Integrating particle swarm optimization with backtracking search optimization feature extraction with two-dimensional convolutional neural network and attention-based stacked bidirectional long short-term memory classifier for effective single and multi-document summarization

Objective

The objective of this research is to create a hybrid deep learning framework that uses PSOBSA for feature extraction and 2DCNN and ABS-BiLSTM for classification to provide high-quality summaries for single and multi-document summarization. The model aims to outperform state-of-the-art techniques like PSO, CSO, LSTM-CNN, SVR, BSA, ACO, and FFA in key evaluation metrics like ROUGE-1 (R-1), ROUGE-2 (R-2), ROUGE-L (R-L), precision, recall, F-measure, cohesion, readability, sensitivity, positive predictive value, best case, average case, worst case, and BLEU score. The system also generates coherent, non-redundant, and linguistically accurate summaries to improve information retrieval, decision-making, and efficiency, across disciplines.

Text preprocessing

Text preprocessing converts text into a predictable and analyzable format for specific tasks. This involves removing extraneous information such as URLs, stop-words, and executing tokenization, stemming, and lemmatization, which eliminates irrelevant data and prepare the dataset for further processing (Mridha et al., 2021). The preprocessing steps in text summarization are determined based on the task’s goal, such as extracting relevant features and removing irrelevant information to improve model’s performance. It is crucial for achieving optimal results in text summarization tasks, as they directly impact the quality and accuracy of the generated summaries. In this work, the preprocessing steps are essential for converting raw text into a structured format suitable for feature extraction and classification. These steps, aligned with methodologies in the literature, such as discussed by Gambhir & Gupta (2017) and Nallapati, Zhai & Zhou (2017), help in dimensionality reduction, eliminate noise, and preserve grammatical structure. Stemming removes word suffixes to reduce words to their root forms, often resulting in non-lexical stems, whereas lemmatization considers the context of a word and reduces to its base or dictionary form, improving the quality of features extracted for models. In our approach, the evolutionary algorithm, PSOBSA is employed to optimize the pre-processed text and extract features effectively. Our preprocessing pipeline is specifically tailored to the needs of the ABS-BiLSTM and 2DCNN classifier models used in the summarization system. The framework effectively extracts linguistic, statistical, and semantic features by implementing chosen preprocessing techniques enhancing summarization accuracy and coherence. Figure 2 below illustrates the preprocessing steps.

Feature extraction using improved PSOBSA

Feature extraction is a pivotal process in machine learning and data analysis, focusing on reducing dimensionality while retaining essential information from raw data. The Enhanced PSOBSA method improves traditional techniques by merging the worldwide search capabilities of PSO with the adaptive refinement mechanisms of BSA. As a result, improved PSOBSA significantly boosts the performance and precision of predictive models across diverse applications (Yadav & Sharan, 2015).

Procedure of PSOBSA

In PSOBSA, the configuration space consists of key elements such as population size, velocity, position update rules, crossover, and mutation rates. The velocity and position update mechanisms are influenced by both the particle’s personal best and the global best positions, ensuring an effective exploration of the solution space. To leverage PSO’s local search, particle positions are updated using neighbourhood-based PSO, while BSA’s global search is applied through mutation and crossover operators. The steps involved to provide a structured approach for exploring the configuration space while maintaining feasible solutions through effective constraint handling in PSOBSA are:

1. Parameters initialization:

   • The initial population size, velocity, and position of particles are defined. Additionally, inertia weight and acceleration coefficients for velocity updates are specified.

   • Crossover and mutation rates based on BSA mechanisms are also set.

2. Representation of scheme:

   • Binary encoding for discrete problems.

   • Real-valued encoding for continuous optimization problems, where each particle represents a vector of decision variables.

3. Initial population evaluation:

   • Each particle’s fitness is calculated using the problem-specific fitness function.

   • Personal best and global best positions are identified in the population.

4. Updating particle velocity and position:

   • Using Eq. (3), each particle’s velocity is updated based on its personal best position, global best position, and current velocity.

   • Particle positions are then updated using the new velocity to explore the solution space.

5. Application of BSA crossover and mutation:

   • Crossover and mutation are applied to introduce diversity within the population.

   • Crossover combines positions of selected particles, while mutation slightly perturbs the positions.

6. Fitness evaluation with constraint handling:

   • The fitness of each particle is evaluated.

   • Penalty function is applied to handle constraints in optimization problems. When a candidate solution (particle) violates a constraint, the penalty function: $f^{\mathrm{\prime }}(x)=f(x)+P(x)$ modifies the fitness value of that solution by adding a penalty, making it less favourable compared to feasible solutions. Where: f(x) is the original fitness function (objective function) that evaluates the quality of the solution x, P(x) is the penalty term added to the fitness function when a constraint is violated and $f^{\mathrm{\prime }}(x)$ is the modified fitness function, incorporating the penalty.

   • Repair mechanisms are used to adjust any infeasible particles, moving them to a feasible region of the search space.

7. Backtracking search algorithm (BSA) component:

   • If a particle violates constraints, backtracking is allowed to a previously feasible position, helping to explore a more promising region of the search space.

8. Adaptive penalty adjustment:

   • Dynamically penalty values are adjusted based on the severity of constraint violations as the algorithm progresses promoting better handling of constraints.

9. Updating global and personal bests:

   • After fitness evaluation, the personal best for each particle and the global best for the population are updated.

10. Checking of termination criteria:

    • If the termination criteria are met (e.g., number of iterations or convergence), the algorithm is stopped.

    • Otherwise, return to step 4, and continue updating particle velocities and positions.

11. Returning of the optimal solution:

    • The global best position and its corresponding fitness value are output as the optimal solution.

This step-by-step procedure highlights the key operations and mechanisms in the PSOBSA algorithm, particularly focusing on the effective use of the penalty function and backtracking for constraint handling, ensuring a robust approach for solving complex optimization problems which is illustrated in Fig. 3.

ABS-BiLSTM and 2DCNN

ABS-BiLSTM and 2DCNN offer advanced techniques for single and multi-document summarization. ABS-BiLSTM leverages bidirectional LSTMs with attention mechanisms to capture contextual information from preceding and succeeding tokens in the text. Meanwhile, 2DCNN applies convolutional operations to detect features across two dimensions, enhancing the model’s capacity to comprehend intricate patterns in the data. Together, these methods improve the accuracy and quality of generated summaries (Sun & Platoš, 2023).

ABS-BiLSTM

A stacked bidirectional LSTM neural network that is attention-based forms the basis of the structure. When converting the text into a fixed-length vector participation, the framework first employs the to clean the text’s data, eliminating particular characters and capitalization and keeping just the details that can be recovered as semantics. To understand the syntax of the sentences in this article, we first analyze the characteristics of the text sequence using stacked Bi-LSTM. Subsequently, a self-attention layer is applied to dynamically weight features, highlighting contextual nuances and addressing ambiguous words. Ultimately, a multilayer perspective is used to create the categorization findings.

Embedding layer: With the use of embeddings, words can now be recognized by computers by being converted into vectors. The neural network language model enables neural network and probabilistic language processing for text models by determining probabilistic language model characteristics through neural networks. When a sentence is lengthy, $h\{C_1,C_2,\dots C_i,\dots C_h\}$, subsequently use the learned word vector modelling to do a word vector mapping, which creates a word vector matrix from the word sequence $\{D_1,D_2,\dots D_i,\dots D_t\}$, whereas matrix size is $h\ast d$, in this case, d is the word vector’s dimension.

Stacked BiLSTM neural network layer: A single neural network cannot effectively extract intricate data characteristics. Deep architectures, like the attention-based stacked Bi-LSTM model, leverage the output of the first Bi-LSTM layer as input for the second. This approach more effectively captures intricate properties, thereby enhancing the model’s feature representation capabilities.

First Bi-LSTM layer $\{D_1,D_2,\dots D_i,\dots D_t\}$, concealed state is mostly transmitted by the first Bi-LSTM layer, where $i$ stands for $i^{th}$ word, the parameters $C_1$, $z_1$ be distributed to the layer’s members. The concatenation of the forward LSTM’s hidden state and the reversed LSTM’s hidden state provides the information for the subsequent Bi-LSTM layer, serving as the output of the first Bi-LSTM layer. Vector of features representation is produced by the second Bi-LSTM layer $t_i$ of text by joining hidden states that are forward and backward. $t_i$ is a graph that includes implicit deep-level relationships in text, which is essential for increasing the accuracy of categorization. The hidden $t_i^1$ state is represented by the following Eqs. (10)–(12):

(10) $$\overrightarrow{t_i^1}=\sigma (C_1[\overrightarrow{t_{i-1}^1,d_i}]+z_1)$$

(11) $$\overrightarrow{t_i^2}=\sigma (C_1[\overrightarrow{t_{i-1}^2,d_i}]+z_1)$$

(12) $$t_i^1=\overrightarrow{t_i^1}\oplus \overleftarrow{t_i^1}.$$

The subsequent formulas determine the stacked Bi-LSTM’s ultimate output in the second layer Eqs. (13)–(15):

(13) $$\overrightarrow{t_i^2}=\sigma (C_2[\overrightarrow{t_{i-1}^2,t_i^1}]+z_2)$$

(14) $$\overrightarrow{t_i^2}=\sigma (C_2[\overrightarrow{t_{i-1}^2,t_i^1}]+z_2)$$

(15) $$t_i^2=\overrightarrow{t_i^2}\oplus \overleftarrow{t_i^2}.$$

Attention layer: After the stacked Bi-LSTM layer, a self-attention mechanism is applied to weigh each word’s context representation vector, creating an attention layer that reflects each word’s significance to sentence semantics. Finally, proportional summation is used to generate the global semantic representation of sentence S.

Step 1: Work out the weight score $e_i$, which calculates how much i-th word contributes to understanding of uncertain terms. Equation (16)’s output $h_i$ across fully linked network having an activation purpose of tanh serves as a calculating procedure.

(16) $$e_i=\tanh (C_c{h}_i+z_c)$$

The calculation of the tanh activation function is Eq. (17):

(17) $$\tanh (d)=\frac{u^d-u^{-d}}{u^d+u^{-d}}$$where $h_i$ is the layer that is hidden the capable of $C_c$ training parameter and the $z_c$ term for bias is the state i-th phrase is used to describe a hyperbolic tangential function, which takes any real integer as input, outputs a value among 0 and 1. It’s utilized for processing information and output; closer output is to 1.0.

Step 2: Determine word’s attention weight ${\alpha }_i$ Softmax function is used following normalization $e_i^T$ and $u_c$ dot creation shown in Eq. (18):

(18) $${\alpha }_i=\frac{\exp (e_i^T{e}_c)}{{\sum }_{i=0}^t\exp (e_i^T{e}_c)}$$where $e_c$ is a learnable context vector with a random start.

Step 3: Equation (19) shows how the calculated weights for attention are applied to hidden layer outputs of stacked Bi-LSTM, or clause vector S.

(19) $$S=\sum _{i=0}^t{\alpha }_i{h}_i$$

For phrases having a self-attention mechanism, the representation vector may be computed using the method described above. Below Fig. 4 shows an architecture of ABS-BiLSTM.

Output Layer: The focal layer outputs a high-dimensional representation, S, which is used as the feature vector. This feature vector is then fed into the fully connected hidden layer, where the softmax activation function, as described in Eq. (20), converts it into an N-dimensional vector.

(20) $$\hat{a}=Softmax(MLP(S)).$$

To train the predictive model, compute the cross-entropy loss function (P) between an anticipated label $\hat{a}$ and ground-truth label $a_i$ using Eq. (21) as follows:

(21) $$P(\theta )=-\sum _{a_i\in a,{\hat{a}}_p\in \hat{a}}{\hat{a}}_i\log {\hat{a}}_i+(1-a_i)\log (1-{\hat{a}}_i).$$

$\theta$ symbolizes the capability of the training parameter vector. Modifying a parameter that reflects the changing slope of the loss operation, $\theta .{\mathrm{\nabla }}_{\theta }P(\theta )$ minimizes the cross-entropy loss function using a gradient descent optimization technique $\theta$. Equation (22) is the value of the gradient descent algorithm update formula.

(22) $$\theta =\gamma .{\mathrm{\nabla }}_{\theta }P(\theta ).$$

$\gamma$ shows the learning rate.

2D convolutional neural network

The primary task of 2DCNN is to learn probable characteristics and hidden representations from a dynamic feature pool, producing feature maps that are then reduced in geographic dimensions through max pooling. Dropout layers are used to prevent overfitting, and batch normalization layers are incorporated to address potential issues. Additionally, data must be normalized before being fed to the next layer due to the sensitivity of deep learning models to different data types. The mathematical formulation of the convolutional layer in a 2DCNN is represented through Eq. (23).

(23) $$a^i={\sigma }_i(c_i\ast d_i+z_i).$$where ${\sigma }_i$ shows the function of sigmoid activation and $a^i$ symbolizes the result of $i^{th}$ convolution layer. $d_i$ references to the 2D data that is the input. Similarly, $c_i$ indicates the weight of $i^{th}$ the layer of convolution and illustrates the bias element. The output $a^i$, after the combining processes are completed, stores the feature maps. After that, a max pooling approach is used to carry out the pooling procedures. The solution for the max-pooling layers is presented through Eq. (24).

(24) $$a^m=ma{x}_{i,j\in J}(a^{i.j}).$$where $a^m$ represents the decreased feature maps obtained from the max pooling layers. To keep the simulation model from overfitting and other problems, the dropout and batch normalization layers are introduced. Additionally, feature map is flattened into a 1-dimensional vector by flatness layer to provide a connection among the subsequent pooling layers impending completely linked layer. Mathematically, Eq. (25) represents the entire linked layer.

(25) $$a^f=g_i(c_i^f\ast a^m+z_i^f)$$here $g_i$ indicates function activation. $c_i^f$ then $z_i^f$ indicates bias factors and weight of the fully connected layer, correspondingly. $a^f$ displays the whole linked layer’s output. Below Fig. 5 shows the architecture of 2DCNN.

Description of dataset

The dataset comprises various collections of documents, including DUC-2002, DUC-2003, DUC-2005, CNN Daily Mail, and the Multi-News dataset. The DUC 2002, 2003, and 2005 databases play a significant role in testing automated text summarizers. These datasets, released by the DUC, were formed by the National Institute of Standards and Technology (NIST) to promote large-scale experimentation among researchers and advance the area of summarization. By providing standardized datasets, DUC facilitates comparative evaluations of summarization algorithms, allowing researchers to assess and compare the effectiveness of different approaches (https://duc.nist.gov).

The multi-news dataset consists of news stories from newser.com and professional human-written summaries of these articles. The dataset includes sources from over 1,500 sites, each appearing at least five times, facilitating the summarization of a variety of documents and summaries (https://www.kaggle.com/datasets/gowrishankarp/newspaper-text-summarization-cnn-dailymail). Table 2 presents single-document dataset description, while Table 3 presents the details of multi-document dataset.

Datasets such as DUC, CNN, and Daily Mail for single/multi-document summarization are preferred based on several key factors:

• Benchmark status: These datasets, particularly DUC (Document Understanding Conference), are widely recognized as standard benchmarks in the text summarization community. Using them allows for direct comparison with a broad range of existing methods, enhancing the comparative value of our results.

• Diverse content: CNN and Daily Mail datasets offer a wide variety of news articles covering multiple topics and writing styles. This diversity helps ensure our model’s robustness across different types of content.

• Multi-document capabilities: The DUC dataset, especially, provides multi-document summarization tasks for both single and multi-document summarization which are crucial for evaluating our model’s versatility.

• Size and quality: These datasets offer a substantial number of document-summary pairs, which is essential for training and evaluating deep learning models.

• Temporal range: By using datasets from different years (e.g., various DUC datasets like DUC 2002, DUC 2003 and DUC 2005), we can assess our model’s performance across evolving language patterns and topics.

The selection of hyperparameters for the ABS-BiLSTM and 2DCNN models in Tables 4–6 was based on a systematic optimization process to achieve the best performance. To balance capturing complicated temporal dependencies and computational efficiency, ABS-BiLSTM used 2–4 stacked layers, 50–512 hidden units, and 0.2–0.5 dropout rate to prevent overfitting. For the 2DCNN, 32–256 filters and 3 × 3, 5 × 5, or 7 × 7 filter sizes were chosen to capture local and wider data correlations. Pooling sizes (2 × 2 or 3 × 3) reduced dimensionality without losing key features, and ReLU was employed as the activation function to mitigate vanishing gradient concerns. The combined model used a batch size of 64 to balance training speed and stability and a learning rate of 0.001 with the Adam optimiser for smooth convergence. To optimise model performance across datasets, these hyperparameters were tested and benchmarked against existing methods.

Hyperparameter tuning studies were conducted to identify optimal configurations for learning rate, batch size, number of layers, and filter sizes. Effects on findings were observed: A deeper ABS-BiLSTM model with more than 512 neurons captures more detailed patterns, which may overfit or increase computing costs. Conversely, fewer neurons reduce model generalisation. In deep architectures like ABS-BiLSTM and 2DCNN, dropout rates below 0.2 provide insufficient regularisation, allowing overfitting. In contrast, a dropout rate above 0.5 deactivates too many neurons, preventing the network from learning by discarding crucial features. Underfitting may result. In 2DCNN, smaller kernels capture finer details, while larger kernels focus on broader patterns, hence tests were needed to establish which size improved feature extraction for summarization. Pooling sizes (2 × 2 or 3 × 3) were changed to reduce dimensionality while keeping crucial information. In combined model construction, too high a learning rate prevents model convergence, while too low delays training, making it harder to get optimal outcomes. Batch size (16 or 32) introduces too much variance in gradient estimates, causing unstable learning, while 128 or 256 batches smooth gradients but require more memory and compute. To calibrate the model for the task, parameter changes were carefully examined for positive and negative consequences.

Figure 6 presents a detailed analysis of various performance metrics for the Multi-Document 2002 dataset. Figure 6A plots precision, recall, and F-measure against epochs, highlighting the model’s convergence and effectiveness. Figure 6B offers a performance analysis of the proposed method against existing techniques like PSO, CSO, LSTM-CNN, SVR, BSA, ACO, and FFA focusing on metrics such as recall, precision, and F-measure. Figure 6C demonstrates that the proposed method achieves higher R-1, R-2, and R-L scores, indicating superior summary quality. Figure 6D provides a comprehensive comparison of metrics like cohesion, sensitivity, positive predictive value, readability, and BLEU score across different methods, showcasing their strengths and weaknesses to aid in selecting the most suitable summarization technique.

Figure 7 provides a comprehensive analysis of the performance metrics for the Multi-Document 2003 dataset. Figure 7A plots precision, recall, and F-measure against epochs, offering insights into the model’s convergence and summarization effectiveness. Figure 7B compares the proposed method with existing techniques like PSO, CSO, LSTM-CNN, SVR, BSA, ACO, and FFA evaluating their performance based on metrics such as recall, precision, and F-measure. Figure 7C contrasts R-1, R-2, and R-L scores between proposed and existing methods, highlighting the superior summarization quality of the proposed approach. Figure 7D presents a detailed analysis of additional performance metrics, including cohesion, sensitivity, positive predictive value, readability, and BLEU score, providing a thorough understanding of each method’s strengths and weaknesses in multi-document summarization.

Figure 8 provides a comprehensive analysis of the performance metrics for the Multi-Document 2005 dataset. Figure 8A plots metrics such as precision, recall, and F-score against epochs, showcasing model’s convergence and summarization effectiveness. Figure 8B compares the proposed method with existing techniques, evaluating their performance based on recall, precision, and F-measure. Figure 8C contrasts R-1, R-2, and R-L scores, highlighting proposed method’s superior alignment with human-generated summaries. Figure 8D offers a detailed comparison of additional performance metrics, revealing the strengths and weaknesses of each technique in multi-document summarization.

Figure 9 provides a detailed analysis of the performance metrics for the multi-document CNN Daily Mail dataset. Figure 9A plots metrics such as precision, recall, and F-score against epochs, showcasing model’s convergence and effectiveness in generating high-quality summaries. Figure 9B compares the proposed method with existing techniques based on recall, precision, and F-measure, revealing their relative effectiveness. Figure 9C contrasts R-1, R-2, and R-L scores, highlighting proposed method’s superior alignment with human-generated summaries. Figure 9D offers a comprehensive comparison of additional metrics like cohesion, sensitivity, positive predictive value, readability, and BLEU score, detailing the strengths and weaknesses of each summarization technique.

Figure 10 provides a comprehensive analysis of performance metrics for the multi-document news dataset. Figure 10A plots precision, recall, and F-score against epochs, showcasing the model’s convergence and effectiveness in generating high-quality summaries. Figure 10B compares the proposed method with existing techniques (PSO, CSO, LSTM-CNN, SVR, BSA, ACO, and FFA) using metrics such as recall, precision, and F-measure, revealing their relative effectiveness. Figure 10C contrasts R-1, R-2, and R-L scores, highlighting proposed method’s superior performance in aligning with human-generated summaries. Figure 10D offers insights into various evaluation criteria, including cohesion, sensitivity, positive predictive value, readability, and BLEU score, aiding in selecting the most suitable summarization technique based on specific requirements.

Figure 11 provides a comprehensive analysis of performance metrics for the single-document 2002 dataset. Figure 11A plots precision, recall, and F-score against epochs, highlighting the model’s convergence and summarization quality. Figure 11B compares the proposed method with existing techniques (PSO, CSO, LSTM-CNN, SVR, BSA, ACO, and FFA) based on recall, precision, and F-measure, revealing their relative effectiveness. Figure 11C contrasts R-1, R-2, and R-L scores, showing proposed method’s superior alignment with human-generated summaries. Figure 11D offers insights into various evaluation criteria, including cohesion, sensitivity, positive predictive value, readability, and BLEU score, aiding in selecting the most suitable summarization technique based on specific requirements.

Figure 12 provides a comprehensive analysis of performance metrics for the single-document 2003 dataset. Figure 12A plots precision, recall, and F-measure against epochs, illustrating the model’s learning progress and convergence, aiding in fine-tuning and optimizing performance. Figure 12B compares the proposed method with existing techniques (PSO, CSO, LSTM-CNN, SVR, BSA, ACO, and FFA), showing superior performance in accuracy, precision, recall, and F-measure. Figure 12C contrasts R-1, R-2, and R-L scores, demonstrating the proposed method’s effectiveness in producing high-quality summaries. Figure 12D offers a detailed comparison of metrics such as cohesion, sensitivity, positive predictive value, readability, and BLEU score, evaluating the strengths and weaknesses of each summarization technique.

Figure 13 provides a comprehensive analysis of performance metrics for the single-document 2005 dataset. Figure 13A plots precision, recall, and F-measure against epochs, illustrating the model’s learning dynamics and convergence, aiding in optimizing model parameters and fine-tuning performance. Figure 13B compares the proposed method with existing techniques (PSO, CSO, LSTM-CNN, SVR, BSA, ACO, and FFA), highlighting its superior accuracy, precision, recall, and F-measure, demonstrating its effectiveness in classification tasks. Figure 13C contrasts R-1, R-2, and R-L scores, showing proposed method’s consistent outperformance in generating high-quality summaries. Figure 13D offers a detailed comparison of cohesion, sensitivity, positive predictive value, readability, and BLEU score, providing a thorough understanding of each technique’s strengths and weaknesses in document summarization.

In Table 7, the performance of the 2DCNN-ABS-BiLSTM approach is evaluated across multiple iterations on different datasets: DUC-2002, DUC-2003, DUC-2005, Multi-News, and CNN/Daily Mail. The table presents precision, recall, and F-measure scores for each and every dataset at various iterations. Overall, the approach demonstrates consistent improvement across iterations, with precision, recall, and F-measure scores generally increasing with each iteration.

Table 8 provides a performance evaluation of several techniques for text summarization across dissimilar datasets: DUC 2002, DUC 2003, DUC 2005, Multi-News, and CNN/Daily Mail. Evaluation metrics include R-1, R-2, and R-L scores. Results show that proposed model consistently outperforms other techniques across all datasets and evaluation metrics, achieving the highest R scores overall. This highlights effectiveness of proposed model in generating quality summaries compared to existing methods across diverse datasets and evaluation criteria.

The proposed model, PSOBSA with 2DCNN and ABS-BiLSTM, offers several advantages over existing techniques for both single and multi-document summarization. By employing PSOBSA, it enhances feature extraction efficiency, improving performance by combining the exploration of PSO with the exploitation of BSA. This hybrid optimization is crucial for accurate feature extraction in summarization tasks. The model’s classification ability is significantly improved by using a 2DCNN and ABS-BiLSTM, allowing it to capture complex text patterns and context more effectively than traditional methods. In single-document summarization, it surpasses methods like PSO, CSO, LSTM-CNN, SVR, BSA, ACO, and FFA across various metrics such as R-1, R-2, and R-L, consistently achieving higher precision, recall, and F-measure scores, sensitivity, readability, positive predictive value, best case, average case, worst case and BLEU score. For multi-document summarization, it outperforms methods like PSO, CSO, LSTM-CNN, SVR, BSA, ACO, and FFA, delivering superior results on datasets like DUC 2002, DUC 2003, DUC 2005, Multi-News, and CNN/Daily Mail with better cohesion, sensitivity, readability, positive predictive value, best case, average case, worst case, Rouge scores and BLEU score. The model’s versatility across multiple benchmark datasets demonstrates its robustness, consistently delivering improved summaries in both single and multi-document tasks. Our proposed model shows performance across the given various metrics, making it a comprehensive solution for generating coherent, accurate, and grammatically sound summaries.