, and the highest five frequent terms were utilized to determine the theme. The ratio of the number of theme words that appear in a sentence $(S)$ over the maximal amount of thematic words in a sentence is used to determine the score for this attribute.

\begin{aligned} F^{T W} = \frac{T h e m a t i c w o r d c o u n t i n S}{M a x i m a l c o u n t o f T h e m a t i c w o r d} \end{aligned}

(16)

The features derived for text summarization are denoted by F. F includes the features namely, BoW feature, thematic feature, and SDC-based TF-IDF features, that is, $F = F^{S D C - T F - I D F} + F^{B o W} + F^{T W}$ .

The extracted features then undergo the text summarization process.

3.4. Hybrid Model: Transfer Learning-Based Modified LSTM and Bi-GRU for Text Summarization

The derived features denoted by F are then subjected to the text summarization phase, where the texts are summarized using a HC that combines both TL-based modified LSTM and Bi-GRU models. By combining diverse classifiers, improved summarization performance can be obtained. Moreover, overfitting issues can be lessened, and model generalization is increased. Here, initially, modified LSTM is pre-trained on the whole dataset and a pre-trained weight is attained. While testing the modified LSTM, the weight that is pre-trained is applied to the modified LSTM instead of starting the learning procedure from scratch with arbitrary weight initialization. Here, both TL-based modified LSTM and Bi-GRU models are fused as a single model, and the summarized texts are attained as shown in Figure 4.

Figure 4.

Proposed text summarization model combining TL-based modified LSTM and Bi-GRU.

3.4.1 Bi-GRU

Bi-GRU (Wang et al., 2022) is an architecture that uses neural networks in GRUs. GRUs are a kind of RNN that are particularly good at identifying sequential relationships in data. Two gates are needed for a typical GRU cell: “an update gate and a reset gate,” as shown in Equations (17) and (18), respectively.

\begin{aligned} v_{t} & = α (W_{v} . [H_{t - 1}, G_{t}^{f}]) \end{aligned}

(17)

\begin{aligned} R_{t} & = α (W_{R} . [H_{t - 1}, G_{t}^{f}]) \end{aligned}

(18)

In Equation (18), W points out the weight, $H_{t - 1}$ points out the hidden state of the prior time stamp, $R_{t}$ points out the reset gate, $v_{t}$ points out the update gate, and $α$ points out the sigmoid function.

In the Bi-GRU design, the gates control the information flow within the cell. The reset gate establishes the degree to which prior information should be disregarded, whereas the update gate controls the proportion of old data that should be retained in relation to the new.

The candidate activation ${\tilde{H}}_{t}$ and the updated hidden state $H_{t}$ are assessed in Equations (19) and (20), wherein $⊙$ implies element-wise multiplication.

\begin{aligned} {\tilde{H}}_{t} & = \tanh (W_{H} . [R_{t} ⊙ H_{t - 1}, G_{t}^{f}]) \end{aligned}

(19)

\begin{aligned} H_{t} & = (1 - v_{t}) ⊙ H_{t - 1} + v_{t} ⊙ {\tilde{H}}_{t} \end{aligned}

(20)

Both backward and forward direction analysis of input sequences is performed by bidirectional RNNs, like bidirectional GRUs. Furthermore, by gathering data from both past and future contexts at the same time, this bi-directional processing helps the network better understand and model complex temporal connections.

3.4.2 TL-based modified LSTM

TL, a technique in ML, involves creating and training a model for a specific case and then deploying it again for another similar task. The modified LSTM architecture is utilized here, whose pre-trained weights are applied to examine another set of data. Hence, TL enables the process to deploy the learnt features from the training dataset. Figure 5 depicts the TL function in the modified LSTM.

Figure 5.

Framework showing TL function in LSTM.

An essential part of text summarization is played by the LSTM (Xiao et al., 2020) classifier. The LSTM classifier operates by receiving features F that are extracted from texts. The LSTM architecture is a kind of RNN designed to handle data sequences with long-range relationships. The information that flows through the memory cells is controlled by the three gates in the LSTM architecture. The forget gate, which is described by Equation (21), controls how much of the past is remembered. The forgetting factor $(f r_{t})$ is computed using the softmax function $(β)$ based on the previous hidden state $(h_{t - 1})$ , and bias term $(A_{f r})$ . The input gate $(i r_{t})$ and an alternative vector $({\tilde{c}}_{t})$ is computed using Equations (22) and (23), where $β$ denotes the softmax function and $ϖ$ denotes the weight. These gates make a decision on how much new data is added, as in Equation (24). The output gate $(o r_{t})$ , shown in Equation (25), calculates an output based on $(h_{t - 1})$ and bias term $(A_{o r})$ . The final output $(h_{t})$ is determined by the output factor $(o r_{t})$ multiplied by the tanh of $(c_{t})$ as in Equation (26).

\begin{aligned} f r_{t} & = β (ϖ_{f r} F + ϖ_{f r} h_{t - 1} + A_{f r}) \end{aligned}

(21)

\begin{aligned} i r_{t} & = β (ϖ_{i r} F + ϖ_{i r} h_{t - 1} + A_{i r}) \end{aligned}

(22)

\begin{aligned} {\tilde{c}}_{t} & = \tanh (ϖ_{c} F + ϖ_{c} h_{t - 1} + A_{c}) \end{aligned}

(23)

\begin{aligned} c_{t} & = f r_{t} \times c_{t - 1} + i r_{t} \times {\tilde{c}}_{t} \end{aligned}

(23)

\begin{aligned} o r_{t} & = β (ϖ_{o r} F + ϖ_{o r} h_{t - 1} + A_{o r}) \end{aligned}

(25)

\begin{aligned} h_{t} & = o r_{t} \times \tanh (c_{t}) \end{aligned}

(26)

The softmax activation function used in LSTM is shown in Equation (27).

\begin{aligned} β = \frac{\exp (F_{I})}{\sum_{j} \exp (F_{j})} \end{aligned}

(27)

Although softmax activation is effective in converting logits into probabilities, it struggles to capture complicated non-linear associations, especially in texts where semantic connotation could be subtle. Moreover, the softmax activation function used in LSTM could be susceptible to extreme values or outliers, since it has a tendency to intensify them. This might result in skewed results. It could cause overemphasis on specific phrases or words in the created summary.

Thereby, a new activation function is proposed that replaces the existing softmax activation function. The modified activation function includes three functions, namely, softmax $(β)$ , modified comb-H-sine $(m c o m b H \sin e)$ and hard Tanh activation functions. The proposed activation function is shown in Equation (28). The modified comb-H-sine function is given in Equation (29), where $δ = 0.3$ .

\begin{aligned} M o d i f i e d a c t i v a t i o n = {\begin{cases} - 1, i f F < - 1 \\ m c o m b - H - \sin e (F), i f - 1 \leq F \leq 1 \\ 1, I F F > 1 \end{cases} \end{aligned}

(28)

\begin{aligned} m c o m b - H - \sin e (F) = s i n h (δ * β) + \sin^{- 1} h (δ * β) \end{aligned}

(29)

The comb-H-sine activation function is modelled to capture the non-linearities much more efficiently by merging the arc sine-H function and the hyperbolic sine function. This could aid in captivating better-complicated associations within the text, which enhances the understanding of models containing nuanced data. Moreover, the modified activation is compared with activation functions such as hard tanh, comb-h-sine, and Softmax, and its graphical representation is in Figure 6.

Figure 6.

Comparison between conventional and modified activation functions.

The modified activation function minimizes the risk of skewed summaries by minimizing the impact of extreme values or outliers.

The LSTM classifier summarizes the texts based on the information taken from the input sequences and the learnt temporal patterns. Both TL-based modified LSTM and Bi-GRU models are fused in a single model, and the summarized texts are finally attained. Figure 7 shows the architecture of the proposed TL-based modified LSTM.

Figure 7.

Architecture of proposed TL-based modified LSTM.

4. Results and Discussion

4.1 Simulation Procedure

The introduced approach for policy categorization and text summarization using HC (TL-based modified LSTM + Bi-GRU) was executed in Python. We have conducted the experiments on two datasets such as APP-350 and BillSum datasets, to prove the HC (TL-based modified LSTM + Bi-GRU) model's superior performance. The evaluation was done to show the better performance of HC (TL-based modified LSTM + Bi-GRU) over LSTM, Bi-GRU, CNN, RNN, NN, VecMap, and BiVec (Wijayanti et al., 2023), CNN + FFNN (Wazery et al., 2023), and POCASUM (Deotale et al., 2021).

4.2. Dataset Description

We have utilized two datasets, namely APP-350 dataset (https://usableprivacy.org/data; Zimmeck et al., 2019) and BillSum dataset (https://www.kaggle.com/datasets/akornilo/billsum) to validate the hybrid DL or HC-based policy categorization and text summarization model, and both datasets are described below.

4.2.1. Description of Dataset 1 (APP-350 dataset)

The input data was taken from the APP-350 dataset (https://usableprivacy.org/data; Zimmeck et al., 2019). The number of data used is 4085, and the number of classes is 5. The class labels are: 0—Collection, 1—Info usage, 2—Location, 3—Share, 4—Contact. The amount of data in each class label is: 0: 1459, 1: 783, 2: 539, 4: 646, 3: 658. The APP-350 Corpus consists of 350 Android app privacy policies annotated with privacy practices (i.e., behavior that can have privacy implications). This is a binary classification task in which the LLM is provided with a clause from a privacy policy, and a description of that clause (e.g., “The policy describes the collection of the user's HTTP cookies, flash cookies, pixel tags, or similar identifiers by a party to the contract.”). The LLM must determine if the description of the clause is correct or incorrect. The dataset is made available for research, teaching, and scholarship purposes only, with further parameters in the spirit of a Creative Commons Attribution-Non-Commercial License. Table 2 shows the training and testing data.

Table 2.
Training and Testing Data.

Training Data (%) Training Data Testing Data

Dataset 1

60 2448 1637

70 2858 1227

80 3266 819

90 3675 410

Dataset 2

60 11369 7580

70 13264 5685

80 15159 3790

90 17054 1895

Training Data (%)	Training Data	Testing Data
Dataset 1
60	2448	1637
70	2858	1227
80	3266	819
90	3675	410
Dataset 2
60	11369	7580
70	13264	5685
80	15159	3790
90	17054	1895

The samples showing input text, actual, and predicted summaries are shown in Table 3.

Table 3.

Samples Showing Input Text, Actual and Predicted Summaries Using HC (TL-based Modified LSTM + Bi-GRU) From Dataset-1.

Input Text	Actual Summary	Predicted Summary
“PERSONAL DATA, INCLUDING YOUR E-MAIL ADDRESS 2.7 You may choose to register with us for a better game experience and/or to create a public or semi-public profile for participating in our social gaming network. This includes, for example: overviews of high scores, achievements listings, and individual game recommendations on our products. When you register with us, you provide us with certain personal information, such as a username, date and month of birth, and your email address. The information you disclose in that respect may also become publicly accessible. We provide you with privacy settings that restrict other users from accessing your information; namely, our network architecture and your privacy settings allow you to make informed choices about who has access to your information. We allow you to choose the information you provide to friends and networks through our products. When your privacy settings permit it, other users may be able to see and supplement your profile (e.g., by posting comments).”	Contact E-mail Address 1stParty Demographic Age 1^st Party'	phone contact email address 1^st party postal contact age 1^st party
Location: If you utilize the Find My Phone feature, the Software may access your phone's location information to assist in locating your phone. We do not utilize your location information for any other purpose.	Location 1^st Party'	phone location 1^st party
Device information. We may collect device-specific information, including your operating system version, system language, and IMEI number.	Identifier_IMEI_1^st Party'	1^st party contact identifier imei 1^st party

4.2.2 Description of Dataset 2 (BillSum)

This BillSum dataset (https://www.kaggle.com/datasets/akornilo/billsum) is a corpus for United States (US) Legislation's automatic summarization. This dataset includes three parts, which were US training bills, California test bills, and US test bills. From the Govinfo offered by the US Government Publishing Office (GPO), these US bills were collected. Moreover, this corpus contains the bills from the 103rd–115th sessions of Congress from 1993–2016. Randomly, this data was divided into 5014 test bills and 28,408 train bills. In the case of California, the 2015–2016 sessions were directly scraped from the legislature's website, and by their Legislative Counsel, the summaries were written. The number of classes utilized is 5, which were 0—Collection, 1—Info usage, 2—Location, 3—Share, 4—Contact. From this dataset, we have utilized the following sample as input, and their actual and predicted summaries are shown in Table 4.

Table 4.
Samples Showing Input Text, Actual and Predicted Summaries Using HC (TL-based Modified LSTM + Bi-GRU) From Dataset-2.

Input Text Actual Summary Predicted Summary

“Section 1. This Act may be cited as the ``Border Hospital Survival and Illegal Immigrant Care Act'‘. Immigration is a federal responsibility. The Immigration and Naturalization Service does not take into custody all aliens who are unlawfully present in the United States.” “Border Hospital Survival and Illegal Immigrant Care Act - Amends the Public Health Service Act to

establish a five-year pilot program of health care provider reimbursement for the costs associated with providing emergency medical and ambulance services in Arizona.” The Public Health Service Act to the Secretary of Health to establish a five-year program of health care provider reimbursement for the costs with emergency medical and ambulance services in Arizona

“Section 1. This Act may be cited as the Seniors’ Retirement Recovery Act of 2002. Social Security benefits in an amount equal to the lesser of one-half of the Social Security benefits received during the taxable year.” “Seniors’ Retirement Recovery Act of 2002 - Amends the Internal Revenue Code to repeal the 1993 income tax increase on Social Security benefits. Distributions must begin from 70 12 to 80.” “Seniors Retirement Recovery Act of 2002 - Amends the Internal Revenue Code to the 1993 income tax increase on Social Security.”

Nothing in the amendments made by this section shall be construed to affect Government obligations mandated by other law, including obligations with respect to Social Security, Medicare, and Medicaid. “Government Shutdown Prevention Act - Amends Federal law to provide continuing appropriations in the absence of regular appropriations.” “Government Shutdown Prevention Act - Amends Federal law to provide for continuing appropriations in the absence of regular appropriations.”

Input Text	Actual Summary	Predicted Summary
“Section 1. This Act may be cited as the ``Border Hospital Survival and Illegal Immigrant Care Act'‘. Immigration is a federal responsibility. The Immigration and Naturalization Service does not take into custody all aliens who are unlawfully present in the United States.”	“Border Hospital Survival and Illegal Immigrant Care Act - Amends the Public Health Service Act to
establish a five-year pilot program of health care provider reimbursement for the costs associated with providing emergency medical and ambulance services in Arizona.”	The Public Health Service Act to the Secretary of Health to establish a five-year program of health care provider reimbursement for the costs with emergency medical and ambulance services in Arizona
“Section 1. This Act may be cited as the Seniors’ Retirement Recovery Act of 2002. Social Security benefits in an amount equal to the lesser of one-half of the Social Security benefits received during the taxable year.”	“Seniors’ Retirement Recovery Act of 2002 - Amends the Internal Revenue Code to repeal the 1993 income tax increase on Social Security benefits. Distributions must begin from 70 12 to 80.”	“Seniors Retirement Recovery Act of 2002 - Amends the Internal Revenue Code to the 1993 income tax increase on Social Security.”
Nothing in the amendments made by this section shall be construed to affect Government obligations mandated by other law, including obligations with respect to Social Security, Medicare, and Medicaid.	“Government Shutdown Prevention Act - Amends Federal law to provide continuing appropriations in the absence of regular appropriations.”	“Government Shutdown Prevention Act - Amends Federal law to provide for continuing appropriations in the absence of regular appropriations.”

4.3. Analysis of Categorization Performance

We have validated the EDCos membership-based FCM clustering algorithm's categorization performance in terms of ROUGE, precision, accuracy, and recall via conducting an ablation study on dataset 1 and dataset 2, and the outcomes are tabulated in Tables 5 and 6. This analysis includes scenarios like (i) HC (TL-based modified LSTM + Bi-GRU) with conventional balanced iterative reducing and clustering using hierarchies (BIRCH), (ii) HC (TL-based modified LSTM + Bi-GRU) with k-means clustering, (iii) HC (TL-based modified LSTM + Bi-GRU) with conventional FCM clustering.

Table 5.
Performance Comparison Between the Developed HC With EDCos Membership-Based FCM and the HC With Other Clustering Algorithms for Dataset-1.

HC with HC with HC with HC with EDCos Membership-

Metrics Conventional BIRCH Conventional K-Means Conventional FCM Based FCM Clustering

Accuracy 0.869 0.871 0.879 0.927

ROUGE 0.732 0.750 0.755 0.794

Precision 0.692 0.706 0.707 0.750

Recall 0.803 0.814 0.817 0.872

F1-score 0.743 0.756 0.768 0.820

	HC with	HC with	HC with	HC with EDCos Membership-
Accuracy	0.869	0.871	0.879	0.927
ROUGE	0.732	0.750	0.755	0.794
Precision	0.692	0.706	0.707	0.750
Recall	0.803	0.814	0.817	0.872
F1-score	0.743	0.756	0.768	0.820

Table 6.

Performance Comparison Between the Developed HC With EDCos Membership-Based FCM and the HC With Other Clustering Algorithms for Dataset-2.

	HC with	HC with	HC with	HC with EDCos Membership-
Metrics	Conventional BIRCH	Conventional K-Means	Conventional FCM	Based FCM Clustering
Accuracy	0.860	0.879	0.876	0.932
ROUGE	0.742	0.757	0.759	0.792
Precision	0.709	0.719	0.720	0.769
Recall	0.817	0.829	0.835	0.864
F1-score	0.759	0.770	0.773	0.814

When the conventional BIRCH technique is utilized for categorization within the HC (TL-based modified LSTM + Bi-GRU), instead of our EDCos membership-based FCM clustering algorithm, only 86% accuracy was attained on dataset-1. At the same time, HC (TL-based modified LSTM + Bi-GRU) with EDCos membership-based FCM clustering algorithm attained a superior accuracy of 92.7%. Moreover, HC (TL-based modified LSTM + Bi-GRU) with k-means clustering (instead of using EDCos membership-based FCM clustering) algorithm attained the precision, ROUGE, and recall ratings of 0.706, 0.750, and 0.814, while HC (TL-based modified LSTM + Bi-GRU) with EDCos membership-based FCM clustering attained 0.750, 0.794, and 0.872. In terms of F1-score, HC (TL-based modified LSTM + Bi-GRU) with conventional FCM obtained 0.768, while HC (TL-based modified LSTM + Bi-GRU) with EDCos membership-based FCM attained 0.820, which was higher than other clustering algorithms.

Not only for dataset-1 but also for dataset-2 HC (TL-based modified LSTM + Bi-GRU) with EDCos membership-based FCM attained superior accuracy, ROUGE, precision, recall, and F1 score ratings of 0.932, 0.792, 0.769, 0.864, and 0.814, which proved the effectiveness and suitability of EDCos membership-based FCM clustering algorithm in the HC (TL-based modified LSTM + Bi-GRU) based policy categorization and text summarization model.

4.4. Analysis of Summarization Performance

We have validated the text summarization performance of the HC (TL-based modified LSTM + Bi-GRU) model with three analyses, which were analysis on performance, performance test analysis, and ablation study, and statistical analysis on dataset-1 and dataset-2, that are briefed below. In addition to that, an analysis of p-test and t-test, along with statistical analysis, was also conducted for dataset-1 and dataset-2 to prove the effectiveness and accuracy of HC (TL-based modified LSTM + Bi-GRU) based policy categorization and text summarization model.

4.4.1 Comparative Analysis on Classifiers

We have validated the HC (TL-based modified LSTM + Bi-GRU) model's overall performance by comparing its performance with other models in terms of accuracy, ROUGE, F1-score, and precision, along with recall for diverse training data (TD) on two datasets. For comparison, we have utilized extant models like BERTSUM, PEGASUS, LSTM, Bi-GRU, CNN, RNN, NN, VecMap, and BiVec (Wijayanti et al., 2023), CNN + FFNN (Wazery et al., 2023), and POCASUM (Deotale et al., 2021); also the outcomes were provided in Figure 8 (for dataset-1) and Figure 9 (for dataset-2). The obtained accuracy should be better for effective text summarization. On examining the graphs in Figure 8 (for dataset-1), it is known that better performance has been obtained by the proposed HC (TL-based modified LSTM + Bi-GRU). Mainly, as shown in Figure 8(a), the accuracy increased with a raise in TDs. When TD is 60%, HC (TL-based modified LSTM + Bi-GRU) attained an accuracy of 0.914, while when TD is 90%, HC (TL-based modified LSTM + Bi-GRU) attained an accuracy of 0.952. For data with 90%, all the methods, including the proposed and compared methods, attained better accuracy than at the initial TDs. When TD = 90%, HC (TL-based modified LSTM + Bi-GRU) attained an accuracy of 0. 956, while BERTSUM, PEGASUS, LSTM, Bi-GRU, CNN, RNN, NN, VecMap, and BiVec (Wijayanti et al., 2023), CNN + FFNN (Wazery et al., 2023), and POCASUM (Deotale et al., 2021), attained the accuracy of 0.907, 0.912, 0.895, 0.886, 0.867, 0.894, 0.892, 0.881, 0.891, and 0.885, respectively. When TD was 70%, models such as BERTSUM, PEGASUS, LSTM, Bi-GRU, CNN, RNN, NN, VecMap, and BiVec (Wijayanti et al., 2023), CNN + FFNN (Wazery et al., 2023), and POCASUM (Deotale et al., 2021), attained the ROUGE values of 0.761, 0.773, 0.752, 0.766, 0.735, 0.759,0.764, 0.752, 0.737, and 0.743, while HC (TL-based modified LSTM + Bi-GRU) attained a higher value of 0.794, which was shown in Figure 8(b). Similarly, for 60%, 80%, and 90% TDs, the developed HC (TL-based modified LSTM + Bi-GRU) also achieved the superior ROUGE ratings of 0.783, 0.806, and 0.817, respectively. Similarly, in the case of F1-score also we have also conducted a performance comparison for diverse DT on dataset-1, and the result was shown in Figure 8(c). For 80% TD, HC (TL-based modified LSTM + Bi-GRU) achieved the F1-score of 0.860, while BERTSUM, PEGASUS, LSTM, Bi-GRU, CNN, RNN, NN, VecMap, and BiVec (Wijayanti et al., 2023), CNN + FFNN (Wazery et al., 2023), and POCASUM (Deotale et al., 2021), attained the ratings of 0.801, 0.800, 0.825, 0.816, 0.791, 0.762, 0.788, 0.802, 0.818, and 0.827. For 60%, 80%, and 90%, better F1-score values are attained by the HC (TL-based modified LSTM + Bi-GRU) model. Likewise, in the case of precision (shown in Figure 8(d)), HC (TL-based modified LSTM + Bi-GRU) attained the highest precision of 0.798 when TD is 90%. On the other hand, BERTSUM, PEGASUS, LSTM, Bi-GRU, CNN, RNN, NN, VecMap, and BiVec (Wijayanti et al., 2023), CNN + FFNN (Wazery et al., 2023), and POCASUM (Deotale et al., 2021), attained lower precision of 0.782, 0.789, 0.776, 0.765, 0.778, 0.723, 0.767, 0.772, 0.763, and 0.767, respectively. The better text summarization performance of the developed HC (TL-based modified LSTM + Bi-GRU) was owing to the TL-based modified LSTM and Bi-GRU for text summarization. Moreover, when TD was 60%, BERTSUM, PEGASUS, LSTM, Bi-GRU, CNN, RNN, NN, VecMap, and BiVec (Wijayanti et al., 2023), CNN + FFNN (Wazery et al., 2023), and POCASUM (Deotale et al., 2021) models attained the recall value of 0.814, 0.806, 0.795, 0.781, 0.791, 0.783, 0.786, 0.786, 0.780, 0.787, while the HC (TL-based modified LSTM + Bi-GRU) model attained 0.857, that was greater than other compared models. Not only for 60% but also for 70%, 80%, and 90%, HC (TL-based modified LSTM + Bi-GRU) model obtained higher recall values of 0.875, 0.893, and 0.914, which proved the effectiveness and suitability of the HC (TL-based modified LSTM + Bi-GRU) model in task summarization tasks.

Figure 8.

Performance analysis comparison of HC (TL-based modified LSTM + Bi-GRU) over extant classifiers on dataset-1 (a) Accuracy, (b) ROUGE, (c) F1-score, (d) precision, and (e) Recall.

Figure 9.

Performance analysis comparison of HC (TL-based modified LSTM + Bi-GRU) over extant classifiers on dataset-2.

Similarly, we have analyzed the performance of the HC (TL-based modified LSTM + Bi-GRU) model by comparing it with other models, and the result is provided in Figure 9. From these outcomes, the HC (TL-based modified LSTM + Bi-GRU) model's outstanding performance was proven, and this betterment was due to the proposed SDC-based TF-IDF feature along with the HCs.

4.4.2. Ablation Study

Tables 7 and 8 validate the ablation study on the betterment of the proposed HC (TL-based modified LSTM + Bi-GRU) over the proposed with existing FCM, the proposed with existing TF-IDF, proposed with existing FCM and existing TF-IDF on dataset-1 and dataset-2. Since we have made certain adjustments in the existing FCM and existing TF-IDF, better performance is attained using improved versions. In Table 7, the proposed HC (TL-based modified LSTM + Bi-GRU) with improved FCM and improved TF-IDF shows a high accuracy of 0.927, while the proposed with existing FCM, the proposed with existing TF-IDF, proposed with existing FCM and existing TF-IDF display less accuracy for dataset-1. The proposed improved FCM approach is found to be effective for datasets with varied feature distributions and scales. It mainly focuses on the directions along with the magnitude. Thus, considering both Euclidean distance and cosine distance, the information about directions as well as magnitude can be captured. This is proved by ROUGE, which attained a value of 0.794 using the developed HC (TL-based modified LSTM + Bi-GRU) with improved FCM and improved TF-IDF, while proposed with existing FCM, proposed with existing TF-IDF, and proposed with existing FCM and existing TF-IDF attained lower ROUGE scores of 0.755, 0.761, and 0.755, respectively. Likewise, the recall and F1-score values are also high using the proposed HC (TL-based modified LSTM + Bi-GRU) over existing ones. The smoothing factor in improved TF-IDF features avoids extreme IDF values by modifying the computation. This leads to much more reliable and stable TF-IDF scores than with extant TF-IDF. Similarly, for dataset-2, an ablation study was conducted for diverse scenarios such as proposed with existing FCM, proposed with existing TF-IDF, and proposed with existing FCM and existing TF-IDF, and the outcomes are compared with other HC (TL-based modified LSTM + Bi-GRU) based models, as given in Table 8. From Table 8, it was proven that with dataset-2, HC (TL-based modified LSTM + Bi-GRU) attained better performance. This enhancement was due to the proposed EDCos membership-based FCM algorithm, SDC-based TF-IDF, and the HCs.

Table 7.
Ablation Analysis of HC (TL-based Modified LSTM + Bi-GRU) for Diverse Scenarios for Dataset-1.

Proposed With Proposed With Proposed With Existing HC (TL-Based Modified

Metrics Existing FCM Existing TF-IDF FCM and Existing TF-IDF LSTM + Bi-GRU)

Accuracy 0.879 0.881 0.875 0.927

ROUGE 0.755 0.761 0.755 0.794

Precision 0.707 0.728 0.706 0.750

Recall 0.817 0.822 0.821 0.872

F1-score 0.768 0.790 0.767 0.820

	Proposed With	Proposed With	Proposed With Existing	HC (TL-Based Modified
Accuracy	0.879	0.881	0.875	0.927
ROUGE	0.755	0.761	0.755	0.794
Precision	0.707	0.728	0.706	0.750
Recall	0.817	0.822	0.821	0.872
F1-score	0.768	0.790	0.767	0.820

Table 8.

Ablation Analysis of HC (TL-based Modified LSTM + Bi-GRU) for Diverse Scenarios for Dataset-2.

	Proposed With	Proposed With	Proposed With Existing	HC (TL-Based Modified
Metrics	Existing FCM	Existing TF-IDF	FCM and Existing TF-IDF	LSTM + Bi-GRU)
Accuracy	0.876	0.883	0.868	0.932
ROUGE	0.759	0.765	0.752	0.792
Precision	0.72	0.723	0.707	0.769
Recall	0.835	0.841	0.81	0.864
F1-score	0.773	0.778	0.755	0.814

4.4.3. Statistical Significance T-Test

The main goal of this significance test is to examine how far the performance evaluation outcomes differ from each classifier model. If it was not significant, that means these algorithms’ performance validation was practically the same. The significance test method applied in this work was the T-Test, and its outcomes for dataset 1 and dataset 2 are offered in Tables 9 and 10. A t-test is an inferential statistic used to determine if there is a significant difference between the two algorithms and how they are related. As per the statistics, if the value of T-Test was more than 0.05, that means there is no significance among the two compared models. We have compared the HC (TL-based modified LSTM + Bi-GRU) model with other extant models. From Tables 9 and 10, it was proven that between the BERTSUM, PEGASUS, LSTM, Bi-GRU, CNN, RNN, NN, VecMap, and BiVec (Wijayanti et al., 2023), CNN + FFNN (Wazery et al., 2023), and POCASUM (Deotale et al., 2021) models, the performance evaluation is not different since the outcomes are not significant.

Table 9.
The Significance Test Outcomes Using T-Test for Analysis of HC (TL-based Modified LSTM + Bi-GRU) Over Other Extant Classifiers for Dataset-1.

HC (TL-Based Modified LSTM + Bi-GRU) Vs T-Test p-Value

BERTSUM .007

PEGASUS .018

LSTM .003

Bi-GRU .002

CNN .000

RNN .005

NN .002

VecMap and BiVec (Wijayanti et al., 2023) .001

CNN + FFNN (Wazery et al., 2023) .003

POCASUM (Deotale et al., 2021) .002

HC (TL-Based Modified LSTM + Bi-GRU) Vs	T-Test p-Value
BERTSUM	.007
PEGASUS	.018
LSTM	.003
Bi-GRU	.002
CNN	.000
RNN	.005
NN	.002
VecMap and BiVec (Wijayanti et al., 2023)	.001
CNN + FFNN (Wazery et al., 2023)	.003
POCASUM (Deotale et al., 2021)	.002

Table 10.

The Significance Test Outcomes Using T-Test for Analysis of HC (TL-based Modified LSTM + Bi-GRU) Over Other Extant Classifiers for Dataset-1.

HC (TL-Based Modified LSTM + Bi-GRU) Vs	T-Test p-Value
BERTSUM	.001
PEGASUS	.006
LSTM	.001
Bi-GRU	4.46×10⁻⁰⁵
CNN	.001
RNN	4.68×10⁻⁰⁵
NN	.002
VecMap and BiVec (Wijayanti et al., 2023)	.001
CNN + FFNN (Wazery et al., 2023)	.002
POCASUM (Deotale et al., 2021)	6.09×10⁻⁰⁵

4.4.4. Statistical Analysis

The statistical analysis in terms of accuracy was conducted for the proposed HC (TL-based modified LSTM + Bi-GRU) over the extant BERTSUM, PEGASUS, LSTM, Bi-GRU, CNN, RNN, NN, VecMap, and BiVec (Wijayanti et al., 2023), CNN + FFNN (Wazery et al., 2023), and POCASUM (Deotale et al., 2021) models on dataset-1 and dataset-2, and their results are given in Tables 11 and 12. For better text summarization, the accuracy values must be high, which is well accomplished by the developed work. The statistical analysis using the proposed HC (TL-based modified LSTM + Bi-GRU) over extant BERTSUM, PEGASUS, LSTM, Bi-GRU, CNN, RNN, NN, VecMap, and BiVec (Wijayanti et al., 2023), CNN + FFNN (Wazery et al., 2023), and POCASUM (Deotale et al., 2021). For better text summarization, the accuracy values must be high, which is well accomplished by the developed work. In Table 11, HC (TL-based modified LSTM + Bi-GRU) discloses a high accuracy value of 0.953 for the maximal case, while BERTSUM, PEGASUS, LSTM, Bi-GRU, CNN, RNN, NN, BiVec (Wijayanti et al., 2023), CNN + FFNN (Wazery et al., 2023), and POCASUM (Deotale et al., 2021) attain relatively lower accuracy values of 0.903, 0.909, 0.894, 0.882, 0.862, 0.892, 0.891, 0.879, 0.887, and 0.883, respectively. For all scenarios, better accuracy is attained using HC (TL-based modified LSTM + Bi-GRU) when evaluated over extant BERTSUM, PEGASUS, LSTM, Bi-GRU, CNN, RNN, NN, BiVec (Wijayanti et al., 2023), CNN + FFNN (Wazery et al., 2023), and POCASUM (Deotale et al., 2021) models.

Table 11.
Statistical Analysis of HC (TL-based Modified LSTM + Bi-GRU) Over Extant Classifiers for Dataset-1.

Methods Minimum Maximum Mean Median SD

BERTSUM 0.866 0.903 0.885 0.885 0.014

PEGASUS 0.868 0.909 0.892 0.896 0.015

LSTM 0.849 0.894 0.871 0.871 0.016

Bi-GRU 0.846 0.882 0.869 0.873 0.014

CNN 0.837 0.862 0.850 0.849 0.012

RNN 0.847 0.892 0.875 0.881 0.017

NN 0.850 0.891 0.867 0.864 0.017

VecMap and BiVec (Wijayanti et al., 2023) 0.851 0.879 0.868 0.871 0.011

CNN + FFNN (Wazery et al., 2023) 0.836 0.887 0.862 0.864 0.020

POCASUM (Deotale et al., 2021) 0.843 0.883 0.865 0.867 0.015

HC (TL-based modified LSTM + Bi-GRU) 0.912 0.953 0.933 0.933 0.015

Methods	Minimum	Maximum	Mean	Median	SD
BERTSUM	0.866	0.903	0.885	0.885	0.014
PEGASUS	0.868	0.909	0.892	0.896	0.015
LSTM	0.849	0.894	0.871	0.871	0.016
Bi-GRU	0.846	0.882	0.869	0.873	0.014
CNN	0.837	0.862	0.850	0.849	0.012
RNN	0.847	0.892	0.875	0.881	0.017
NN	0.850	0.891	0.867	0.864	0.017
VecMap and BiVec (Wijayanti et al., 2023)	0.851	0.879	0.868	0.871	0.011
CNN + FFNN (Wazery et al., 2023)	0.836	0.887	0.862	0.864	0.020
POCASUM (Deotale et al., 2021)	0.843	0.883	0.865	0.867	0.015
HC (TL-based modified LSTM + Bi-GRU)	0.912	0.953	0.933	0.933	0.015

Table 12.

Statistical Analysis of HC (TL-based Modified LSTM + Bi-GRU) Over Extant Classifiers for Dataset-2.

Methods	Minimum	Maximum	Mean	Median	SD
BERTSUM	0.874	0.898	0.884	0.882	0.009
PEGASUS	0.869	0.907	0.890	0.892	0.014
LSTM	0.856	0.876	0.868	0.870	0.007
Bi-GRU	0.858	0.865	0.861	0.860	0.003
CNN	0.848	0.872	0.859	0.859	0.010
RNN	0.847	0.861	0.856	0.858	0.006
NN	0.850	0.890	0.874	0.879	0.016
VecMap and BiVec (Wijayanti et al., 2023)	0.838	0.882	0.858	0.856	0.018
CNN + FFNN (Wazery et al., 2023)	0.854	0.878	0.867	0.867	0.011
POCASUM (Deotale et al., 2021)	0.854	0.868	0.859	0.858	0.005
HC (TL-based modified LSTM + Bi-GRU)	0.916	0.948	0.934	0.935	0.012

Similarly, on dataset-2, a statistical analysis in terms of accuracy was also conducted, and the outcomes in Table 12 proved the HC (TL-based modified LSTM + Bi-GRU) model's superior performance and higher accuracy. This is mainly due to the proposed new activation function in the modified LSTM that replaces the existing function. The new activation function minimizes the risk of skewed summaries by minimizing the impact of extreme values or outliers, thus enhancing the performance of text summarization.

4.4.5. K-Fold Analysis

By combining TL-based modified LSTM and Bi-GRU architectures, the suggested HC model consistently outperforms the other models, as shown by the K-Fold analysis performed on the APP-350 and BillSum datasets in Table 13. BERTSUM, PEGASUS, CNN, RNN, and POCASUM are among the baseline models that the HC model continuously outperforms as the number of folds increases from k = 2 to k = 6, reaching top performance scores of 0.803407 and 0.79547 on dataset 1 and dataset 2, respectively. This pattern indicates that the model's architecture well captures semantic and contextual subtleties in policy texts, highlighting its robustness and generalizability across different data splits. The HC model's consistent improvement with larger k-values suggests improved learning stability and lower variance in contrast to classic models that exhibit varying or plateauing performance. These results provide a trustworthy standard for further study in policy summary and categorization tasks, in addition to validating the effectiveness of the suggested method.

Table 13.
Analysis of K-Fold.

HC

(TL-based

VecMap modified

Bi- and CNN + LSTM

BERTSUM PEGASUS LSTM GRU CNN RNN NN BiVec FFNN POCASUM + Bi-GRU)

Dataset 1

k = 2 0.754801 0.768771 0.7327 0.725023 0.718295 0.721898 0.719015 0.7258 0.719153 0.735263 0.778922

k = 3 0.765334 0.770929 0.73515 0.73611 0.726556 0.738841 0.734692 0.731162 0.722981 0.740314 0.794268

k = 4 0.767554 0.783615 0.743464 0.741724 0.746934 0.751897 0.735079 0.734663 0.747882 0.742704 0.796347

k = 5 0.774948 0.789594 0.747396 0.742678 0.749658 0.767746 0.764472 0.75413 0.75558 0.762871 0.803164

k = 6 0.777656 0.794875 0.763251 0.77405 0.760376 0.776308 0.764745 0.773766 0.763759 0.778296 0.803407

Dataset 2

k = 2 0.742275 0.75458 0.740027 0.728303 0.732387 0.735005 0.728923 0.730795 0.734393 0.74675 0.778223

k = 3 0.752052 0.760085 0.742202 0.756924 0.736126 0.738752 0.74042 0.735534 0.747458 0.747877 0.790312

k = 4 0.756711 0.765523 0.760007 0.763878 0.736499 0.742276 0.751707 0.753772 0.765702 0.748398 0.79143

k = 5 0.775449 0.775372 0.761402 0.770354 0.739741 0.750461 0.757978 0.758006 0.77252 0.757217 0.792169

k = 6 0.781106 0.780942 0.764009 0.774144 0.751934 0.751033 0.760241 0.758236 0.77751 0.767615 0.79547

											HC
Dataset 1
k = 2	0.754801	0.768771	0.7327	0.725023	0.718295	0.721898	0.719015	0.7258	0.719153	0.735263	0.778922
k = 3	0.765334	0.770929	0.73515	0.73611	0.726556	0.738841	0.734692	0.731162	0.722981	0.740314	0.794268
k = 4	0.767554	0.783615	0.743464	0.741724	0.746934	0.751897	0.735079	0.734663	0.747882	0.742704	0.796347
k = 5	0.774948	0.789594	0.747396	0.742678	0.749658	0.767746	0.764472	0.75413	0.75558	0.762871	0.803164
k = 6	0.777656	0.794875	0.763251	0.77405	0.760376	0.776308	0.764745	0.773766	0.763759	0.778296	0.803407
Dataset 2
k = 2	0.742275	0.75458	0.740027	0.728303	0.732387	0.735005	0.728923	0.730795	0.734393	0.74675	0.778223
k = 3	0.752052	0.760085	0.742202	0.756924	0.736126	0.738752	0.74042	0.735534	0.747458	0.747877	0.790312
k = 4	0.756711	0.765523	0.760007	0.763878	0.736499	0.742276	0.751707	0.753772	0.765702	0.748398	0.79143
k = 5	0.775449	0.775372	0.761402	0.770354	0.739741	0.750461	0.757978	0.758006	0.77252	0.757217	0.792169
k = 6	0.781106	0.780942	0.764009	0.774144	0.751934	0.751033	0.760241	0.758236	0.77751	0.767615	0.79547

4.5 Practical Implications

This study has important implications for improving user understanding and interaction with privacy policies on digital platforms. The suggested hybrid methodology gives users, particularly those with little technical literacy, the ability to make knowledgeable decisions regarding personal data protection by automating the classification and summarizing of long and intricate policy texts. This method can be incorporated into websites, mobile platforms, and social media apps to offer easily accessible, real-time summaries that emphasize important policy features including location tracking, data gathering, usage, sharing, and contact details. By providing regulations in an easier-to-understand manner, firms can also use this paradigm to increase transparency, lower legal risks, and build user trust.

5. Conclusion

This study developed a comprehensive policy categorization and text summarization model, beginning with policy categorization, where policies were split into paragraphs using an EDCos membership-based FCM algorithm. Using this categorized data, pre-processing was conducted, which includes tokenization and stemming processes. Afterwards, features including SDC-based TF-IDF, Thematic features, and BOW are extracted. Finally, the summarized text was produced using a HC combining TL-based modified LSTM and Bi-GRU models. Experiments on the APP-350 and BillSum datasets showed that the suggested HC strategy was effective, with superior accuracy of 95% and 95.6%, respectively. Through the integration of TL-based modified LSTM and Bi-GRU architectures, this model achieves a significant improvement over conventional ML and DL techniques in both policy classification and text summarization tasks. By capturing both magnitude and directional similarities, the use of EDCos membership-based FCM clustering improves classification accuracy. Additionally, sophisticated feature extraction techniques like SDC-based TF-IDF, thematic features, and Bag of Words (BOW) help to produce coherent and contextually relevant summaries. In terms of academia, this work furthers the subject of intelligent document summarization, namely in the analysis of privacy policies, and provides a framework that is scalable and interpretable for use in other fields such as financial, legal, and medical texts. However, the model may not work well in multilingual or low-resource environments because it is primarily focused on English-language datasets. Furthermore, the dependence on accuracy measurements restricts a more thorough assessment of semantic fidelity and user comprehension, and computational complexity may increase with larger datasets. In order to improve user engagement during policy acceptance, future research should look into cross-lingual and multilingual extensions, integrate transformer-based architectures or reinforcement learning, create user-centric evaluation frameworks, and examine real-time deployment strategies for web and mobile platforms.

Footnotes

Funding

The authors received no financial support for the research, authorship, and/or publication of this article.

Declaration of Conflicting Interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Appendix

References

Abdi

Hasan

Shamsuddin

S. M.

Idris

Piran

(2021). A hybrid deep learning architecture for opinion-oriented multi-document summarization based on multi-feature fusion. Knowledge-Based Systems, 213, 106658. https://doi.org/10.1016/j.knosys.2020.106658

Alahmadi

Wali

Alzahrani

(2022). TAAM: Topic-aware abstractive Arabic text summarisation using deep recurrent neural networks. Journal of King Saud University-Computer and Information Sciences, 34(6), 2651–2665. https://doi.org/10.1016/j.jksuci.2022.03.026

Alami

Mallahi

M. E.

Amakdouf

Qjidaa

(2021a). Hybrid method for text summarization based on statistical and semantic treatment. Multimedia Tools and Applications, 80(13), 19567–19600. https://doi.org/10.1007/s11042-021-10613-9

Alami

Meknassi

En-nahnahi

El Adlouni

Ammor

(2021b). Unsupervised neural networks for automatic Arabic text summarization using document clustering and topic modeling. Expert Systems with Applications, 172, 114652. https://doi.org/10.1016/j.eswa.2021.114652

Alqaisi

Ghanem

Qaroush

(2020). Extractive multi-document Arabic text summarization using evolutionary multi-objective optimization with K-medoid clustering. IEEE Access, 8, 228206–228224. https://doi.org/10.1109/ACCESS.2020.3046494

Basheer

Anbarasi

Sakshi

D. G.

Vinoth Kumar

(2020). Efficient text summarization method for blind people using text mining techniques. International Journal of Speech Technology, 23(4), 713–725. https://doi.org/10.1007/s10772-020-09712-z

Belwal

R. C.

Rai

Gupta

(2021). A new graph-based extractive text summarization using keywords or topic modeling. Journal of Ambient Intelligence and Humanized Computing, 12(10), 8975–8990. https://doi.org/10.1007/s12652-020-02591-x

Bhuyan

S. S.

Mahanta

S. K.

Pakray

Favre

(2023). Textual entailment as an evaluation metric for abstractive text summarization. Natural Language Processing Journal, 4, 100028. https://doi.org/10.1016/j.nlp.2023.100028

Cheng

Zhang

Guo

(2020). A syntax-augmented and headline-aware neural text summarization method. IEEE Access, 8, 218360–218371. https://doi.org/10.1109/ACCESS.2020.3042886

10.

Deo

Banik

Pattnaik

P. K.

(2025). Customized long short-term memory architecture for multi-document summarization with improved text feature set. Data & Knowledge Engineering, 159, 102440. https://doi.org/10.1016/j.datak.2025.102440

11.

Deotale

Rawat

Vijayarajan

Surya Prasath

V. B.

(2021). POCASUM: Policy categorizer and summarizer based on text mining and machine learning. Soft Computing, 25(14), 9365–9375. https://doi.org/10.1007/s00500-021-05916-w

12.

Dixit

R. S.

Apte

S. S.

(2012). Improvement of text summarization using fuzzy logic based method. IOSR Journal of Computer Engineering (IOSRJCE), 5(6), 5–10. https://doi.org/10.9790/0661-0560510

13.

Huo

(2020). News text summarization based on multi-feature and fuzzy logic. IEEE Access, 8, 140261–140272. https://doi.org/10.1109/ACCESS.2020.3007763

14.

Ertam

Aydin

(2022). Abstractive text summarization using deep learning with a new Turkish summarization benchmark dataset. Concurrency and Computation: Practice and Experience, 34(9), e6482. https://doi.org/10.1002/cpe.6482

15.

Ghodratnama

Beheshti

Zakershahrak

Sobhanmanesh

(2020). Extractive document summarization based on dynamic feature space mapping. IEEE Access, 8, 139084–139095. https://doi.org/10.1109/ACCESS.2020.3012539

16.

Hazra

Ghosh

Mukherjee

Saha

Nandi

Ghosh

Chakraborty

(2021). Sustainable text summarization over mobile devices: An energy-aware approach. Sustainable Computing: Informatics and Systems, 32, 100607. https://doi.org/10.1016/j.suscom.2021.100607

17.

Hernández-Castañeda

García-Hernández

R. A.

Ledeneva

Millán-Hernández

C. E.

(2020). Extractive automatic text summarization based on lexical-semantic keywords. IEEE Access, 8, 49896–49907. https://doi.org/10.1109/ACCESS.2020.2980226

18.

Hernández-Castañeda

García-Hernández

R. A.

Ledeneva

Millán-Hernández

C. E.

(2022). Language-independent extractive automatic text summarization based on automatic keyword extraction. Computer Speech & Language, 71, 101267. https://doi.org/10.1016/j.csl.2021.101267

19.

Hou

(2020). Knowledge-guided unsupervised rhetorical parsing for text summarization. Information Systems, 94, 101615. https://doi.org/10.1016/j.is.2020.101615

20.

https://usableprivacy.org/data

21.

https://www.kaggle.com/datasets/akornilo/billsum

22.

Jang

Kim

(2021). Reinforced abstractive text summarization with semantic added reward. IEEE Access, 9, 103804–103810. https://doi.org/10.1109/ACCESS.2021.3097087

23.

Jiang

Zhang

Dai

Zhao

Feng

Ganchev

(2021). Enhancements of attention-based bidirectional lstm for hybrid automatic text summarization. IEEE Access, 9, 123660–123671. https://doi.org/10.1109/ACCESS.2021.3110143

24.

Jindal

S. G.

Kaur

(2020). Automatic keyword and sentence-based text summarization for software bug reports. IEEE Access, 8, 65352–65370. https://doi.org/10.1109/ACCESS.2020.2985222

25.

Khaleghi

Fakhredanesh

Hourali

(2021). MSCSO: Extractive multi-document summarization based on a new criterion of sentences overlapping. Iranian Journal of Science and Technology, Transactions of Electrical Engineering, 45(1), 195–205. https://doi.org/10.1007/s40998-020-00361-1

26.

Peng

Tang

Zhang

(2020). Text summarization method based on double attention pointer network. IEEE Access, 8, 11279–11288. https://doi.org/10.1109/ACCESS.2020.2965575

27.

Mohamed

(2011). The effect of automatic tokenization, vocalization, stemming, and POS tagging on Arabic dependency parsing. In Proceedings of the Fifteenth Conference on Computational Natural Language Learning. p. 10–18.

28.

Phani

Abdul

Prasad

M. K.

Reddy

V. D.

(2025). MATSFT: User query-based multilingual abstractive text summarization for low resource Indian languages by fine-tuning mT5. Alexandria Engineering Journal, 127, 129–142. https://doi.org/10.1016/j.aej.2025.04.031

29.

Rahul Raj

Haroon

R. P.

Sobhana

N. V.

(2020). A novel extractive text summarization system with self-organizing map clustering and entity recognition. Sādhanā, 45(1), 32. https://doi.org/10.1007/s12046-019-1248-0

30.

Roul

R. K.

Sahoo

J. K.

Arora

(2017). Modified TF-IDF term weighting strategies for text categorization. 2017 14th IEEE India council international conference (INDICON). p. 1–6.

31.

Rupapara

Rustam

Shahzad

H. F.

Mehmood

Ashraf

Choi

G. S.

(2021). Impact of SMOTE on imbalanced text features for toxic comments classification using RVVC model. IEEE Access, 9, 78621–78634. https://doi.org/10.1109/ACCESS.2021.3083638

32.

Sanchez-Gomez

J. M.

Vega-Rodríguez

M. A.

Pérez

C. J.

(2021). The impact of term-weighting schemes and similarity measures on extractive multi-document text summarization. Expert Systems with Applications, 169, 114510. https://doi.org/10.1016/j.eswa.2020.114510

33.

M. H.

C. H.

Cheng

H. T.

(2020). A two-stage transformer-based approach for variable-length abstractive summarization. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 28, 2061–2072. https://doi.org/10.1109/TASLP.2020.3006731

34.

Tomer

Kumar

(2020). Improving text summarization using ensembled approach based on fuzzy with LSTM. Arabian Journal for Science and Engineering, 45(12), 10743–10754. https://doi.org/10.1007/s13369-020-04827-6

35.

Verma

Thomas

Verma

R. K.

(2025). Generative artificial intelligence-based modified abstractive cross attention enabled sequence to sequence model for abstractive Hindi text summarization. Engineering Applications of Artificial Intelligence, 158, 111478. https://doi.org/10.1016/j.engappai.2025.111478

36.

Wang

Shao

Zhang

Zheng

Meng

(2022). Traffic flow prediction using bi-directional gated recurrent unit method. Urban Informatics, 1(1), 16. https://doi.org/10.1007/s44212-022-00015-z

37.

Wazery

Y. M.

Saleh

M. E.

Ali

A. A.

(2023). An optimized hybrid deep learning model based on word embeddings and statistical features for extractive summarization. Journal of King Saud University-Computer and Information Sciences, 35(7), 101614. https://doi.org/10.1016/j.jksuci.2023.101614

38.

Wijayanti

Khodra

M. L.

Surendro

Widyantoro

D. H.

(2023). Learning bilingual word embedding for automatic text summarization in low resource language. Journal of King Saud University-Computer and Information Sciences, 35(4), 224–235. https://doi.org/10.1016/j.jksuci.2023.03.015

39.

Xiao

Sotelo

M. A.

Cao

Zhou

Wang

(2020). An improved LSTM model for behavior recognition of intelligent vehicles. IEEE Access, 8, 101514–101527. https://doi.org/10.1109/ACCESS.2020.2996203

40.

Lee

Zhang

(2020). Multi-task learning for abstractive text summarization with key information guide network. EURASIP Journal on Advances in Signal Processing, 2020(1), 16. https://doi.org/10.1186/s13634-020-00674-7

41.

Yang

Shen

Zhao

Chen

(2020). Hierarchical human-like deep neural networks for abstractive text summarization. IEEE Transactions on Neural Networks and Learning Systems, 32(6), 2744–2757. https://doi.org/10.1109/TNNLS.2020.3008037

42.

Yao

Zhang

Luo

Tao

(2018). Dual encoding for abstractive text summarization. IEEE transactions on Cybernetics, 50(3), 985–996. https://doi.org/10.1109/TCYB.2018.2876317

43.

You

Zhao

Chen

(2020). A topic information fusion and semantic relevance for text summarization. IEEE Access, 8, 178946–178953. https://doi.org/10.1109/ACCESS.2020.2999665

44.

Zhu

L. F.

Wang

J. S.

Wang

H. Y.

(2019). A novel clustering validity function of FCM clustering algorithm. IEEE Access, 7, 152289–152315.

45.

Zimmeck

Story

Smullen

Ravichander

Wang

Reidenberg

Russell

N. C.

Sadeh

(2019). Maps: Scaling privacy compliance analysis to a million apps. Proceedings on Privacy Enhancing Technologies, 66–86. https://doi.org/10.2478/popets-2019-0037

											HC
											(TL-based
								VecMap			modified
				Bi-				and	CNN +		LSTM
	BERTSUM	PEGASUS	LSTM	GRU	CNN	RNN	NN	BiVec	FFNN	POCASUM	+ Bi-GRU)
Dataset 1
k = 2	0.754801	0.768771	0.7327	0.725023	0.718295	0.721898	0.719015	0.7258	0.719153	0.735263	0.778922
k = 3	0.765334	0.770929	0.73515	0.73611	0.726556	0.738841	0.734692	0.731162	0.722981	0.740314	0.794268
k = 4	0.767554	0.783615	0.743464	0.741724	0.746934	0.751897	0.735079	0.734663	0.747882	0.742704	0.796347
k = 5	0.774948	0.789594	0.747396	0.742678	0.749658	0.767746	0.764472	0.75413	0.75558	0.762871	0.803164
k = 6	0.777656	0.794875	0.763251	0.77405	0.760376	0.776308	0.764745	0.773766	0.763759	0.778296	0.803407
Dataset 2
k = 2	0.742275	0.75458	0.740027	0.728303	0.732387	0.735005	0.728923	0.730795	0.734393	0.74675	0.778223
k = 3	0.752052	0.760085	0.742202	0.756924	0.736126	0.738752	0.74042	0.735534	0.747458	0.747877	0.790312
k = 4	0.756711	0.765523	0.760007	0.763878	0.736499	0.742276	0.751707	0.753772	0.765702	0.748398	0.79143
k = 5	0.775449	0.775372	0.761402	0.770354	0.739741	0.750461	0.757978	0.758006	0.77252	0.757217	0.792169
k = 6	0.781106	0.780942	0.764009	0.774144	0.751934	0.751033	0.760241	0.758236	0.77751	0.767615	0.79547

Clustering Technique for Policy Categorization and Text Summarization based on Deep Learning Approaches

Abstract

Keywords

1 Introduction

2 Literature Review

2.1 Related Works

3. HC-Based Policy Categorization and Summarization Model

3.1.1. Data Annotation Using EDCos Membership-based FCM

3.3. Feature Extraction

3.3.1. SDC-based TF-IDF features

3.3.3. Thematic features

4.1 Simulation Procedure

4.2. Dataset Description

4.2.1. Description of Dataset 1 (APP-350 dataset)

Table 2. Training and Testing Data. Training Data (%) Training Data Testing Data Dataset 1 60 2448 1637 70 2858 1227 80 3266 819 90 3675 410 Dataset 2 60 11369 7580 70 13264 5685 80 15159 3790 90 17054 1895

4.4.1 Comparative Analysis on Classifiers

5. Conclusion

Footnotes

Funding

Declaration of Conflicting Interests

Appendix

References

Table 2.
Training and Testing Data.

Training Data (%) Training Data Testing Data

Dataset 1

60 2448 1637

70 2858 1227

80 3266 819

90 3675 410

Dataset 2

60 11369 7580

70 13264 5685

80 15159 3790

90 17054 1895