Text summarization using modified generative adversarial network 1

Abstract

Text summarizing (TS) takes key information from a source text and condenses it for the user while retaining the primary material. When it comes to text summaries, the most difficult problem is to provide broad topic coverage and diversity in a single summary. Overall, text summarization addresses the fundamental need to distill large volumes of information into more manageable and digestible forms, making it a crucial technology in the era of information abundance. It benefits individuals, businesses, researchers, and various other stakeholders by enhancing efficiency and comprehension in dealing with textual data. In this paper, proposed a novel Modified Generative adversarial network (MGAN) for summarize the text. The proposed model involves three stages namely pre-processing, Extractive summarization, and summary generation. In the first Phase, the Text similarity dataset is pre-processed using Lowering Casing, Tokenization, Lemmatization, and, Stop Word Removal. In the second Phase, the Extractive summarization is done in three steps Generating similarity metrics, Sentence Ranking, and Sentence Extractive. In the third stage, a generative adversarial network (GAN) employs summary generation to jointly train the discriminative model D and the generative model G. To classify texts and annotate their syntax, Generative Model G employs a convolutional neural network called Bidirectional Gated Recursive Unit (CNN-BiGRU). The performance analysis of the proposed MGAN is calculated based on the parameters like accuracy, specificity, Recall, and Precision metrics. The proposed MGAN achieves an accuracy range of 99%. The result shows that the proposed MGAN improves the overall accuracy better than 9%, 6.5% and 5.4% is DRM, LSTM, and CNN respectively.

Keywords

Text summarization convolutional neural network bidirectional gated recurrent unit

1 Introduction

Text summarization tools are widely used in Sentiment Analysis that is used to create a brief summaries of Customer reviews, forum discussions, and social media posts which can then be analyzed for sentiment. TS is a complex task, ideally requiring extensive natural language processing skills [1]. To simplify the problem, current research focuses on generating abstract summaries [2]. Sentence-based extract summarization techniques are commonly used in automatic text summarization to create extract summaries [3, 4]. Traditional summarization methods use sentence features to assess the importance of sentences in a document [5]. That limit has nothing to do with calculating the semantic similarity of sentences [6].

In general, a summary can be defined as a text that is constructed from one or more texts and conveys the most crucial details from the source material while being brief enough [7]. Summarization methods can be split into two groups based on the volume of documents like one document and multiple documents [8]. While multi-document summarizing can condense a collection of papers into a summary, single-document summarization can only do so for a single document [9].

ATS is the process of generating concise, accurate summaries of human-generated text [10]. NLP (Natural Language Processing) is a subset of AI (Artificial Intelligence) that deals with human language comprehension and generation [11]. By using NLP techniques, algorithms can identify the most important information and sentences within a text, and generate a summary that captures the main points [12]. An abstractive summarizing strategy aims to comprehend the key ideas in a document using linguistic techniques and then to communicate those ideas succinctly and clearly [13]. According to linguistic and statistical characteristics of sentences, an extractive summary approach chooses all significant sentences from the original text and then concatenates them exactly as they were arranged in the original text [14]. A good summary should select the material from the document that is most pertinent to the reader while also containing little redundant information [15].

The text summarization cuts down on the amount of time needed to understand long articles without losing important information, Flexible and speedy. Our hypothesis is that GAN based unsupervised learning process for routinely noticing the groups of same sentences from which is selected the most illustrative sentence; reducing in this way the length of paragraph in the final summary.

In this paper, a novel MGAN for summarize the text has been proposed. The main contribution of the work has been followed by:

In the first Phase, the Text similarity dataset is pre-processed using Lowering Casing, Tokenization, Lemmatization, and, Stop Word Removal.

In the second Phase, the Extractive summarization is done by three steps Generating similarity metrics, Sentence Ranking, and Sentence Extractive.

In the third stage, a generative adversarial network (GAN) employs summary generation to jointly train the discriminative model D and the generative model G. To classify texts and annotate their syntax, Generative Model G employs a convolutional neural network called Bidirectional Gated Recursive Unit (CNN-BiGRU).

The performance analysis of the proposed MGAN is calculated based on parameter like accuracy, specificity, Recall, and Precision metrics.

The remaining portions of the research can be followed by: Section II illustrates the literature review in detail. Section III represents the proposed technique. section IV represents the result and section V represents the conclusion.

2 Literature review

TS is a technique for condensing long texts so that the summary contains all of the important points from the original document. This section discusses several text summarization methods that have been performed in this area due to the importance of Automatic text summarization.

In 2019 Liu Y, Lapata M [16] suggested a unique document-level encoder and a broad framework for both abstractive and extractive summarizing. The handling problem is identified by the deep reinforced model (DRM) for abstractive summarization employing an interdependence-based architecture where the decoder pays attention to previously produced words. Experiment outcomes using a wide range of datasets reveal that the technique continuously achieves state-of-the-art results when evaluated using fully automated and human-based protocols.

In 2019 Fang, L., et al., [17] suggested a sample-based representation of finite difference distributions for natural language. As just a consequence, implied dormant characteristics could provide more flexible things up a notch than Gaussian-based posteriors. Then they design an LVM to fit the aggregate information backward to the previous. They demonstrate how our models perform in a variety of text generation scenarios, including computational linguistics, unaligned implemented -, and conversation response generation.

Table 1

SI.No Year and Author Proposed Advantages Disadvantages

1 In 2019 Y, Liu A novel document-level encoder based on BERT A two-stage fine-tuning approach can further boost the quality of the generated summaries. Yet this model does not utilize the capabilities of BERT for language generation.

2 In 2019 Fang, L., suggested sample-based representation of finite difference distributions for natural language The success of our model in a variety of generative tasks, including language modeling, unsupervised style transfer, and dialog answer creation, is evidence of its efficacy. This model achieves a little lower Acc due to much more content reserving, even word reserving, of the source sentences.

3 In 2022 A. Jain, A. Hindi language utilizing a Real Coded Genetic Algorithm (RCGA) There is a critical need to locate pertinent information quicker due to the ever-growing volume of text material available online. The produced features’ size and complexity provide challenges for Hindi ATS processing.

4 In 2022 Pasquadibisceglie, V., PROMISE+ Finding more precise and straightforward process models is made possible by this summary phase. This approach may result in the identification of intricate and simple process models that conceal the accurate and pertinent behavior of the underlying process.

5 In 2022 S. Xu, X. SeqCo This model offers an effective sequence-to-sequence text generation method for three various aggregate datasets. This model does not perform in multi-lingual or cross-lingual text generation

6 In 2022 W. Du, Latent structured variable model This model provides better contextual representation that leads to higher-generation quality However, this model requires more computation during training

7 In 2021 S.V. Moravvej, supervised extractive summarization method based on conditional generative adversarial networks using convolutional neural networks. proposed models achieved 5 % better results than other methods Yet this, model attains low coherence of the summary sentence

8 In 2020 J. Cheng, Seq2Seq The suggested technique produces summaries with performance on par with the standard extraction method. The attention mechanism, however, may focus too much on some specific words/phrases and generate them in the summary repeatedly, and incurs the redundancy problem.

9 In 2022 P. Verma, A. developed a two-stage hybrid method. Comparing the output from this model to the reference summaries, the final summaries cover around 50% of the pertinent material. Utilizing classification and clustering techniques to increase system overview coverage and redundancy.

10 In 2023 H. Abo-Bakr, extractive ATS system that aims to extract a small subset of sentences from a large multi-document text. It is simpler and faster than the abstractive method. improve extractive ATS systems.

SI.No	Year and Author	Proposed	Advantages	Disadvantages
1	In 2019 Y, Liu	A novel document-level encoder based on BERT	A two-stage fine-tuning approach can further boost the quality of the generated summaries.	Yet this model does not utilize the capabilities of BERT for language generation.
2	In 2019 Fang, L.,	suggested sample-based representation of finite difference distributions for natural language	The success of our model in a variety of generative tasks, including language modeling, unsupervised style transfer, and dialog answer creation, is evidence of its efficacy.	This model achieves a little lower Acc due to much more content reserving, even word reserving, of the source sentences.
3	In 2022 A. Jain, A.	Hindi language utilizing a Real Coded Genetic Algorithm (RCGA)	There is a critical need to locate pertinent information quicker due to the ever-growing volume of text material available online.	The produced features’ size and complexity provide challenges for Hindi ATS processing.
4	In 2022 Pasquadibisceglie, V.,	PROMISE+	Finding more precise and straightforward process models is made possible by this summary phase.	This approach may result in the identification of intricate and simple process models that conceal the accurate and pertinent behavior of the underlying process.
5	In 2022 S. Xu, X.	SeqCo	This model offers an effective sequence-to-sequence text generation method for three various aggregate datasets.	This model does not perform in multi-lingual or cross-lingual text generation
6	In 2022 W. Du,	Latent structured variable model	This model provides better contextual representation that leads to higher-generation quality	However, this model requires more computation during training
7	In 2021 S.V. Moravvej,	supervised extractive summarization method based on conditional generative adversarial networks using convolutional neural networks.	proposed models achieved 5 % better results than other methods	Yet this, model attains low coherence of the summary sentence
8	In 2020 J. Cheng,	Seq2Seq	The suggested technique produces summaries with performance on par with the standard extraction method.	The attention mechanism, however, may focus too much on some specific words/phrases and generate them in the summary repeatedly, and incurs the redundancy problem.
9	In 2022 P. Verma, A.	developed a two-stage hybrid method.	Comparing the output from this model to the reference summaries, the final summaries cover around 50% of the pertinent material.	Utilizing classification and clustering techniques to increase system overview coverage and redundancy.
10	In 2023 H. Abo-Bakr,	extractive ATS system that aims to extract a small subset of sentences from a large multi-document text.	It is simpler and faster than the abstractive method.	improve extractive ATS systems.

In 2022 Jain, A., et al., [18] suggested a Hindi Health Data (HHD) corpus, an ATS method for the Hindi language. ATS employs the Real Coded Genetic Algorithm (RCGA), which uses choice, designed to simulate Boolean Crossover (SBX), and Quadratic formula Mutation to optimize feature weights. The ATS extraction system reduces summary by 65% when compared to current summarizing techniques.

In 2022 Pasquadibisceglie, V., et al., [19] offered the PROMISE+ technique, which uses abstractions from predictive process mining to provide summaries of lengthy events. You may find more precise and straightforward process models by using this summary step. Experiments utilizing various benchmark event logs and process identification techniques show how successful the suggested approach.

In 2022 Xu, S., et al., [20] provide a sequence-level contrastive learning model for text summarizing that aims to reduce the training gap between an original document’s summary and the final summary. SeqCo consistently beats the high-performance Seq2Seq text generation model in testing on three summary datasets (CNNDM, NYT, and XSum). Multiple contrast lenses were not shown to improve the outcomes in experiments. Creating plans to regularize various, opposing aims is something we’re interested.

Fig. 1

Proposed MGAN summary generation.

In 2022 Du, W., et al., [21] introduced a novel latent structured variable model for producing high-quality text by improving the contextual representation learning of the encoder/decoder model. They offer a GP before reflecting the dependence among random context variables and estimate the GP posterior using an effective amortized variational inference approach to train this stochastic function. Experimental findings show that our technique can develop a better contextual representation, resulting in greater generation quality.

In 2021, Moravvej, S.V. et al., [22] suggested, a supervised extractive summarization approach based on conditional generative adversarial networks with convolutional neural networks (CNN). A novel way for picking phrases as opposed to prior models, which frequently use greedy algorithms. Research on a medical dataset demonstrates that the suggested method works beats rival models by a factor of two. 5% on average and is more comparable to the reference summary.

In 2020, Cheng, J. et al., [23] suggested an automated text summarizing model that uses a heading-aware decoder and a syntax-enhanced encoder to expand the conventional sequence-to-sequence neural text summarization model (Seq2Seq). A sentence’s syntactic structure and word information are both encoded by the encoder in the sentence embedding. The experiment’s findings demonstrate that the suggested method beats summarization, basis models, in terms of ROUGE assessment methodologies and achieves summary generation performance on par with the extractive baseline method.

In 2022, P Verma., et al., [24] created a Fuzzy evolutionary and clustering algorithm used in this method of extractive text summarization. Comparing the proposed strategy to reference summaries, it covers roughly 50% of the relevant content. The outcomes on these common datasets demonstrate the effectiveness of the suggested strategies.

In 2023, Abo-Bakr, H. and Mohamed, S.A., [25] proposed a large-scale sparse multi-objective optimization technique for automatic text summarization of many documents. The integrative ATS system proposed in this paper tries to extract a small selection of phrases from a huge multi-document text. The proposed ATS system has been investigated with the aid of (DUC) datasets.

According to the literature review, most approaches fail to achieve in TS, it is highly challenging for a user to understand the pertinent and significant information from the accessible media or to obtain an accurate overview of it. Also, text summarization contents content election, abstractive vs extraction, Handling Long Documents and Handling ambiguity. To overcome these drawbacks, MGAN has been proposed. The advanced work related to the proposed method is that the proposed MGAN will experiment with other datasets, incorporate other features into the objective function that can choose the sentences most relevant to the document content, and develop other algorithms for solving the optimization problem.

3 Proposed MGAN

In this section, proposed a novel Modified Generative adversarial network (MGAN) to summarize the text. The proposed model involves three stages namely pre-processing, Extractive summarization, and summary generation. In the first Phase, the Text similarity dataset is pre-processed using Lowering Casing, Tokenization, Lemmatization, and, Stop Word Removal. In the second Phase, the Extractive summarization is done in three steps Generating similarity metrics, Sentence Ranking, and Sentence Extractive. In the third stage, a generative adversarial network (GAN) employs summary generation to jointly train the discriminative model D and the generative model G. To classify texts and annotate their syntax, Generative Model G employs a convolutional neural network called Bidirectional Gated Recursive Unit (CNN-BiGRU). The innovation in this paper is the hybridization of GAN with BiGRU for generating text summarization. We not simply combine existing methods instead we fused both the advantages of MGAN and BiGRU can generate more diverse and creative tasks.

3.1 Pre-processing

In this research, pre-processing includes four categories: Lowering Casing, Tokenization, Lemmatization, and, Stop Word Removal which called as LTLS. Pre-processing each text document using the Lowering Casing, Tokenization, Lemmatization, and, Stop Word Removal as shown in Fig. 2.

Fig. 2

LTLS Pre-processing.

3.1.1 Lowering casing

Lowercase the term (NLP to nlp). While words like “book” and “book” have the same meaning, if they are not written in lowercase, they are represented in a vector space model as two separate words, which adds more dimensions.

3.1.2 Tokenization

Tokenization is separating the input document into individual words.

3.1.3 Lemmatization

Lemmatization performs a morphological study of the words, which makes it preferable to stemming.

3.1.4 Stop word removal

Stop words are words that appear frequently in written work (for example, a, an, the, etc.). The inability to discern between the two texts renders these sentences meaningless.

3.2 GAN discriminator

GAN-based text creation approaches use a binary text classifier as an input text created by a generative model G or as a discriminative model, D to identify the input text originally written by a person. To do. However, when utilizing binary classifiers as discriminative models, it is frequently difficult to produce adequate results. Because the discriminative model can predict the output summaries with high confidence, binary classifiers can achieve high accuracy with almost minimal reward. Instead of utilizing a binary classifier as a discriminative model D, we create D using a language model and utilize the output of the language model as the reward for driving the generative model. By maximizing the ground truth summary reward while decreasing the summary reward provided by G, the settings of discriminator D are maximized. As a result, a low-quality summary provided by G may be easily spotted and rewarded. As a result, the generative model G generates summaries that resemble basic (human-written) summaries. The settings of discriminator D and generator G are alternatively iteratively improved. You can retrain the discriminant model once the generative model delivers higher quality and more valid summaries.

$\begin{matrix} \min_{\emptyset} - F_{X \sim qdata} [log D_{\emptyset} (X)] \\ - E_{X \sim T_{⊖}} [log (1 - D_{\emptyset} (X))] \end{matrix}$ (1)

3.2.1 GAN generator

when the discriminant method D has been optimized and modified, they are ready to update the method G. An objective function for the generation G is defined. The Architecture of MGAN is shown In Fig. 3. According to the policy gradient theorem, compute the gradient,

Fig. 3

MGAN Architecture.

$\nabla_{\emptyset Jmixed} = \frac{1}{G} \sum_{g = 1}^{G} \sum_{\hat{X_{g}}} R_{g} . \nabla_{\emptyset} (T_{\emptyset} ({\hat{x}}_{g} | {\hat{X}}_{1 : g - 1,} Y))$ (2)

$= \frac{1}{G} \sum_{g = 1}^{G} E_{\hat{{xg}_{ɛ T \emptyset}}} [R_{g} \nabla_{\emptyset} log \tilde{Q_{g}^{sum}} ({\hat{x}}_{g} | {\hat{X}}_{1 : g - 1,} Y)]$ (3) where T denotes the length of the generated sequence. ^∧Y 1: t is the partial summary produced by G up to time step t. R t denotes the total reward from step t. Specifically, given the generated summaries, the sentence-level reward ^∧Y 1: T is the average reward for each word ^∧y t and is defined as follows:

$R ({\hat{X}}_{1 : G}) = - \frac{1}{G} \sum_{g - 1}^{G} log E_{\emptyset} ({\hat{x}}_{g} | {\hat{X}}_{1 : g - 1})$ (4)

Various words in the abstract should have various rewards. Therefore, we also compute word-level rewards for each token ^∧y t.

$R ({\hat{x}}_{g}) = - log E_{\emptyset} ({\hat{x}}_{g} | {\hat{X}}_{1 : g - 1})$ (5)

Finally, the total reward R t starting from step t is computed by:

$R_{g} = \sum_{k = g}^{G} R ({\hat{X}}_{1 : G}) R ({\hat{x}}_{k})$ (6)

In this GAN generator, Learn more about the CNN biGRU model. This method consists of some layers. Each of the layers is described as follows.

a) Word Embedding

The model’s initial step involves matching words to their vector representations. A continuous vector representation associates words or sentences with comparable vectors based on their geometrical similarity. The finding that identical words appear in similar situations is the basis for the majority of methods for constructing vector representations. Word2Vec was employed to pre-train the word vectors using external knowledge base materials. As a result, a k-dimensional word embedding vector may be utilized to illustrate each word in the text summary. The initialization seed of the model is improved over a randomly initialized model by employing pre-trained Word2Vec. The biGRU encoder accepts these embeddings as inputs. The word embeddings are extended using each word’s biGRU encoder output.

b) BiGRU Encoder

GRU is recognized for successively learning short and lengthy information. The biGRU encoder in this part includes both forward and backward GRU. Both use the identical architectural paradigm, with the exception of reverse GRU, which processes sequences in reverse order. When applied to sentences, biGRU may learn word representations that contain contextual information. As a result, depending on the context, the same word might be rendered differently in two phrases. Let (Y1, Y2,..., Yn) be the input message’s string of word embedding vectors. Each GRU entity analyses these vectors progressively, computing the sequence from the initial hidden state hs0 (hs₁, hs₂,..., hs_n) as follows:

$w = σ (X_{w} Y_{r} + V_{w} {hs}_{r - 1} + c_{w})$ (7)

$g = σ (X_{g} Y_{r} + V_{r} {hs}_{r - 1} + c_{g})$ (8)

${\tilde{hs}}_{r} = tan hs (X_{{hsY}_{r}} + V_{hs} ({gohs}_{r - 1}) + C_{w})$ (9)

${hs}_{r} = {wohs}_{r - 1} + (1_{- w}) o {\tilde{c}}_{2}$ (10) where w and t are the update and reset gates, ${\tilde{hs}}_{r}$ is the candidate’s hidden state ie is converted to the actual hidden state hs_r. X_w, X_g and X_hs matrices. and V_w, V_r, and V_hs are matrices.

c) Sentence Matrix

Text summarization is word2vec and biGRU encoder. By concatenating the biGRU-encoded word representations, for pre-trained word embeddings, we can assume that the vectors represent both the context information and the words themselves.

d) Convolutional Layer

A CNN in NLP encodes many hs words like n-grams at once. CNNs tend to memorize certain patterns over n-grams. Applying a CNN to a sentence matrix, we can assume that the network remembers the contextual information for certain patterns. After randomly initializing the filters, we can assume that each of the k filters is assigned its own role according to the weights during training.

e) Max-pooling Layer

The pooling layer receives the output of the convolutional layer after passing it via a nonlinear activation function. The input is gathered and the representation is flattened by the pooling layer in order to provide an output matrix with a defined size. This makes it possible to utilize and convert various size sets to fixed-size vectors. Each dimension in the vector denotes the number of filters that are active in the sequence, each with a distinct role. The cost is reduced while still preserving relevant functionality in this way. The outcome of the pooling procedure appears as follows:

$B_{pooled} = (\begin{matrix} \begin{matrix} pool (α (B_{1} + c_{1} * d)) \\ \dots \end{matrix} \\ pool (α (B_{n} + c_{n} * d)) \end{matrix})$ (11) where B_n is the i-th convolutional feature map c, a bias term C_n is added, and d is a unit vector of the same size as B_n. B_n is passed through the activation function α (). The most common options for the pooling process are max pooling and average pooling.

f) Softmax Layer

The output of the fifth and final pooling layer, y, is then sent to a fully linked softmax layer. When classifying multiclass data, the softmax function is frequently utilized. Following is how the probability distribution across the classes is calculated.

$\begin{matrix} q (x = i | y, a) & = \underset{i}{softmax} (y^{F} v + a) \\ = \frac{d^{y^{F} v_{i} + a_{i}}}{\sum_{l = 1}^{I} d^{y^{F}} v_{l} + a_{l}} \end{matrix}$ (12)

Where, V_i and a_i are the weight vector and bias, respectively.

4 Result and discussion

The results of the proposed method to Predict Text summarization are discussed in this section. To evaluate the performance of our proposed work, Accuracy, Recall, precision, and Specificity are employed. The proposed technique was simulated using MATLAB 2019b on a machine with 8 GB RAM, 1 TB HD, a 2.60 GHz Intel i8 CPU, and Windows 10.

The following techniques used to verify our hypotheses formulated in the previous section.

Lowering Casing: Lowercase the term (NLP to nlp). While words like “book” and “book” have the same meaning, if they are not written in lowercase, they are represented in a vector space model as two separate words, which adds more dimensions.

Tokenization: It is separating the input document into individual words.

Lemmatization: It performs a morphological study of the words, which makes it preferable to stemming.

Stop Word Removal: Stop words are words that appear frequently in written work. The inability to discern between the two texts renders these sentences meaningless.

Dataset description

This experiment made use of the BBC News dataset. The dataset can be accessed at http://mlg.ucd.ie/datasets/bbc.html. This dataset comprises 2225 documents (all news stories) published during 2004 and 2005 on the BBC News website. The news stories are divided into 5 (five) thematic groups. The dataset is divided into the following categories: business, entertainment, politics, sports, and technology. News article datasets originating from BBC News provide a benchmark for machine learning research. In this work corpus contains nearly 1000 articles in various fields. It contains class labels business, entertainment, politics, sports, etc.

Figure 4 compares the proposed method’s performance with the existing techniques in terms of accuracy, Specificity, recall, and Precision. The proposed method outperforms other models like DRM, LSTM, and CNN by achieving an accuracy of 99% for BBC News datasets, respectively. For both datasets, the proposed method outperforms existing models in terms of accuracy.

Fig. 4

Comparison of existing and proposed model based on the BBC News dataset.

4.1 Performance metrics

For a comparison study, performance was measured using several assessment metrics, including accuracy, precision, specificity, recall, and a statistical analysis of the parameters.

Accuracy:

The accuracy of all correctly predicted categories compared to the dataset’s actual classifications represents the prediction algorithm’s accuracy. Equation (13) determines the model’s accuracy.

$Accuracy = \frac{TP + TN}{totalno . ofsamples}$ (13)

Specificity:

It is the ability to recognize secure instances with accuracy. The calculation is done by determining the percentage of genuine negativity in the underlying data. Equation (14) determines the model Specificity.

$Specificity = \frac{TN}{TN + FP}$ (14)

Recall:

The amount of accurate phishing URL predictions made over all URLs in the dataset is known as the prediction algorithm recall. Equation (15) determines the model recall.

$recall = \frac{TP}{TP + FN}$ (15)

Precision:

Precision is an exact definition of the frequency of positive abnormalities in a particular picture. The higher proportion of information is highlighted by precision. Equation (16) determines the model Precision ratio.

$precision = \frac{TP}{TP + FP}$ (16) where TP, FP, TN, and FN illustrate the number of true positives, false positives, true negatives, and false negatives. By increasing the acc value, prediction performance can be improved.

Figure 5 depicts how, given a starting value of B, the weights of the four phrases in the document vary. The number of iterations is shown on the horizontal axis. The vertex weights converge fairly fast in undirected weighted graphs, while the vertical axis shows the set weight of the current state. The change in vertex weight value is substantially lower than after a few rounds. If the value of the similarity calculation exceeds a certain threshold, we combine the sentences and create a new node. The starting weight vector has no bearing on the graph ranking model’s convergence outcome.

Fig. 5

Example of convergence curves.

Figure 6 illustrates the response of testees regarding three various Text summarizations. The terms “extractive 1” and “extractive 2” relate to the extractive summarization methods of “cosine similarity” and “TF and IDF,” respectively. Each article receives a score between 1 to 5, with 1 denoting a mediocre summary and 5 denoting a superb one. According to the results, the abstractive technique was considered to be the least effective and accurate way, while the extractive summarization method employing cosine similarity was considered to be the most effective and accurate method.

Fig. 6

Summarization methods.

For example, the recommended abstract size ratio should be 33–40%, but some abstracts have a size ratio of up to 80% of the specified text. About 40 items are divided into various sections, as shown in Fig. 7.

Fig. 7

Raw text line counts with line count output summaries.

Figure 8 shows the Sentence via Text Rank Score. The graph-based ranking method forms the foundation of the Text Rank algorithm. Google’s web searches are where it is most commonly utilized, but it has numerous other uses as well. To determine a vertex’s relevance, graph-based ranking algorithms consider data about the entire graph rather than just vertex-specific data. The knowledge on the connections (edges) between the vertices would be a typical piece of information. they must specify the vertices and edges we intend to use in the NLP scenario. In this instance, sentences will serve as the vertices and words as the edges of the connections. As a result, sentences containing words that are used in numerous other sentences are given more weight.

Fig. 8

Sentence via Text Rank Score.

In comparison to proposed, traditional networks like DRM, LSTM, and CNN gain less accuracy. Proposed achieves a high accuracy range of 99 %. Figure 9 shows that the accuracy obtained by DRM, LSTM, and CNN is 90%, 92.5%, and 93.6% respectively, and the Specificity obtained by DRM, LSTM, and CNN is 80%, 85%, and 86% respectively, and the Recall obtained by DRM, LSTM, and CNN is 70%, 76%, and 78% respectively, and the precision obtained by DRM, LSTM, and CNN is 68%, 69%, and 70% respectively.

Fig. 9

Performance Metrics.

Figure 10 illustrates that the suggested technique performed well in both training and testing. Increasing the epoch value would increase method performance. Figure 11 depicts the loss curve and epochs, demonstrating how the model loss decreases with increasing epoch size. As a result, the model’s prediction of outcome was extremely accurate.

Fig. 10

Training and testing accuracy of proposed technique.

Fig. 11

Training and testing loss of the proposed method.

5 Conclusion

In this paper, proposed a novel MGAN for summarize the text. The proposed model involves three stages namely pre-processing, Extractive summarization, and summary generation. In the first Phase, the Text similarity dataset is pre-processed using Lowering Casing, Tokenization, Lemmatization, and, Stop Word Removal. In the second Phase, the Extractive summarization is done in three steps Generating similarity metrics, Sentence Ranking, and Sentence Extractive. In the third stage, a generative adversarial network (GAN) employs summary generation to jointly train the discriminative model D and the generative model G. To classify texts and annotate their syntax, Generative Model G employs a convolutional neural network called Bidirectional Gated Recursive Unit (CNN-BiGRU). The performance analysis of the proposed MGAN is calculated based on parameters like accuracy, specificity, Recall, and Precision metrics. The proposed MGAN achieves an accuracy range of 99%. The result shows that the proposed MGAN improves the overall accuracy better than 9%, 6.5% and 5.4% is DRM, LSTM, and CNN respectively. The limitations when choosing a text summarization method for a specific application and being aware of the trade-offs between different approaches. Additionally, ongoing research aims to address many of these challenges and improve the capabilities of text summarization models. Another way to achieve the objective of the proposed method is for summary extraction BiLSTM or LSTM used and for summary generation gated CNN or optimization-based Deep Learning techniques have been utilized. In future research, will experiment with other datasets, incorporate other features into the objective function that can choose the sentences most relevant to the document content, and develop other algorithms for solving the optimization problem.

References

López-Úbeda

, Martín-Noguerol

, Aneiros-Fernández

, Luna

, Natural Language Processing in Pathology: Current Trends and Future Insights, The American Journal of Pathology (2022).

Duan

, Lu

, Yang

, Wang

and Wang

, An Abstract Summarization Method Combining Global Topics, Applied Sciences 12(20), pp. 10378.

Dhankhar

and Gupta

M.K.

, A statistically based sentence scoring method using mathematical combination for extractive Hindi text summarization, Journal of Interdisciplinary Mathematics 25 (2022), 773–790.

Jiang

, Zou

, Xu

and Zhang

, GATSum: Graph-Based Topic-Aware Abstract Text Summarization, Information Technology and Control 51 (2022), 345–355.

Wang

, Shi

, Liu

and Yan

, A joint FrameNet and element focusing Sentence-BERT method of sentence similarity computation, Expert Systems with Applications 200 (2022), 117084.

Joye

, Maeseele

, Critical discourse analysis: the articulation of power and ideology in texts. In Qualitative Data Analysis, Key Approaches. Sage. (2022).

Wang

, Liu

, Klabjan

, Large-Scale Multi-Document Summarization with Information Extraction and Compression. (2022).

Srivastava

, Singh

, Rana

K.P.S.

and Kumar

, A topic modeled unsupervised approach to single document extractive text summarization, Knowledge-Based Systems 246 (2022), 108636.

El-Kassas

W.S.

, Salama

C.R.

, Rafea

A.A.

and Mohamed

H.K.

, Automatic text summarization: A comprehensive survey, Expert Systems with Applications 165 (2021), 113679.

10.

Alami Merrouni

, Frikh

and Ouhbi

, EXABSUM: a new text summarization approach for generating extractive and abstractive summaries, Journal of Big Data 10(1) (2023), 1–34.

11.

Gómez-Pérez

J.M.

, García-Silva

, Leone

, Albani

, Fontaine

, Poncet

, Summerer

, Donati

, Roma

, Scaglioni

, Artificial Intelligence and Natural Language Processing and Understanding in Space: Four ESA Case Studies. 2022.

12.

Zhang

, Wu

, Zhou

, Duan

, Zhao

, Wang

, SG-Net: Syntax guided transformer for language representation, IEEE Transactions on Pattern Analysis and Machine Intelligence (2020).

13.

Mao

, Yang

, Huang

, Liu

and Li

, Extractive summarization using supervised and unsupervised learning, Expert Systems with Applications 133 (2019), 173–181.

14.

Aries

, Hidouci

W.K.

, Automatic text summarization: What has been done and what has to be done (2019).

15.

Abazari Kia

, Question-driven text summarization with extractive-abstractive frameworks (Doctoral dissertation, University of Essex). 2023.

16.

Liu

, Lapata

, Text summarization with pretrained encoders, arXiv preprint arXiv:1908.08345. 2019 Aug 22.

17.

Fang

, Li

, Gao

, Dong

, Chen

, Implicit deep latent variable models for text generation. 2019.

18.

Jain

, Arora

, Morato

, Yadav

and Kumar

K.V.

, Automatic text summarization for Hindi using real coded genetic algorithm, Applied Sciences 12 (2022), 6584.

19.

Pasquadibisceglie

, Appice

, Castellano

and van der Aalst

, PROMISE: Coupling predictive process mining to process discovery, Information Sciences 606 (2022), 250–271.

20.

, Zhang

, Wu

and Wei

, Sequence level contrastive learning for text summarization. In, Proceedings of the AAAI Conference on Artificial Intelligence 36 (2022), 11556–11565.

21.

, Zhao

, Wang

, Ji

, Diverse Text Generation viaVariational Encoder-Decoder Models with Gaussian Process Priors (2022).

22.

Moravvej

S.V.

, Mirzaei

, Safayani

, Biomedical text summarization using conditional generative adversarial network (CGAN). arXiv preprint arXiv:2110.11870. 2021.

23.

Cheng

, Zhang

and Guo

, A syntax-augmented and headline-aware neural text summarization method, IEEE Access 8 (2020), 218360–218371.

24.

Verma

, Verma

and Pal

, An approach for extractive text summarization using fuzzy evolutionary and clustering algorithms, Applied Soft Computing 120 (2022), 108670.

25.

Abo-Bakr

, Mohamed

S.A.

, Automatic multidocuments text summarization by a large-scale sparse multi-objective optimization algorithm, Complex & Intelligent Systems (2023), 1–16.