Confidence-aware semi-supervised siamese graph networks for semantic text similarity in low-resource Urdu Authors: Anil Ahmed, Lei Guo, Degen Huang, Salahuddin Unar, Mobeen Nazar Published: 2026 ABSTRACT Accurate detection of semantically similar texts underpins applications such as plagiarism detection and content recommendation, yet remains challenging in low-resource languages like Urdu due to limited labeled data and noisy annotations. We introduce Siamese Graph-Based Semi-Supervised Learning (SGSSL), a framework that expands training diversity through back-translation and masked-word prediction, and learns from unlabeled data via confidence-aware pseudo-labeling with consistency regularization. At its core, SGSSL employs a Siamese Graph Neural Network that fuses Graph Convolutional Networks and Graph Attention Networks to capture fine-grained relational signals between sentences. A bottleneck feature-refinement layer and a Transformer Encoder with multi-head self-attention mitigate GNN over-smoothing while modeling long-range dependencies. We further leverage Urdu RoBERTa embeddings for deep contextual semantics. Our hybrid confidence-based filtering integrates probabilistic confidence, semantic similarity, and model uncertainty to select high-quality pseudo-labels. Evaluated on six benchmark datasets for text reuse and paraphrase detection (binary and ternary classification), SGSSL consistently outperforms strong baselines. On the USTRC dataset, SGSSL attains an accuracy of 0.722, an 8.1% absolute gain over the best baseline, demonstrating effectiveness for low-resource settings. Introduction Automatic paraphrase detection is a critical task in natural language processing (NLP), essential in identifying paraphrased content across different linguistic levels, from single words to entire sentences. It serves as a classification task, distinguishing between paraphrased, non-paraphrased, and partially paraphrased texts. This capability underpins a wide range of NLP applications, such as query ranking1,2, question answering3,4, document summarization 5,6, and plagiarism detection7,8. Among these, detecting paraphrased plagiarism, where text is reworded to maintain its original meaning, is especially important as plagiarism has become increasingly prevalent due to easy access to online content. The rise of plagiarism is a pressing concern in academic, publishing, and professional spheres. Surveys indicate that up to 90% of students have engaged in plagiarism, with 50% admitting to doing so in assignment submissions9,10. High-profile cases have been uncovered globally, including over 200 documented instances of academic plagiarism in Germany. A report also found that 66% of student work at leading U.S. universities contained plagiarized content. Furthermore, the use of AI-generated text in academic assignments is becoming more widespread, exacerbating the issue. Undetected plagiarism undermines academic integrity, leads to unearned professional recognition, and misallocates research funding. Consequently, detecting paraphrased plagiarism is critical yet challenging, especially given the surge of digital content and advanced AI tools, such as Jasper, Writesonic, and Rytr, which make paraphrasing and text generation more accessible. Alongside paraphrase detection (P-D), text reuse (T-R) has gained attention, focusing on intentionally recycling text to generate new intellectual content. T-R spans direct copying, paraphrasing, and more subtle forms of content reuse. Local reuse integrates existing text into new content, often intersecting with P-D tasks. In contrast, global reuse, which involves detecting nearly identical documents across various locations, is crucial for duplication detection and version control tasks. T-R and P-D are complex due to the vast variability in linguistic expression, driving ongoing research in NLP, particularly in multilingual contexts. Despite advancements in NLP for languages such as English and European languages, there is a significant gap in resources for under-resourced languages like Urdu. Urdu’s intricate morphology, rich vocabulary, and flexible syntax introduce challenges for automated processing tasks, including T-R and P-D. As a derivative of Hindustani, Urdu allows flexible word order, which leads to ambiguities in computational text analysis. While Deep Neural Networks (DNNs) have shown great promise in Urdu language processing11–15, their performance relies heavily on large annotated datasets, which are often unavailable for under-resourced languages. Traditional models and DNNs require significant labeled data, but acquiring such data is costly and time-intensive. In order to address the above problems, leveraging semi-supervised learning (SSL) approaches has emerged as a promising strategy. SSL utilizes limited labeled data in combination with abundant unlabeled data, reducing dependency on large annotated corpora16. Additionally, mitigating overfitting, which is a common challenge when working with small datasets, can be achieved by augmenting the training data. This study proposes an innovative approach to tackle the challenges of P-D and T-R in Urdu, a language that has not been sufficiently explored in this context. This study proposes a novel Siamese graph-based semi-supervised learning (SGSSL) approach that integrates Graph Convolutional Networks (GCN) and Graph Attention Networks (GAT). Our method leverages the strengths of GCN and GAT to enhance textual data representation, focusing specifically on T-R and P-D tasks in Urdu. These tasks are crucial in linguistic research, aiming to discern semantic similarities and differences between sentences. Our model incorporates Transformer Encoder (TE) modules enriched with MHSA, capturing intricate language patterns and dependencies across varying sentence lengths and complex content in Urdu texts. The integration of the TE helps mitigate the over-smoothing issue in GCN and GAT, as suggested in a similar study17,18. We utilize a bottleneck layer to consolidate features extracted by GCN and GAT before inputting them into the TE module. This structured integration boosts computational efficiency and enhances classification accuracy by effectively leveraging semantic relationships encoded in the graph structure. Furthermore, we employ advanced Data Augmentation (DA) techniques such as Back Translation (BT) and Masked Word Prediction (MWP) to enrich the training corpus and improve model robustness against variations in linguistic expression. This addresses a significant research gap in Urdu NLP, where prior studies have not systematically explored their impact on enhancing classification performance. We employ a hybrid confidence-based filtering (HCBF) strategy to tackle noise in pseudo-labeled samples. This method integrates probabilistic confidence, semantic similarity, and model uncertainty to compute hybrid confidence scores. Samples exceeding a predefined confidence threshold are retained for subsequent training, ensuring the inclusion of high-confidence pseudo-labeled data. We implement consistency regularization to bolster model robustness against minor input perturbations and mitigate the influence of noisy labels. Experimental evaluations on benchmark datasets underscore the efficacy of our approach, demonstrating superior performance compared to existing methods in T-R and P-D tasks, relying on pre-trained contextual embedding. Our contributions are the following:We utilize pre-trained Urdu Roberta embeddings to generate deep contextualized representations for T-R and P-D in Urdu, improving model accuracy on ambiguous input sentences.We introduce a robust hybrid confidence-based filtering mechanism to handle noisy pseudo-labeled data, enhancing the quality of training samples.Our unified sentence-level and word-level DA techniques, including BT and MWP, enrich the dataset and enhance model robustness.We propose SGSSL, a novel Siamese architecture combining GCN, GAT, TE, and MHSA for T-R and P-D tasks, leveraging both labeled and unlabeled data.We provide a comprehensive analysis of the impact of DA and HCBF techniques and various model components on the performance of Urdu T-R and P-D tasks.Empirical evaluations across six benchmark datasets show significant performance improvements over state-of-the-art (SOTA) baseline methods in SSL text classification.The structure of this paper is outlined as follows: “Related works” presents a review of the literature relevant to our research. In “Proposed framework”, we introduce our proposed framework. “Experiments and analysis” elaborates on the experimental setup and provides a comparative analysis with SOTA baseline methods. “Ablation study” details the ablation study conducted. “Comparison with baselines” discusses the implications and insights derived from our findings. Finally, “Discussion” concludes the paper by summarizing our contributions and suggesting directions for future research. We utilize pre-trained Urdu Roberta embeddings to generate deep contextualized representations for T-R and P-D in Urdu, improving model accuracy on ambiguous input sentences. We utilize pre-trained Urdu Roberta embeddings to generate deep contextualized representations for T-R and P-D in Urdu, improving model accuracy on ambiguous input sentences. We introduce a robust hybrid confidence-based filtering mechanism to handle noisy pseudo-labeled data, enhancing the quality of training samples. We introduce a robust hybrid confidence-based filtering mechanism to handle noisy pseudo-labeled data, enhancing the quality of training samples. Our unified sentence-level and word-level DA techniques, including BT and MWP, enrich the dataset and enhance model robustness. Our unified sentence-level and word-level DA techniques, including BT and MWP, enrich the dataset and enhance model robustness. We propose SGSSL, a novel Siamese architecture combining GCN, GAT, TE, and MHSA for T-R and P-D tasks, leveraging both labeled and unlabeled data. We propose SGSSL, a novel Siamese architecture combining GCN, GAT, TE, and MHSA for T-R and P-D tasks, leveraging both labeled and unlabeled data. We provide a comprehensive analysis of the impact of DA and HCBF techniques and various model components on the performance of Urdu T-R and P-D tasks. We provide a comprehensive analysis of the impact of DA and HCBF techniques and various model components on the performance of Urdu T-R and P-D tasks. Empirical evaluations across six benchmark datasets show significant performance improvements over state-of-the-art (SOTA) baseline methods in SSL text classification. Empirical evaluations across six benchmark datasets show significant performance improvements over state-of-the-art (SOTA) baseline methods in SSL text classification. Related works In recent years, DNNs have significantly advanced NLP, particularly in tasks like T-R and P-D, with notable progress in high-resource languages such as English. This progress has been driven by the development of sophisticated models and the availability of large datasets. However, advancements remain limited for under-resourced languages like Urdu due to the scarcity of annotated corpora, as shown in Table 1. The proliferation of machine-generated content, especially through models like GPT-4, has amplified the need for advanced T-R and P-D techniques. P-D, crucial for detecting semantic equivalence, is essential for maintaining originality amid the increasing use of AI-generated text. Recent research leveraging the Turku Paraphrase Corpus has shown that span-based models outperform sentence-level methods like BERT and SBERT, particularly in token-level F1 scores, enhancing paraphrase retrieval in low-resource settings19. Despite these advances, challenges persist in distinguishing human-generated from machine-generated paraphrases, as the former tend to exhibit greater semantic complexity. While modern transformer models like BERT and T5 surpass traditional approaches, the need for more sophisticated machine-generated datasets remains, particularly for under-resourced languages like Urdu. Short text similarity and classification have been widely explored in Natural Language Processing, particularly due to challenges such as limited context, ambiguity, and sparse features. Traditional machine learning approaches, including Support Vector Machines and Random Forests, rely on handcrafted features but often fail to capture deep semantic relationships. Deep learning models have significantly improved performance by learning distributed representations. Recurrent Neural Networks (RNNs), especially Long Short-Term Memory (LSTM) and Bidirectional LSTM (BiLSTM), are widely used for capturing sequential dependencies in text. Faseeh et al. proposed a BiLSTM-based model with attention mechanisms and multiple pre-trained embeddings (GloVe, FastText, Paragram), demonstrating strong performance in duplicate question detection by effectively modeling semantic similarity20. Hybrid architectures combining CNN and RNN have also shown promising results. CNNs extract local spatial features, while LSTMs capture long-range dependencies. In this direction, Alkaabi et al. introduced a dual attention-based hybrid deep learning framework that integrates CNN and BiLSTM with location-based and co-attention mechanisms to improve feature extraction in short texts21. Their model addresses the limitations of short text by combining spatial and sequential representations, achieving high accuracy while maintaining computational efficiency. Attention mechanisms have further enhanced model performance by focusing on important textual features. Various studies incorporate self-attention and cross-attention to improve semantic understanding. Additionally, knowledge-enhanced attention models utilize external information to enrich representations and handle ambiguity in short texts. Recently, multimodal approaches have gained attention for leveraging complementary information beyond text. Jiang et al. proposed a dual-similarity enhanced hybrid orthogonal fusion framework that integrates textual and visual features using self-attention and cross-modal attention mechanisms22. This approach improves semantic representation and reduces feature redundancy, though challenges such as modality imbalance and semantic distortion remain. Ma et al.23 proposed an LLM-Enhanced Decoupled Learner and Course Modeling (LE-DLCM) framework for course recommendation, which introduces a dual-channel architecture to separately model learners and courses. Their approach utilizes knowledge-enhanced LLM-based course modeling to capture structural relationships among courses and interaction-enhanced learner modeling to address sparse user interactions. By incorporating similar learners and constructing course relation graphs, the model effectively mitigates cold-start and data sparsity issues, achieving significant improvements in recommendation performance. However, despite its effectiveness, the approach still faces challenges related to computational complexity, scalability, and reliance on prompt design for optimal performance. Graph-based semi-supervised learning Graph-based SSL has made significant strides, driving innovation across multiple domains by enhancing task performance through the effective use of both labeled and large amounts of unlabeled data. Recent advances, such as the integration of GCNs with deep learning techniques, have significantly boosted classification accuracy by maintaining a balance between global and local graph structures24. Additionally, the Dynamic Mutual Training method has further improved performance by refining noisy pseudo-labels. This approach leverages inter-model disagreements, proving beneficial in tasks like image classification and semantic segmentation25. However, despite these advances, Graph Neural Networks (GNNs) remain vulnerable to adversarial attacks. The EpoAtk method addresses this issue by employing an exploratory strategy to strengthen the resilience of GNNs in semi-supervised node classification scenarios26. For graph-level classification, GraphSpa offers a cost-effective alternative to traditional graph kernels, efficiently selecting informative unlabeled graphs to improve classification outcomes27. Another challenge in semi-supervised GNNs is the over-smoothing problem, where node features become indistinguishable. To counter this, the Graph Alignment Neural Network preserves feature diversity and structural distinctiveness, ensuring better task performance28. Innovations in SSL have also included methods like ScanMix, which integrates SSL with semantic clustering to handle high-noise environments, and the Cross-modal Graph Convolutional Network (MCGCN), which employs adversarial learning to extract both modality-specific and shared information, particularly useful in multimedia retrieval tasks29. The Adaptive Graph Smoothing Network offers another key advancement, dynamically adjusting edge weights to improve node classification accuracy, achieving impressive results on citation and Wiki networks30. Among the newer approaches is the Semi-Supervised Graph Contrastive Learning (SGCL) algorithm, which combines supervised and unsupervised learning techniques with a Confidence Propagation Algorithm, significantly enhancing pseudo-label accuracy and setting new benchmarks31. For time series classification, SSL methods that rely on pseudo-labeling and consistency regularization often face difficulties in capturing effective temporal representations. Solutions such as bidirectional consistency and temporal-aware contrastive learning aim to enhance intra-class compactness and inter-class separability, resulting in improved classification performance32. Additionally, GraphixMatch, which integrates the FixMatch framework with both weak and strong graph augmentations, has emerged as a leading SSL approach33. Similarly, the Masked Dual-Temporal Autoencoder addresses label sparsity by effectively leveraging both labeled and unlabeled data for semi-supervised time series classification34. Recent advancements in semi-supervised node classification have sought to overcome persistent challenges such as over-smoothing, insufficient supervision, and the integration of global structural information. The Graph Convolutional Kernel Machine (GCKM), for instance, incorporates unsupervised kernel layers for feature propagation, outperforming traditional GNNs, especially in situations with limited labeled data35. Another innovative approach, Optimized Simple Graph Convolution (OGC), revisits the cluster assumption in graph-based learning, offering a solution to the over-smoothing problem36. HyperGraph Transformer (HyperGT) further advances the field by incorporating global information through Transformer-based architectures, addressing over-smoothing in hypergraph models37. To prevent representation collapse, Mixed Graph Contrastive Networks (MGCN) use interpolation-based augmentation, which significantly enhances the discriminative power of node embeddings38. Nazir et al. proposed an enhanced extractive text summarization framework for low-resourced Urdu language, integrating both supervised and unsupervised models with feature engineering and hybrid deep learning architectures, achieving notable improvements in ROUGE scores compared to prior methods39. Their work highlights the importance of combining linguistic features with deep learning; however, challenges such as dataset scarcity and language-specific complexities still limit performance. In the domain of graph neural networks, Chitla et al. introduced GDNConv, a novel graph deformation network designed to improve robustness in noisy graph structures by incorporating dynamic edge weighting, attention-based deformation, and multi-level aggregation40. This approach significantly enhances representation learning and outperforms traditional GNN models, though it still faces challenges related to computational complexity in large-scale graphs. Similarly, Chitla et al. proposed a Modified Graph Convolutional Network (MGCN) that integrates adaptive edge weighting, residual connections, and scalable attention mechanisms to address over-smoothing and rigid graph structures41. While effective, these models indicate that balancing scalability and robustness remains an open problem in graph-based learning. For information retrieval tasks, Panchumarthi et al. developed a Modified Cross Encoder for two-stage passage ranking, leveraging pre-trained transformer models and a hybrid scoring mechanism combining sigmoid and cosine similarity42. Their model demonstrated improved ranking accuracy over traditional approaches such as BM25 and bi-encoders, though the computational cost of cross-encoder architectures continues to be a limitation for real-time systems. In text classification using graph-based learning, Wang et al. proposed PaSIG, a pre-trained semantic interaction-based inductive graph neural network that models semantic relationships between texts using heterogeneous graph structures and gated fusion mechanisms43. This method effectively improves classification accuracy and supports inductive learning; however, the complexity of graph construction and dependency on pre-trained models may impact efficiency. Furthermore, Yin et al. introduced TextGT, a double-view graph transformer that combines graph neural networks with transformer architectures to jointly model syntactic and sequential representations for aspect-based sentiment analysis44. This hybrid approach successfully alleviates the over-smoothing problem common in GNNs and improves performance, though integrating dual representations increases model complexity. Lastly, the Neighbours-to-Neighbours Contrastive GCN (NNC-GCN) improves classification performance through local neighborhood contrast learning without the need for complex architectures45. These innovations collectively push the boundaries of semi-supervised learning, tackling both local and global challenges in graph-structured data. They underscore the evolving potential of SSL to produce more accurate, robust, and efficient models across a variety of applications. Oversmoothing mechanism Oversmoothing in GCNs is a phenomenon that arises when multiple layers are stacked. This stacking results in the representations of nodes from distinct clusters becoming excessively similar, ultimately degrading the overall performance of the model. The problem of over-smoothing was first highlighted in the work of Li and colleagues46. Following this, Xu and others observed that the rate of oversmoothing can differ based on the characteristics of the nodes; specifically, nodes located near the edges of the graph tend to experience slower oversmoothing effects47. To address the over-smoothing issue, several solutions have been proposed. Notably, Kipf and collaborators, along with Xu and others, introduced methods that draw inspiration from residual networks. These approaches incorporate residual connections to help mitigate the over-smoothing effect48,49. Although these techniques demonstrate potential, they are not entirely sufficient for completely resolving the problem. Another approach to tackling over-smoothing is presented in the context of FAGCN50. Unlike previous methods that focus primarily on residual connections, FAGCN enhances node representation by aggregating information from first-order neighbors across both high and low frequencies. This strategy enables the central node to capture more informative features, thereby improving its distinction from neighboring nodes. By integrating diverse frequency information, FAGCN not only alleviates the effects of over-smoothing but also enriches node representations. This leads to enhanced model performance and clearer differentiation among clusters. The over-smoothing effect can significantly hinder a network’s ability to distinguish between different data points. Nodes, regardless of their unique attributes or community affiliations, tend to converge toward similar representations. This issue becomes particularly pronounced in semi-supervised learning scenarios, where a lack of labeled data presents a major challenge. Consequently, achieving high classification accuracy becomes increasingly difficult, as the subtle distinctions critical for effective learning are diminished. To combat the over-smoothing problem, the structure of the proposed model integrates TE with Graph GNNs. This design aims specifically to mitigate the adverse effects of over-smoothing, thereby improving the model’s capability to learn from data. Siamese networks and plagiarism detection Several studies have leveraged Siamese network architectures for P-D tasks in the context of low-resource languages. One example is the work on Indonesian poetry, where a Siamese MaLSTM model was used to detect sentence similarity51. The model achieved 91.6% accuracy in detecting semantic similarities between lines of poetry, demonstrating the effectiveness of Siamese networks in identifying plagiarism. This model combined word embeddings (FastText) with tokenization and stop-word removal techniques. Although successful, the authors noted the need for refinement in understanding poetic structures and styles, which are often more nuanced than traditional prose. A study by52 proposed a modularity-based Siamese Simple Graph Convolutional Network (MS-SGC) to address the over-smoothing issue in traditional GCNs. Over-smoothing occurs as the model depth increases, leading to diminished performance. Inspired by the modular structure of brain networks, the authors integrated spectral clustering to detect and preserve modularity in graphs, reducing cross-module connection weights and improving both locality and global balance. Extensive experiments showed that MS-SGC improved node classification accuracy by up to 2% on the Pubmed dataset, outperformed several SOTA models, and reduced computational complexity in node clustering and community prediction tasks. Additionally53, proposed an enhanced, attentive Siamese LSTM network for low-resource text plagiarism detection, particularly focusing on Tibetan-Chinese cross-lingual cases. The study addressed the challenges of limited training data by utilizing translation-based DA to expand the corpus, thus improving the model’s performance in detecting plagiarism, especially in cases involving semantic rewriting and translation-based plagiarism. The paper also introduced a pre-detection module that employs abstract document vectors to enhance detection efficiency. Experimental results demonstrated that the proposed model achieved strong performance, with a significant Pearson correlation coefficient in both Tibetan and Chinese text plagiarism detection tasks, highlighting its potential effectiveness in low-resource language settings. Urdu text reuse and paraphrase detection Urdu, being a low-resource language, has witnessed relatively few advancements in T-R and P-D. To the best of our knowledge, all previous efforts in this domain have relied on supervised learning (SL) approaches, and no attempts have been made to explore semi-supervised learning (SSL) in this area. One of the pioneering efforts was the development of the Urdu Paraphrase Plagiarism Corpus (UPPC)54, which contains 160 documents, including 75 paraphrased (P) and 65 non-paraphrased (NP) texts. Despite its value, UPPC has limitations, such as a small dataset size, a focus on celebrity-related texts, and controlled experimental conditions that may not represent real-world plagiarism scenarios. More recent efforts by Hafeez et al.12 introduced the Urdu Sentential Paraphrase (USP) corpus, comprising 4900 sentences with labeled paraphrased and non-paraphrased data. The researchers used a feature-fusion technique combining sentence transformers and n-grams, achieving an F1 score of 0.855, indicating its effectiveness in detecting paraphrases in Urdu text. A significant leap in Urdu P-D came with the development of the Semi-Automatically Generated Urdu Sentential Paraphrase Corpus (SUSPC)55, which employed DNNs-based models such as WENGO and D-TRAPPD. D-TRAPPD, a hybrid model combining convolutional neural networks (CNNs) and long short-term memory (LSTM) networks, significantly outperformed previous methods, achieving an F1 score of 96.80% for P-D. Aslam et al.7 proposed a BiLSTM-based framework for Urdu paraphrase detection to address morphological complexity and limited resources. They also developed the Urdu Paraphrased Corpus (UPC) and demonstrated improved performance over CNN, LSTM, and BERT models. This study laid the foundation for future research, emphasizing the importance of deep learning (DL) architectures in improving Urdu NLP tasks. In the context of T-R, significant progress has been made with corpora, such as the Urdu Short Text Reuse Corpus (USTRC)56 and the COUNTER corpus57. The USTRC contains 2684 manually labeled short text pairs, categorized into verbatim, paraphrased, and non-paraphrased texts, enabling improved T-R detection in Urdu. The study evaluating USTRC found that character n-gram overlap, combined with the J48 classifier, was the most effective approach for detecting T-R. On the other hand, the COUNTER corpus advanced T-R detection in Urdu news articles, with the GST mML1 model achieving an F1 score of 0.81 for binary classification. A more recent development is the creation of the UTRD-Phr-23 corpus58, a large-scale dataset containing 25001 manually annotated text pairs designed to train and evaluate T-R models. The authors proposed a novel model integrating eight pre-trained sentence transformers, achieving an F1 score of 0.63 in T-R tasks. This demonstrates the robustness of transformer-based models in detecting T-R in low-resource languages like Urdu. While supervised learning has been the dominant approach in Urdu T-R and P-D research, there is a growing interest in SSL due to its ability to leverage both labeled and unlabeled data. SSL has shown promise in reducing the reliance on large labeled datasets, often scarce in low-resource languages. However, SSL techniques face challenges, such as error propagation from noisy labeled data and managing noisy predictions59. Recent advancements in GNNs for SSL tasks provide new opportunities for improving T-R and P-D tasks in Urdu. GNNs, including GCNs and GATs, have successfully enhanced feature extraction and the learning of relationships between text nodes. Specifically, models such as Text-GCN60 and TextING61, which construct graphs representing documents and words, allow for the joint learning of word and document embeddings, thereby improving T-R and P-D tasks. A recent study62 that combined a Siamese Network with graph clustering techniques demonstrated improvements in detecting query similarities, achieving a 6% performance improvement on benchmark datasets. Despite significant advances in NLP for high-resource languages, Urdu remains under-researched in the domains of T-R and P-D. Although some progress has been made with corpora like UPPC, USP, and USTRC, the lack of large-scale, diverse datasets and benchmark models has limited the development of robust T-R and P-D systems for Urdu. The application of SSL, particularly in conjunction with GNNs, presents a promising direction for future research. This approach offers the potential to overcome current limitations and improve the performance of T-R and P-D tasks in low-resource languages.Table 1Summary of Urdu T-R and P-D corpora.TasksCorpusCorpus typeObscurity degreesSource textT-RCounterDocumentWholly derived (135), partially derived (288), non-derived (177)JournalismT-RUSTRCSentenceVerbatim (496), P (1329), NP (859)JournalismT-RUTRD-Phr-23PhrasalDerived (15,105), Non-derived (9,896)JournalismP-DUPPCDocumentP (75), NP (65)AcademicP-DUSPSentenceP (2,941), NP (1,959)AcademicP-DSUSPCSentenceP (854), NP (2,293)Academic Table 1Summary of Urdu T-R and P-D corpora.TasksCorpusCorpus typeObscurity degreesSource textT-RCounterDocumentWholly derived (135), partially derived (288), non-derived (177)JournalismT-RUSTRCSentenceVerbatim (496), P (1329), NP (859)JournalismT-RUTRD-Phr-23PhrasalDerived (15,105), Non-derived (9,896)JournalismP-DUPPCDocumentP (75), NP (65)AcademicP-DUSPSentenceP (2,941), NP (1,959)AcademicP-DSUSPCSentenceP (854), NP (2,293)Academic Summary of Urdu T-R and P-D corpora. Proposed framework In this section, we start by outlining the problem formulation. Next, we detail the proposed HBCF mechanism, which selectively incorporates high-confidence pseudo-labeled data into the training set. Following this, we describe our approach to edge construction and pre-processing. We then discuss the DA techniques employed. Finally, we provide a comprehensive explanation of the proposed SGSSL model. Problem formulation In the domain of T-R and P-D tasks, challenges persist due to sparse digital resources, especially in languages like Urdu. The proposed SGSSL model addresses low accuracy caused by limited annotated data and noise in pseudo-labeled samples by employing an HCBF strategy. SGSSL is an SSL method that integrates a DA technique to enhance the training dataset and improve overall performance. It represents sentence pairs as graphs G=(V, E), leveraging contextual embeddings x from Urdu Roberta and establishing edges E based on a semantic similarity threshold θ =0.5.Fig. 1The proposed SGSSL model integrates GCN and GAT within a Siamese framework for SSL text classification. A bottleneck layer consolidates features from GCN and GAT before they are processed by TE modules enriched with MHSA, enhancing the model’s ability to capture intricate language patterns. Advanced DA techniques like BT and MWP and an HCBF strategy ensure robust training and high classification accuracy by selectively retaining high-confidence pseudo-labeled data. The proposed SGSSL model integrates GCN and GAT within a Siamese framework for SSL text classification. A bottleneck layer consolidates features from GCN and GAT before they are processed by TE modules enriched with MHSA, enhancing the model’s ability to capture intricate language patterns. Advanced DA techniques like BT and MWP and an HCBF strategy ensure robust training and high classification accuracy by selectively retaining high-confidence pseudo-labeled data. The model uses a Siamese network architecture with a bottleneck layer to distill critical features from the graph embeddings. It then enhances these representations using a TE module with MHSA, dropout, Root Mean Square Normalization (RMSNorm), and feedforward layers. This architecture refines semantic dependencies and improves feature extraction, enabling more accurate paraphrase classification and T-R instance detection. Figure 1 illustrates the contribution of the proposed methodology. The holistic approach of SGSSL mitigates data scarcity challenges in T-R and P-D tasks, which is particularly beneficial in resource-constrained language environments like Urdu. HCBF mechanism This study introduces a novel SGSSL approach with a hybrid confidence-based filtering (HCBF) mechanism, as shown in Algorithm 1. The purpose of HCBF is to improve pseudo-label quality by jointly considering three complementary signals: (i) predictive confidence, (ii) semantic similarity, and (iii) model uncertainty. In contrast to confidence-only filtering, this design reduces the chance of retaining overconfident but unreliable pseudo-labels. First, the SGSSL model \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {M}$$\end{document} is trained on the labeled set \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {L}$$\end{document}. The trained model is then used to generate pseudo-labels \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{y}_i$$\end{document} for each unlabeled sample \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x_i \in \mathcal {U}$$\end{document}. To explicitly quantify uncertainty, we use Monte Carlo (MC) Dropout63. During inference, dropout remains enabled and the same sample \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x_i$$\end{document} is forwarded through the model N times, producing a set of predictive probability vectors:1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned}{\bf p}_i^{(n)} = \mathcal {M}^{(n)}(x_i), \quad n=1,2,\dots ,N, \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {M}^{(n)}$$\end{document} denotes the nth stochastic forward pass. The predictive mean is computed as:2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \bar{\bf{p}}_i = \frac{1}{N} \sum _{n=1}^{N}{\bf p}_i^{(n)}. \end{aligned}$$\end{document}The pseudo-label is assigned using the maximum mean predictive probability:3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \hat{y}_i = \arg \max _k \bar{p}_{i,k}, \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\bar{p}_{i,k}$$\end{document} is the mean probability of class k. We define the probabilistic confidence as:4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} c_i = \max _k \bar{p}_{i,k}. \end{aligned}$$\end{document}To measure predictive uncertainty, we compute the variance of the MC predictions across stochastic forward passes:5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} u_i = \frac{1}{K}\sum _{k=1}^{K}\textrm{Var}\left( \{p_{i,k}^{(n)}\}_{n=1}^{N}\right) , \end{aligned}$$\end{document}where K is the number of classes. A lower value of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$u_i$$\end{document} indicates more stable predictions across MC samples and therefore higher reliability. In parallel, we compute a semantic similarity score \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$s_i$$\end{document} using Sentence Transformer embeddings. Let \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\bf e}_i$$\end{document} denote the embedding of sample \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x_i$$\end{document}. The semantic consistency score is obtained by cosine similarity between \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\bf e}_i$$\end{document} and the prototype embedding of its pseudo-labeled class:6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} s_i = \cos ({\bf e}_i,{\bf z}_{\hat{y}_i}), \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\bf z}_{\hat{y}_i}$$\end{document} is the centroid of labeled samples belonging to class \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hat{y}_i$$\end{document}. Finally, the hybrid confidence score is computed as:7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} h_i = \alpha c_i + \beta s_i - \gamma u_i, \end{aligned}$$\end{document}where α, β, and γ are balancing coefficients satisfying \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha +\beta +\gamma =1$$\end{document}. The score increases with higher predictive confidence and stronger semantic agreement, while it decreases with larger predictive uncertainty. After computing \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$h_i$$\end{document} for all unlabeled samples, we retain only the top-scoring pseudo-labeled instances using a threshold τ:8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \mathcal {U}_{\text {high}} = \{x_i \in \mathcal {U} \mid h_i > \tau \}. \end{aligned}$$\end{document}This selected subset is then used with consistency regularization, and finally merged with the original labeled set to retrain the SGSSL model. In this way, HCBF acts as a reliability-aware pseudo-label selection mechanism that is more robust than filtering based only on prediction confidence. Uncertainty estimation and hybrid score construction To clarify the “hybrid” nature of HCBF, the uncertainty term is explicitly estimated using Monte Carlo Dropout rather than a deterministic confidence proxy. Specifically, dropout is kept active at inference time, and each unlabeled instance is evaluated through N stochastic forward passes. This produces a distribution of predicted class probabilities, from which predictive variance is computed as the uncertainty estimate. Therefore, the uncertainty term captures the model’s epistemic uncertainty, i.e., instability caused by uncertainty in model parameters when labeled data are limited. This choice is particularly suitable for semi-supervised learning because pseudo-labels may appear highly confident in a single forward pass while remaining unstable across repeated stochastic passes. By including the variance-based uncertainty term in Eq. (7), HCBF penalizes such unreliable samples even when their maximum softmax confidence is high. Consequently, the filtering mechanism becomes more conservative and better aligned with pseudo-label reliability. The final hybrid confidence score integrates three complementary signals: (i) confidence \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$c_i$$\end{document}, which reflects the model’s belief in the assigned pseudo-label; (ii) semantic similarity \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$s_i$$\end{document}, which captures consistency with class-level semantic structure; and (iii) uncertainty \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$u_i$$\end{document}, which quantifies prediction instability under MC Dropout. This formulation makes the filtering criterion mathematically transparent and explains why HCBF performs better than confidence-only pseudo-label selection. Algorithm 1Hybrid confidence-based filtering (HCBF). Hybrid confidence-based filtering (HCBF). Constructing graph edges In the edge creation process for each pair of sentences, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$k \in \{1, 2\}$$\end{document}, contextual embeddings \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$E_k$$\end{document} for each word are computed using pre-trained Urdu RoBERTa models. An empty adjacency matrix \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$A_k$$\end{document} is initialized for each sentence. For each pair of words i and j in the sentence, the cosine similarity \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$s_{ij}$$\end{document} between their embeddings \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$E_k[i]$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$E_k[j]$$\end{document} is computed as Eq. (1):9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} s_{ij} = \frac{E_k[i] \cdot E_k[j]}{\Vert E_k[i]\Vert \Vert E_k[j]\Vert } \end{aligned}$$\end{document}The adjacency matrix element \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$A_k[i,j]$$\end{document} is defined as in Eq. (2):10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} A_k[i,j] = {\left\{ \begin{array}{ll} 1 & \text {if } s_{ij} \ge \mathcal {T}, \\ 0 & \text {otherwise}. \end{array}\right. } \end{aligned}$$\end{document}Here, Eq. (2) defines \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$s_{ij}$$\end{document} as the similarity score between words i and j, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {T}$$\end{document} is the threshold value, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$A_k[i,j]$$\end{document} is the adjacency matrix element for the k-th graph. This process is applied to both sentences in the pair for T-R and P-D tasks. After constructing the adjacency matrices, GAT and GCN are used to process these graphs. GAT leverages attention mechanisms to assign different weights to the nodes in the neighborhood, enhancing the model’s ability to focus on the most relevant parts of the graph. Conversely, GCN performs convolution operations over the graph structure, aggregating features from neighboring nodes to generate a comprehensive graph representation. Data processing To efficiently analyze text, it is essential to convert raw text into numerical data. We enhanced T-R and P-D models with high-quality training data and diverse DA through BT and MWP. Text preprocessing included cleaning input by removing punctuation, whitespace, numbers, and stop words, which optimized tasks and reduced algorithm iterations. We used Python tools NLTK (https://www.nltk.org/) and SpaCy (https://spacy.io/) to handle complex text and ensure the semantic meaning of sentences was preserved. The preprocessed text was tokenized, integer-encoded, and padded into fixed-length vectors for model training, ensuring the models perform well, especially in complex languages like Urdu. Postgraduate experts selected the most suitable augmented samples to maximize model accuracy. Data augmentation DA entails the application of semantically invariant transformations as a regularization technique. These transformations generate additional data samples while preserving the integrity of class labels. The scope of DA goes beyond image classification and extends to various domains such as speech recognition, machine translation, and, more recently, text classification. Augmenting data is a critical approach to improving the overall adaptability of models and preventing overfitting. We present our enhanced approach for T-R and P-D by discussing the proposed DA strategy. The annotation is costly and time-consuming, while DL approaches demand a large corpus. To address this problem, we developed methods for generating additional annotations efficiently. Our study utilized two adopted methods: BT and MWP. Back-translation DA enhances DNN performance by generating new samples from existing data, which is especially useful when dealing with small datasets or low-resource languages. BT is a common DA technique where sentences are translated from one language to another and then back to the original language to create paraphrases. In this study, we apply BT to the T-R and P-D corpus by translating Urdu sentences into English and back into Urdu, as this method effectively preserves meaning. We employed multiple BT strategies using the Transformers library, which includes machine translation models like Marian (https://huggingface.co/docs/transformers/en/model_doc/marian/) and Flore 200 (https://github.com/facebookresearch/flores/), as well as the Google Translation API. For each original sentence \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$S_{\text {Original}}$$\end{document}, three back-translated sentences were generated as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$BT = \{S_{t1}, S_{t2}, S_{t3}\}$$\end{document}. Mask word prediction We present a DA technique utilizing Masked Language Modeling (MLM) for Urdu based on two pre-trained models from the Hugging Face library: Roberta Urdu (https://huggingface.co/urduhack/roberta-urdu-small/) and Albert Urdu (https://huggingface.co/sibt-rj/albert-large-urdu/). The Roberta Urdu model is highly effective in processing Urdu news articles, featuring 125.98 million parameters and a large vocabulary of 52,000. It is designed for tasks like sentiment analysis and summarization within the Urdu news domain. Conversely, the Albert Urdu model, with 11.69 million parameters and a vocabulary of 30,020, is optimized for educational and encyclopedic content, particularly Urdu Wikipedia articles. Both models share an architecture of 12 attention heads and a hidden dimension of 768, facilitating broad contextual understanding. Our DA methodology begins by masking random words in sentences from our Urdu corpus, replacing them with a token but keeping the target word intact. This approach allows the models to predict suitable replacements from five contextually relevant options, enhancing linguistic comprehension. We augment this by increasing batch sizes and prolonging training durations, further tuning the models with diverse parameters to optimize performance. By predicting ten potential replacements (five from each model) for every masked word and subsequently transforming these predictions into embeddings using the Roberta-Urdu model, we can compute similarity scores between the embeddings of the original masked words and their predicted counterparts. The most contextually appropriate word is then selected to replace the original, creating varied training sentences that significantly boost the efficacy of Urdu T-R and P-D tasks. SGSSL framework We propose a novel SSL approach named SGSSL, as represented in Fig. 2, utilizing Siamese GCN and GAT to classify sentence pairs effectively. Each sentence pair is embedded using the Urdu Roberta model, and edges in the graph are constructed based on semantic similarities between words, enabling robust contextual representation. To comprehensively evaluate our model, we conduct experiments using various combinations of GCN and GAT architectures (i.e., GAT-GAT, GCN-GCN, GAT-GCN, and GCN-GAT) across six widely used T-R and P-D benchmark datasets. Our SSL approach integrates labeled and unlabeled data to enhance the model’s learning process and improve performance in scenarios with sparsely labeled data. The concatenated features from the Siamese network are processed through a bottleneck layer, implemented as a linear transformation, to reduce dimensionality and mitigate the risk of overfitting. This dimensionality reduction is crucial for handling high-dimensional data efficiently. Following this, the features are fed into a TE module, which includes MHSA, dropout, RMSNORM, and Feedforward layers. These components collectively enhance the model’s ability to capture intricate dependencies and patterns within the data. The final prediction is achieved through a SoftMax layer, providing a probabilistic classification output, as detailed in Algorithm 2. Our proposed method demonstrates significant advantages, including improved accuracy, robustness to sparse data, and efficient handling of complex sentence structures. The SSL framework boosts performance by leveraging unlabeled data and ensures scalability and adaptability to various datasets and applications. This innovative approach holds substantial impact potential for advancing sentence pair classification and related NLP tasks. Graph convolutional networks GCNs play a crucial role in GNNs, contrasting with conventional graph embedding methods by conducting end-to-end training to derive compact vector representations. These representations are utilized for tasks like node classification, graph classification, and link prediction. Over recent years, integrating neural network techniques has driven substantial progress in applying GNNs to recommender systems, facilitating the modeling of associative connections between users and items where traditional algorithms face challenges. Moreover, GNNs have shown significant effectiveness in tasks within NLP. Originally proposed by Kipf et al.48, GCNs operate directly on graph structures, generating embedding vectors for nodes based on their neighborhood properties and learning graph-based structural information. The GAT enhances this by learning node attribute associations through attention mechanisms, improving classification capabilities. A GCN is a neural network architecture designed to handle graph-structured data. It operates on a graph G = (V, E), where V is the set of nodes, and E is the set of edges. Each node v in the graph has a corresponding feature vector \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x_v$$\end{document}, typically initialized with pre-trained or context-specific embeddings. The adjacency matrix A represents the connection between nodes, while the degree matrix D contains the number of edges connected to each node. The normalized adjacency matrix \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\tilde{A} = D^{-\frac{1}{2}}AD^{-\frac{1}{2}}$$\end{document} scales the feature vector appropriately. A GCN layer combines the features of a node and its neighbors using a spectral filter. The output \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$L^{(1)}$$\end{document} of a single-layer GCN can be expressed as in Eq. (3):11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} L^{(1)} = \sigma (\tilde{A}XW^{(0)}) \end{aligned}$$\end{document}Where X is the feature matrix, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$W^{(0)}$$\end{document} is a trainable weight matrix, and σ is an activation function such as ReLU. Stacking multiple GCN layers allows the model to capture information from progressively larger neighborhoods, making it particularly effective for T-R and P-D. For this study, we implement a two-layer GCN. Initially, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$H^{(0)}$$\end{document} is the feature matrix. The first layer, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$H^{(1)}$$\end{document}, captures the intermediate node embeddings. The second layer, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$H^{(2)}$$\end{document}, provides the final node embeddings used for model predictions. We apply ReLU activation in the first layer and log softmax in the second layer. To mitigate overfitting, a dropout layer is inserted between the two layers. We minimize the cross-entropy loss over the training dataset during training, incorporating the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$L_2$$\end{document}-norm for regularization. Formally, the loss function is defined as:12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} L = - \sum _{i \in D_{train}} \sum _{j=1}^{C} Y_{ij} \log P_{ij} + \lambda \Vert \theta \Vert ^2 \end{aligned}$$\end{document}Here, C denotes the number of classes, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$D_{train}$$\end{document} is the training dataset, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Y_{ij}$$\end{document} is the label matrix, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P_{ij}$$\end{document} is the predicted probability matrix, and λ is the regularization factor. Graph attention networks GAT differs significantly from GCN and other popular GNN architectures, which treat all neighbors equally using aggregation methods like mean and max-pooling. GAT64 introduces an attention mechanism that assigns different weights to nodes within a neighborhood, allowing the model to focus on the most relevant nodes. The core of GAT is a scoring function \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$a: \mathbb {R}^d \times \mathbb {R}^d \rightarrow \mathbb {R}$$\end{document} that computes a score for each edge \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$(E_i, E_j)$$\end{document}, indicating the significance of neighbor \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$E_j$$\end{document}’s features to node \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$E_i$$\end{document}:13\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \alpha (E_i, E_j) = \text {LeakyReLU}\left( a^T [Wh_i \parallel Wh_j]\right) \end{aligned}$$\end{document}Here, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$a \in \mathbb {R}^{2d}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$W \in \mathbb {R}^{d \times d}$$\end{document} are trainable parameters, and ∥ denotes vector concatenation. These attention scores are normalized across all neighbors using a Softmax function, which is defined as follows:14\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} a_{ij} = \frac{\exp (\alpha (E_i, E_j))}{\sum _{k \in N_i} \exp (\alpha (E_i, E_k))} \end{aligned}$$\end{document}The coefficient matrix \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$A \in \mathbb {R}^{N \times N}$$\end{document} captures these attention scores, with entries \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$a_{ij}$$\end{document} if \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha (E_i, E_j) \in \epsilon$$\end{document} and 0 otherwise. GAT then computes a weighted average of the transformed features of the neighbor nodes, followed by a non-linearity σ, to produce the new representation of node \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$E_i$$\end{document}:15\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} h_i = \sigma \left( \sum _{j \in N_i} a_{ij} \cdot Wh_j\right) \end{aligned}$$\end{document}In this context, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$h_i$$\end{document} denotes the representation of the node \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$E_i$$\end{document} in the next layer. In a GAT layer, multiple attention heads can capture different aspects of neighborhood information. The output from each head is concatenated or averaged and passed through a non-linear activation function.Fig. 2Proposed siamese SGSSL architecture. Proposed siamese SGSSL architecture. Transformer encoder The Transformer model employs TE blocks to efficiently convert vectors from \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathbb {R}^{t \times \text {dim}}$$\end{document} for sequence processing. Adding the TE module to our SGSSL approach addresses the limitations commonly observed in GNNs, such as over-smoothing17, where node features become overly similar in deeper GNN layers. Our study introduces encoder blocks composed of MHSA, dropout, RMSNorm, and SwiGLU layers, which enhance feature extraction and model stability. Algorithm 2Proposed SGSSL architecture. Proposed SGSSL architecture. The MHSA mechanism correlates Query (Q), Key (K), and Value (V) components to capture complex dependencies across nodes, mitigating the over-smoothing effect by maintaining both local and global feature distinctions. The adjustable MHSA layers in the encoder allow flexible weighting of node features, facilitating the effective extraction of structural and contextual information in graph data. Additionally, the encoder’s weight matrices and activation functions optimize the attention and feedforward networks, making the model more robust in handling complex graph structures. Overall, the Transformer’s design, incorporating TE blocks with MHSA, dropout, RMSNorm, and feedforward layers, overcomes the limitations of traditional GNNs and enhances feature extraction and stability, making it highly effective for Urdu T-R and P-D tasks requiring long-range relationship analysis and segment prioritization. Bottleneck feature-refinement layer To improve feature efficiency and reduce model complexity, we introduce a bottleneck feature-refinement layer within the SGSSL framework. This layer performs dimensionality reduction on high-dimensional node embeddings obtained from the Siamese GNN encoder before passing them to subsequent modules. Formulation: let \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\bf h}_i \in \mathbb {R}^{d}$$\end{document} denote the input feature representation of node i. The bottleneck layer applies a linear transformation defined as:16\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned}{\bf z}_i ={\bf W}_b{\bf h}_i +{\bf b}_b \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\bf W}_b \in \mathbb {R}^{d' \times d}$$\end{document} is the projection matrix, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\bf b}_b \in \mathbb {R}^{d'}$$\end{document} is the bias term, and d' < d represents the reduced dimensionality. Normalization and regularization: to improve stability and generalization, the projected features are optionally processed using layer normalization and dropout:17\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \tilde{{\bf z}}_i = \text {Dropout}(\text {LayerNorm}({\bf z}_i)) \end{aligned}$$\end{document}Design choice: unlike conventional bottleneck layers that employ non-linear activation functions, we adopt a linear bottleneck to preserve feature interpretability and avoid information distortion. This design reduces computational overhead while maintaining essential semantic information. Integration in SGSSL: the bottleneck layer is placed between the Siamese GNN encoder and the Transformer Encoder (TE) module. It compresses redundant features, enabling more efficient attention computation in the subsequent TE layer. Impact: as demonstrated in the ablation study (Table 10), the bottleneck layer yields an average improvement of 1.2% in F1 score while reducing feature dimensionality by approximately 32%, contributing to both efficiency and performance gains (Table 2).Table 2Configuration of the bottleneck feature-refinement layer.ComponentSpecificationInput dimension (d)768 (Urdu RoBERTa embeddings)Output dimension (d')256TransformationLinear (fully connected layer)Activation functionNone (linear bottleneck)NormalizationLayerNormRegularizationDropout (rate = 0.1)Parameter reduction∼32% Table 2Configuration of the bottleneck feature-refinement layer.ComponentSpecificationInput dimension (d)768 (Urdu RoBERTa embeddings)Output dimension (d')256TransformationLinear (fully connected layer)Activation functionNone (linear bottleneck)NormalizationLayerNormRegularizationDropout (rate = 0.1)Parameter reduction∼32% Configuration of the bottleneck feature-refinement layer. Experiments and analysis Datasets We evaluated our approach on six publicly available benchmark datasets covering Urdu text reuse detection (T-R) and paraphrase detection (P-D) at document, sentence, and phrasal levels. These datasets were selected to ensure diversity in text granularity, task formulation, and label complexity. Specifically, UPPC and COUNTER contain document-level pairs, USP, SUSPC, and USTRC contain sentence-level pairs, and UTRD-Phr-23 focuses on phrase-level similarity. Among these datasets, UPPC, USP, SUSPC, and UTRD-Phr-23 are evaluated in the binary setting, whereas COUNTER and USTRC are evaluated in both binary and ternary settings. The ternary setting is more challenging because it requires discrimination among fine-grained categories such as verbatim, partially paraphrased, and paraphrased. To ensure a uniform semi-supervised learning (SSL) protocol, each dataset was partitioned into a testing subset (10%) and a training subset (90%). The training subset was further divided into 20% labeled and 80% unlabeled samples. The unlabeled subset was incorporated into training through SSL, while the labeled subset was used to guide supervised optimization. In addition, data augmentation (DA) was applied only to the training portion of the data, thereby preserving the integrity of the test set. Table 3 summarizes the corpus size, SSL split, graph statistics, and number of classes for each benchmark, while Table 4 provides a more detailed description of the dataset source type, task formulation, text granularity, and label setting.Table 3Summary of the benchmark datasets utilized in our experiments.Corpus nameText length#Sentences#Total sentences (with DA)#Training (labeled)#Training (unlabeled)#Testing#Nodes#Edges#ClassesUPPCDocument160480863464815,77325,7052USPSentence4,90014,7002,3529,4081,47044,78394,8762SUSPCSentence3,1479,4411,5016,0051,47037,47376,7442COUNTERDocument1,2003,6005762,304360131,771422,3483USTRCSentence2,6848,0041,2815,12340161,685144,5363UTRD-Phr-23Phrasal25,00175,00312,00148,0027,501178,391830,4782Table 4Detailed description of the benchmark datasets used in this study.DatasetTaskGranularitySource typeLabel settingRemarksUPPCParaphrase detectionDocument-levelPublic Urdu benchmark corpusBinaryLong-form text pairs for semantic similarity analysisUSPParaphrase detectionSentence-levelPublic Urdu benchmark corpusBinarySentence pairs with lexical and semantic variationSUSPCParaphrase detectionSentence-levelPublic Urdu benchmark corpusBinaryShort Urdu sentence pairs with paraphrase/non-paraphrase labelsCOUNTERText reuse detectionDocument-levelPublic Urdu benchmark corpusBinary + TernaryIncludes finer-grained reuse distinctions in ternary settingUSTRCText reuse detectionSentence-levelPublic Urdu benchmark corpusBinary + TernaryContains exact, partial, and paraphrased reuse patternsUTRD-Phr-23Paraphrase detectionPhrase-levelPublic Urdu benchmark corpusBinaryVery short text pairs, making contextual modeling more difficult Table 3Summary of the benchmark datasets utilized in our experiments.Corpus nameText length#Sentences#Total sentences (with DA)#Training (labeled)#Training (unlabeled)#Testing#Nodes#Edges#ClassesUPPCDocument160480863464815,77325,7052USPSentence4,90014,7002,3529,4081,47044,78394,8762SUSPCSentence3,1479,4411,5016,0051,47037,47376,7442COUNTERDocument1,2003,6005762,304360131,771422,3483USTRCSentence2,6848,0041,2815,12340161,685144,5363UTRD-Phr-23Phrasal25,00175,00312,00148,0027,501178,391830,4782 Summary of the benchmark datasets utilized in our experiments. Table 4Detailed description of the benchmark datasets used in this study.DatasetTaskGranularitySource typeLabel settingRemarksUPPCParaphrase detectionDocument-levelPublic Urdu benchmark corpusBinaryLong-form text pairs for semantic similarity analysisUSPParaphrase detectionSentence-levelPublic Urdu benchmark corpusBinarySentence pairs with lexical and semantic variationSUSPCParaphrase detectionSentence-levelPublic Urdu benchmark corpusBinaryShort Urdu sentence pairs with paraphrase/non-paraphrase labelsCOUNTERText reuse detectionDocument-levelPublic Urdu benchmark corpusBinary + TernaryIncludes finer-grained reuse distinctions in ternary settingUSTRCText reuse detectionSentence-levelPublic Urdu benchmark corpusBinary + TernaryContains exact, partial, and paraphrased reuse patternsUTRD-Phr-23Paraphrase detectionPhrase-levelPublic Urdu benchmark corpusBinaryVery short text pairs, making contextual modeling more difficult Detailed description of the benchmark datasets used in this study. As shown in Table 4, the selected benchmarks cover complementary Urdu T-R and P-D scenarios, ranging from document-level semantic similarity to short phrase-level matching. This diversity allows us to evaluate the SGSSL framework under varying levels of contextual richness and classification difficulty. Data preprocessing and reproducibility protocol Before model training, all datasets were processed using a unified preprocessing pipeline to ensure consistency across experiments. First, all text pairs were normalized to a consistent Unicode form to reduce variation caused by Urdu script encoding differences. We then removed extra white spaces, normalized punctuation, and filtered malformed or empty samples. Duplicate text pairs, if present, were removed to avoid unintended information leakage across training and testing partitions. For sentence- and phrase-level datasets, token sequences were preserved as much as possible in order to retain lexical and syntactic variation important for paraphrase and text reuse detection. For document-level datasets, text was segmented into the corresponding pair representation used by the encoder. No test sample was augmented or modified during preprocessing. All data augmentation techniques, including back translation (BT) and masked word prediction (MWP), were applied only to the training data after the train/test split. To improve reproducibility, all baseline models and all SGSSL variants were evaluated using the same preprocessing pipeline, identical train/test partitions, and the same labeled/unlabeled division strategy. The labeled and unlabeled subsets were created only from the training split, ensuring that no test instances contributed to graph construction, pseudo-label generation, or augmentation. This protocol preserves experimental fairness and prevents data leakage. After preprocessing, each dataset was converted into a graph representation for SGSSL. Nodes correspond to textual units and their augmented counterparts, while edges were constructed using semantic similarity scores derived from contextual embeddings. The same graph construction procedure and thresholding strategy were applied consistently across datasets unless otherwise noted in the sensitivity analysis. Baseline methods We selected baseline models categorized into graph-based and large language models (LLMs) to ensure a comprehensive comparison. The baseline models include: BERT65: a deep, multi-layer bidirectional transformer achieving SOTA performance across various tasks.ALBERT66: a lightweight version of BERT, reducing model size while maintaining high performance.RoBERTa67: a refined adaptation of BERT with enhanced pretraining methodologies.XLM-RoBERTa68: a multilingual variant trained on data across 100 languages.ELECTRA69: introduces token replacement during training for greater efficiency.Simple transformer70: uses MHSA to capture long-range dependencies in text.mT571: a multilingual text-to-text transformer model for various NLP tasks.GPT-272: employs a Transformer-based decoder architecture for generative language modeling.LLaMA 373: meta’s SOTA pre-trained LLaMA 3 Urdu model available on Hugging Face as the baseline for our Urdu classification task.TextGT74: applies transformer architecture to graph-structured data for tasks like paraphrase detection.TextING61: tackles inductive text classification using GNNs.Text-GCN60: transforms text into a graph structure and applies GCN for classification.HGAT75: uses dual-level attention on heterogeneous networks for text embedding and classification.PTE76: encodes labeled data and word co-occurrence into a heterogeneous network for text classification.HGAN77: embeds heterogeneous networks using attention mechanisms across multiple sub-networks.Text-GAT64: enhances GCN by integrating attention to prioritize important nodes.All models, including GPT-based approaches, were tested on the same six datasets used for the SGSSL model, employing the same data partitioning strategy to ensure fairness. We included GPT-based models by leveraging pre-trained GPT architectures and fine-tuning them on our dataset. Despite their inclusion, GPT-based models required significant computational resources and were limited by the scarcity of large labeled datasets in Urdu for effective fine-tuning. Additionally, our research focuses on graph-based architectures that are better aligned with the specific requirements of the T-R and P-D tasks. The results in “Comparison with baselines” (Table 18) show that our proposed SGSSL framework outperforms all baseline models, demonstrating its ability to leverage unlabeled data and optimize performance in resource-constrained environments. BERT65: a deep, multi-layer bidirectional transformer achieving SOTA performance across various tasks. BERT65: a deep, multi-layer bidirectional transformer achieving SOTA performance across various tasks. ALBERT66: a lightweight version of BERT, reducing model size while maintaining high performance. ALBERT66: a lightweight version of BERT, reducing model size while maintaining high performance. RoBERTa67: a refined adaptation of BERT with enhanced pretraining methodologies. RoBERTa67: a refined adaptation of BERT with enhanced pretraining methodologies. XLM-RoBERTa68: a multilingual variant trained on data across 100 languages. XLM-RoBERTa68: a multilingual variant trained on data across 100 languages. ELECTRA69: introduces token replacement during training for greater efficiency. ELECTRA69: introduces token replacement during training for greater efficiency. Simple transformer70: uses MHSA to capture long-range dependencies in text. Simple transformer70: uses MHSA to capture long-range dependencies in text. mT571: a multilingual text-to-text transformer model for various NLP tasks. mT571: a multilingual text-to-text transformer model for various NLP tasks. GPT-272: employs a Transformer-based decoder architecture for generative language modeling. GPT-272: employs a Transformer-based decoder architecture for generative language modeling. LLaMA 373: meta’s SOTA pre-trained LLaMA 3 Urdu model available on Hugging Face as the baseline for our Urdu classification task. LLaMA 373: meta’s SOTA pre-trained LLaMA 3 Urdu model available on Hugging Face as the baseline for our Urdu classification task. TextGT74: applies transformer architecture to graph-structured data for tasks like paraphrase detection. TextGT74: applies transformer architecture to graph-structured data for tasks like paraphrase detection. TextING61: tackles inductive text classification using GNNs. TextING61: tackles inductive text classification using GNNs. Text-GCN60: transforms text into a graph structure and applies GCN for classification. Text-GCN60: transforms text into a graph structure and applies GCN for classification. HGAT75: uses dual-level attention on heterogeneous networks for text embedding and classification. HGAT75: uses dual-level attention on heterogeneous networks for text embedding and classification. PTE76: encodes labeled data and word co-occurrence into a heterogeneous network for text classification. PTE76: encodes labeled data and word co-occurrence into a heterogeneous network for text classification. HGAN77: embeds heterogeneous networks using attention mechanisms across multiple sub-networks. HGAN77: embeds heterogeneous networks using attention mechanisms across multiple sub-networks. Text-GAT64: enhances GCN by integrating attention to prioritize important nodes. Text-GAT64: enhances GCN by integrating attention to prioritize important nodes. Evaluation methodology In this study, we employed Accuracy (ACC) and Macro-F1 as our evaluation metrics. Accuracy, denoted as ACC, measures the proportion of correctly predicted instances out of the total number of instances. In the context of multi-class classification, the F1 score can be decomposed into Micro-F1 and Macro-F1, each calculated using different methodologies. Micro-F1 aggregates precision and recall across all classes, treating them collectively as a single entity. It is calculated as:18\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} & \text {Micro-Precision} = \frac{\sum _i TP_i}{\sum _i (TP_i + FP_i)} \end{aligned}$$\end{document}19\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} & \text {Micro-Recall} = \frac{\sum _i TP_i}{\sum _i (TP_i + FN_i)} \end{aligned}$$\end{document}20\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} & \text {Micro-F1} = \frac{2 \times (\text {Micro-Precision} \times \text {Micro-Recall})}{\text {Micro-Precision} + \text {Micro-Recall}} \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$TP_i$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$FP_i$$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$FN_i$$\end{document} represent true positives, false positives, and false negatives for class i, respectively. In contrast, Macro-F1 independently computes the F1 score for each class and averages these scores to derive an overall measure. This is given by:21\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} & F1_i = \frac{2 \times (\text {Precision}_i \times \text {Recall}_i)}{\text {Precision}_i + \text {Recall}_i} \end{aligned}$$\end{document}22\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} & \text {Macro-F1} = \frac{1}{n} \sum _{i=1}^{n} F1_i \end{aligned}$$\end{document}where n is the number of classes. This metric, Macro-F1, is particularly useful for evaluating classification performance in scenarios where the class distribution is imbalanced, as it treats each class equally, regardless of the number of instances. Parameters setting Our research comprehensively investigated the influence of different configurations on our system’s performance. We selected particular settings, detailed in 5. We executed our experiments on a Linux platform, employing PyTorch, Python 3.6, and an Nvidia T4 GPU with 16 GB of memory. Our model employs Urdu Roberta embedding for sentence processing, producing a 768-dimensional output.Table 5Hyper-parameter settings for six standard datasets.ParametersUPPCUSPSUSPCCOUNTERUSTRCUTRD-Phr-23GAT heads888888Batch size8321681664Embedding size768768768768768768Num TE heads444444Num TE blocks222222Learning rate0.000040.000050.000050.00040.000050.00005Epochs606060606060Weight decay0.00010.00010.00010.00010.00010.0001Dropout0.10.20.20.20.20.3OptimizerAdamWAdamWAdamWAdamWAdamWAdamWLoss functionCross-entropy loss Table 5Hyper-parameter settings for six standard datasets.ParametersUPPCUSPSUSPCCOUNTERUSTRCUTRD-Phr-23GAT heads888888Batch size8321681664Embedding size768768768768768768Num TE heads444444Num TE blocks222222Learning rate0.000040.000050.000050.00040.000050.00005Epochs606060606060Weight decay0.00010.00010.00010.00010.00010.0001Dropout0.10.20.20.20.20.3OptimizerAdamWAdamWAdamWAdamWAdamWAdamWLoss functionCross-entropy loss Hyper-parameter settings for six standard datasets. Experimental results and analysis We evaluated our SSL model on Urdu T-R and P-D tasks using six benchmark datasets for binary and ternary classification. The SGSSL framework, combining GCN, GAT, and a Siamese architecture with TE units and MHSA, was tested under various conditions, including with and without DA techniques (BT and MWP) and HCBF. Tables 6, 7, 8 and 9 present the results. Table 6 shows the performance of different models without DA, where the GAT-GAT+TE model excels, particularly in binary classification on the UPPC dataset with F1 and accuracy scores of 0.741 and 0.764, respectively. Table 7 illustrates significant performance improvements after integrating DA techniques. The GAT-GAT+TE model achieves an F1 of 0.759 and accuracy of 0.795 on the UPPC dataset, highlighting DA’s role in enhancing model generalization. Table 8 evaluates the impact of applying HCBF without DA, showing enhanced performance across all models compared to the SGSSL baselines. The GAT-GAT+TE model achieves an F1 of 0.782 and accuracy of 0.799 on the UPPC dataset, demonstrating HCBF’s effectiveness in reducing label noise. Table 9 reveals that combining HCBF and DA yields the most significant performance enhancements, with the GAT-GAT+TE configuration achieving an F1 of 0.796 and accuracy of 0.822 on the UPPC dataset. These results confirm that the combination of HCBF and DA significantly boosts SSL model performance across various datasets.Table 6SGSSL test Acc and F1 of different models on six standard datasets without DA.ModelsUPPCUSPSUSPCCOUNTER (binary)COUNTER (ternary)USTRC (binary)USTRC (ternary)UTRD-Phr-23F1AccF1AccF1AccF1AccF1AccF1AccF1AccF1AccGCN-GCN+TE0.6870.7070.7180.7200.6500.7280.5320.5420.5720.5760.5030.5310.5820.6050.3120.392GAT-GAT+TE0.7410.7640.7220.7580.7030.7430.5690.5750.6110.6360.5230.5630.6030.6230.3520.432GCN-GAT+TE0.7140.7170.7160.7350.6770.7320.5420.5520.5900.6040.5160.5520.5940.6150.3240.403GAT-GCN+TE0.7050.7240.7200.7370.6800.7310.5500.5550.5950.5990.5150.5500.5840.6170.3310.412Table 7SGSSL test Acc and F1 of different models on six standard datasets with DA.ModelsUPPCUSPSUSPCCOUNTER (binary)COUNTER (ternary)USTRC (binary)USTRC (ternary)UTRD-Phr-23F1AccF1AccF1AccF1AccF1AccF1AccF1AccF1AccGCN-GCN+TE0.7060.7510.7330.7500.6920.7310.5520.5500.5890.5900.5170.5390.5820.6450.3550.402GAT-GAT+TE0.7590.7950.7360.7630.7280.7680.5730.5840.6320.6440.5300.5750.6200.6730.3820.462GCN-GAT+TE0.7340.7720.7300.7550.7100.7400.5490.5600.6000.6240.5230.5620.6060.6500.3600.444GAT-GCN+TE0.7350.7800.7310.7530.7090.7390.5600.5660.6040.6220.5250.5630.6020.6560.3690.447Table 8SGSSL-HCBF test Acc and F1 of different models on six standard datasets without DA.ModelsUPPCUSPSUSPCCOUNTER (binary)COUNTER (ternary)USTRC (binary)USTRC (ternary)UTRD-Phr-23F1AccF1AccF1AccF1AccF1AccF1AccF1AccF1AccGCN-GCN+TE0.7290.7530.7630.7920.6660.6700.5400.5460.6010.6010.5170.5390.6300.6640.3780.457GAT-GAT+TE0.7820.7990.7940.8320.7290.7520.5750.5820.6400.6350.6020.6210.6660.6870.4050.486GCN-GAT+TE0.7590.7700.7750.8110.6970.7390.5590.5570.6110.6220.5890.6110.6400.6650.3890.477GAT-GCN+TE0.7510.7810.7700.8070.7400.7690.5690.5750.6520.6590.6010.6440.6850.7060.4670.502Table 9SGSSL-HCBF test Acc and F1 of different models on six standard datasets with DA.ModelsUPPCUSPSUSPCCOUNTER (binary)COUNTER (ternary)USTRC (binary)USTRC (ternary)UTRD-Phr-23F1AccF1AccF1AccF1AccF1AccF1AccF1AccF1AccGCN-GCN+TE0.7480.7750.7900.8010.7210.7480.5640.5590.6550.6560.5890.6240.6710.6960.4000.485GAT-GAT+TE0.7960.8220.8140.8520.7530.7870.5800.5910.6720.6750.6270.6590.6930.7220.4860.553GCN-GAT+TE0.7770.8050.8050.8240.7460.7790.5640.5730.6600.6620.5900.6450.6800.7070.4440.524GAT-GCN+TE0.7740.8000.8100.8310.7400.7690.5690.5750.6520.6590.6010.6440.6850.7060.4670.502 Table 6SGSSL test Acc and F1 of different models on six standard datasets without DA.ModelsUPPCUSPSUSPCCOUNTER (binary)COUNTER (ternary)USTRC (binary)USTRC (ternary)UTRD-Phr-23F1AccF1AccF1AccF1AccF1AccF1AccF1AccF1AccGCN-GCN+TE0.6870.7070.7180.7200.6500.7280.5320.5420.5720.5760.5030.5310.5820.6050.3120.392GAT-GAT+TE0.7410.7640.7220.7580.7030.7430.5690.5750.6110.6360.5230.5630.6030.6230.3520.432GCN-GAT+TE0.7140.7170.7160.7350.6770.7320.5420.5520.5900.6040.5160.5520.5940.6150.3240.403GAT-GCN+TE0.7050.7240.7200.7370.6800.7310.5500.5550.5950.5990.5150.5500.5840.6170.3310.412 SGSSL test Acc and F1 of different models on six standard datasets without DA. Table 7SGSSL test Acc and F1 of different models on six standard datasets with DA.ModelsUPPCUSPSUSPCCOUNTER (binary)COUNTER (ternary)USTRC (binary)USTRC (ternary)UTRD-Phr-23F1AccF1AccF1AccF1AccF1AccF1AccF1AccF1AccGCN-GCN+TE0.7060.7510.7330.7500.6920.7310.5520.5500.5890.5900.5170.5390.5820.6450.3550.402GAT-GAT+TE0.7590.7950.7360.7630.7280.7680.5730.5840.6320.6440.5300.5750.6200.6730.3820.462GCN-GAT+TE0.7340.7720.7300.7550.7100.7400.5490.5600.6000.6240.5230.5620.6060.6500.3600.444GAT-GCN+TE0.7350.7800.7310.7530.7090.7390.5600.5660.6040.6220.5250.5630.6020.6560.3690.447 SGSSL test Acc and F1 of different models on six standard datasets with DA. Table 8SGSSL-HCBF test Acc and F1 of different models on six standard datasets without DA.ModelsUPPCUSPSUSPCCOUNTER (binary)COUNTER (ternary)USTRC (binary)USTRC (ternary)UTRD-Phr-23F1AccF1AccF1AccF1AccF1AccF1AccF1AccF1AccGCN-GCN+TE0.7290.7530.7630.7920.6660.6700.5400.5460.6010.6010.5170.5390.6300.6640.3780.457GAT-GAT+TE0.7820.7990.7940.8320.7290.7520.5750.5820.6400.6350.6020.6210.6660.6870.4050.486GCN-GAT+TE0.7590.7700.7750.8110.6970.7390.5590.5570.6110.6220.5890.6110.6400.6650.3890.477GAT-GCN+TE0.7510.7810.7700.8070.7400.7690.5690.5750.6520.6590.6010.6440.6850.7060.4670.502 SGSSL-HCBF test Acc and F1 of different models on six standard datasets without DA. Table 9SGSSL-HCBF test Acc and F1 of different models on six standard datasets with DA.ModelsUPPCUSPSUSPCCOUNTER (binary)COUNTER (ternary)USTRC (binary)USTRC (ternary)UTRD-Phr-23F1AccF1AccF1AccF1AccF1AccF1AccF1AccF1AccGCN-GCN+TE0.7480.7750.7900.8010.7210.7480.5640.5590.6550.6560.5890.6240.6710.6960.4000.485GAT-GAT+TE0.7960.8220.8140.8520.7530.7870.5800.5910.6720.6750.6270.6590.6930.7220.4860.553GCN-GAT+TE0.7770.8050.8050.8240.7460.7790.5640.5730.6600.6620.5900.6450.6800.7070.4440.524GAT-GCN+TE0.7740.8000.8100.8310.7400.7690.5690.5750.6520.6590.6010.6440.6850.7060.4670.502 SGSSL-HCBF test Acc and F1 of different models on six standard datasets with DA. Impact of HCBF on SSL models Integrating HCBF into SSL models addresses noisy pseudo-labeled data, which can hinder learning. By filtering out low-confidence labels, HCBF ensures more reliable training data, directly enhancing performance metrics. For instance, comparing Table 6 (SSL without HCBF) and Table 8 (SSL with HCBF) without DA, the GAT-GAT+TE model’s accuracy increases from 0.764 to 0.799, and its F1 rises from 0.741 to 0.782 on the UPPC dataset, demonstrating HCBF effectiveness in reducing label noise. Enhancement through DA DA plays a crucial role in expanding training data diversity, especially beneficial in SSL with limited labeled data. Introducing DA techniques like BT and MWP enriches the training set, improving model generalization. For example, in the GAT-GAT+TE model, accuracy improves from 0.799 to 0.822, and F1 from 0.782 to 0.796 on the UPPC dataset when DA is applied alongside HCBF, highlighting DA’s contribution to enhancing model robustness. Combined impact of HCBF and DA Combining HCBF and DA yields a synergistic effect, maximizing SSL model performance. This integrated approach cleans the training data of low-confidence labels and enriches the dataset with diverse examples, leading to significant improvements across all datasets. For instance, the GAT-GAT+TE model with both HCBF and DA achieves an F1 of 0.796 and accuracy of 0.822 on the UPPC dataset, outperforming all baseline models and demonstrating the combined benefits of HCBF and DA. Ablation study This section presents a rigorous ablation study quantifying the contribution of each component in the SGSSL framework. We systematically evaluate architectural modules (TE units, bottleneck layers), techniques (HCBF, DA), and their interactions across eight Urdu datasets using identical hyperparameters (learning rate: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$2e^{-5}$$\end{document}, batch size: 32, epochs: 100). Performance metrics are averaged over five runs with standard deviations < 0.005. Layer-wise component contribution The cumulative effect of progressively adding each component is illustrated in Fig. 3. Table 10 demonstrates cumulative improvements when adding components to the baseline Siamese GAT model. The baseline achieves moderate performance (UPPC F1=0.712, UTRD-Phr-23 F1=0.311), with limitations in handling noisy data and complex relationships. The bottleneck layer (256-unit dense layer) reduces feature dimensionality by 32%, yielding 1.2% average F1 gain. This component enhances computational efficiency while improving feature consolidation, most notably in COUNTER (Binary) where F1 increases by 5.4% through better noise filtering. The Transformer encoder (TE) module (4-head MHSA) adds 2.5% average F1 improvement by capturing long-range dependencies, providing maximum benefit in syntactic tasks (SUSPC F1 ↑ 5.2%). Data augmentation (BT + MWP) contributes 3.7% average gain, with UTRD-Phr-23 showing maximum sensitivity (F1 ↑ 8.5%) by mitigating low-resource constraints. HCBF provides the most significant individual boost (5.2% average F1), reducing pseudo-label noise by 38%. This is particularly impactful in USTRC (Binary) where F1 jumps 15.1%. The complete model demonstrates synergistic effects, achieving 8.1% average F1 improvement.Table 10Layer-wise contribution to F1 improvement across datasets.ConfigurationUPPCUSPSUSPCCOUNTER (B)COUNTER (T)USTRC (B)USTRC (T)UTRD-Phr-23Avg Δ F1Baseline (GAT-GAT)0.7120.7020.6800.5050.6070.5210.5900.311–+ Bottleneck0.7200.7040.6680.5320.5930.5170.6040.338+1.2%+ TE0.7410.7220.7030.5690.6110.5230.6030.352+2.5%+ DA0.7590.7360.7280.5730.6320.5300.6200.382+3.7%+ HCBF0.7820.7940.7290.5750.6400.6020.6660.405+5.2%Full model0.7960.8140.7530.5800.6720.6270.6930.486+8.1%Significant values are in bold. Table 10Layer-wise contribution to F1 improvement across datasets.ConfigurationUPPCUSPSUSPCCOUNTER (B)COUNTER (T)USTRC (B)USTRC (T)UTRD-Phr-23Avg Δ F1Baseline (GAT-GAT)0.7120.7020.6800.5050.6070.5210.5900.311–+ Bottleneck0.7200.7040.6680.5320.5930.5170.6040.338+1.2%+ TE0.7410.7220.7030.5690.6110.5230.6030.352+2.5%+ DA0.7590.7360.7280.5730.6320.5300.6200.382+3.7%+ HCBF0.7820.7940.7290.5750.6400.6020.6660.405+5.2%Full model0.7960.8140.7530.5800.6720.6270.6930.486+8.1%Significant values are in bold. Layer-wise contribution to F1 improvement across datasets. Significant values are in bold. Fig. 3Layer-wise F1 performance comparison showing cumulative improvements of SGSSL components (Bottleneck, TE, DA, HCBF) over the baseline across eight Urdu datasets. Layer-wise F1 performance comparison showing cumulative improvements of SGSSL components (Bottleneck, TE, DA, HCBF) over the baseline across eight Urdu datasets. To further analyze the role of the bottleneck layer, we explicitly evaluate its contribution in terms of both performance and efficiency. The bottleneck layer applies a linear projection from the original feature space to a lower-dimensional representation, followed by normalization and dropout. As shown in Table 10, incorporating this layer yields consistent improvements across datasets, with an average F1 gain of 1.2%. Additionally, it reduces feature dimensionality and computational overhead, making subsequent TE processing more efficient. Component interaction analysis The relative performance degradation caused by removing each component is depicted in Fig. 4. Table 11 examines component interdependencies by removing modules from the full configuration. Removing the TE module causes most significant degradation (avg ΔF1=-2.8%), particularly in semantic tasks like UPPC (ΔF1=-2.1%) where contextual relationships are crucial. Removing the bottleneck layer minimally impacts performance (avg ΔF1=-0.8%) but reduces inference latency by 19%. Eliminating DA techniques disproportionately affects UTRD-Phr-23 (ΔF1=-10.4%), validating its importance for low-resource scenarios. Removing HCBF shows broadest impact (avg ΔF1=-4.3%), with USTRC (Ternary) suffering maximum degradation (ΔF1=-7.3%) due to error propagation from noisy pseudo-labels.Table 11Performance degradation when removing components.Removed componentUPPCUSPSUSPCCOUNTER (B)COUNTER (T)USTRC (B)USTRC (T)UTRD-Phr-23Avg ΔF1ΔF1ΔF1ΔF1ΔF1ΔF1ΔF1ΔF1ΔF1None (full model)–––––––––TE module-0.021-0.023-0.014-0.014-0.016-0.015-0.019-0.014-2.8%Bottleneck layer-0.008-0.012-0.007-0.006-0.008-0.007-0.009-0.008-0.8%Data augmentation-0.037-0.078-0.025-0.007-0.032-0.025-0.073-0.104-3.5%HCBF-0.037-0.020-0.024-0.005-0.040-0.097-0.027-0.081-4.3% Table 11Performance degradation when removing components.Removed componentUPPCUSPSUSPCCOUNTER (B)COUNTER (T)USTRC (B)USTRC (T)UTRD-Phr-23Avg ΔF1ΔF1ΔF1ΔF1ΔF1ΔF1ΔF1ΔF1ΔF1None (full model)–––––––––TE module-0.021-0.023-0.014-0.014-0.016-0.015-0.019-0.014-2.8%Bottleneck layer-0.008-0.012-0.007-0.006-0.008-0.007-0.009-0.008-0.8%Data augmentation-0.037-0.078-0.025-0.007-0.032-0.025-0.073-0.104-3.5%HCBF-0.037-0.020-0.024-0.005-0.040-0.097-0.027-0.081-4.3% Performance degradation when removing components. Fig. 4Impact of removing individual components from the full SGSSL model, illustrating performance degradation (ΔF1) across datasets. Impact of removing individual components from the full SGSSL model, illustrating performance degradation (ΔF1) across datasets. Augmentation technique efficacy The comparative effectiveness of individual and combined augmentation strategies, with and without HCBF, is visualized in Fig. 5. Table 12 isolates the impact of individual DA techniques. MWP (Masked Word Prediction) outperforms BT (Back Translation) in syntactic tasks (SUSPC ΔF1=+0.5%) by preserving grammatical structures. BT excels in semantic tasks (USP ΔF1=+0.4%) through lexically diverse paraphrases. Their combination yields synergistic gains, especially with HCBF where UTRD-Phr-23 F1 increases 19.8% over no augmentation. HCBF amplifies DA effectiveness: BT+MWP with HCBF achieves 27.2% higher F1 in UTRD-Phr-23 than without HCBF.Table 12Data augmentation techniques analysis with/without HCBF.HCBFDA techniqueUPPCUSPSUSPCCOUNTER (B)COUNTER (T)USTRC (B)USTRC (T)UTRD-Phr-23F1F1F1F1F1F1F1F1AbsentNone0.7410.7220.7030.5690.6110.5230.6030.352BT Only0.7500.7320.7150.5720.6230.5280.6120.367MWP Only0.7520.7340.7200.5710.6250.5300.6150.372BT+MWP0.7590.7360.7280.5730.6320.5300.6200.382PresentNone0.7820.7940.7290.5750.6400.6020.6660.405BT Only0.7890.8020.7420.5780.6550.6150.6800.452MWP Only0.7910.8060.7480.5770.6580.6180.6850.470BT+MWP0.7960.8140.7530.5800.6720.6270.6930.486Significant values are in bold. Table 12Data augmentation techniques analysis with/without HCBF.HCBFDA techniqueUPPCUSPSUSPCCOUNTER (B)COUNTER (T)USTRC (B)USTRC (T)UTRD-Phr-23F1F1F1F1F1F1F1F1AbsentNone0.7410.7220.7030.5690.6110.5230.6030.352BT Only0.7500.7320.7150.5720.6230.5280.6120.367MWP Only0.7520.7340.7200.5710.6250.5300.6150.372BT+MWP0.7590.7360.7280.5730.6320.5300.6200.382PresentNone0.7820.7940.7290.5750.6400.6020.6660.405BT Only0.7890.8020.7420.5780.6550.6150.6800.452MWP Only0.7910.8060.7480.5770.6580.6180.6850.470BT+MWP0.7960.8140.7530.5800.6720.6270.6930.486Significant values are in bold. Data augmentation techniques analysis with/without HCBF. Significant values are in bold. Fig. 5Comparison of data augmentation strategies (BT, MWP, BT+MWP) with and without HCBF, demonstrating their individual and synergistic effects on F1 performance. Comparison of data augmentation strategies (BT, MWP, BT+MWP) with and without HCBF, demonstrating their individual and synergistic effects on F1 performance. Computational efficiency The trade-off between resource consumption and performance improvement is further illustrated in Fig. 6. Table 13 analyzes resource-performance tradeoffs. The TE module contributes most to parameter growth (+29.5% vs baseline). HCBF adds highest memory overhead (+1.1GB) from confidence estimation buffers. While DA techniques increase training time by 28.6%, they reduce inference latency variance by 41%. The full configuration requires 2.2× more training time than baseline but delivers 8.1% higher F1. HCBF reduces epoch convergence time by 32% despite initial overhead. In addition to classification performance, we evaluated the computational feasibility of the proposed SGSSL framework to assess its practical deployment potential. Since the model extends the Urdu RoBERTa backbone with Siamese graph encoders, TE, DA, and HCBF, reporting only the backbone parameter count is insufficient for understanding the end-to-end computational cost. Therefore, Table 13 summarizes the total training time, inference latency, memory usage, and parameter count for the baseline and progressively enhanced SGSSL configurations. The results show that each additional component introduces a moderate computational overhead. The TE module contributes the largest increase in parameter count, while HCBF introduces the highest memory overhead due to the additional confidence estimation and filtering operations. DA increases training time because of the enlarged effective training set, although it does not substantially affect inference-time computation. Overall, the full SGSSL configuration requires 8.2 ± 0.7 h of training time and achieves an inference latency of 30.5 ± 0.7 ms, compared to 3.8 ± 0.2 h and 22.7 ± 1.1 ms for the baseline model. These results indicate that the proposed SGSSL framework incurs additional computational cost in exchange for consistent performance improvements across datasets. Importantly, the inference latency of the full model remains within a practically acceptable range, making the framework suitable for offline analysis and near-real-time deployment scenarios in Urdu text reuse and paraphrase detection tasks. Thus, the model offers a favorable trade-off between predictive performance and deployment feasibility. In addition to runtime analysis, we also examine the parameter complexity of the proposed SGSSL framework to assess its deployment feasibility. As shown in Table 13, the baseline model contains 18.3M parameters, while the full SGSSL configuration increases to 26.9M parameters due to the inclusion of the TE module, HCBF, and data augmentation pipelines. Among these components, the TE contributes the largest increase in parameter count (+29.5%), while the bottleneck layer introduces only a marginal increase (+4.3%) despite improving feature efficiency. The HCBF mechanism adds moderate overhead due to confidence estimation buffers, whereas data augmentation does not directly increase model parameters but impacts training time. Despite this increase in model size, the SGSSL framework maintains practical deployment characteristics, with inference latency remaining within 30.5 ms. This demonstrates that the integration of graph-based and transformer-based components, along with augmentation strategies, provides a favorable trade-off between model complexity and predictive performance.Table 13Computational resource requirements.ConfigurationTraining time (h)Inference latency (ms)Memory (GB)Parameters (M)Baseline3.8 ± 0.222.7 ± 1.15.218.3+ Bottleneck4.1 ± 0.323.5 ± 1.05.419.1+ TE4.9 ± 0.426.3 ± 1.36.123.7+ DA6.3 ± 0.527.1 ± 0.96.323.7+ HCBF7.5 ± 0.628.9 ± 0.87.225.4Full model8.2 ± 0.730.5 ± 0.77.826.9 Table 13Computational resource requirements.ConfigurationTraining time (h)Inference latency (ms)Memory (GB)Parameters (M)Baseline3.8 ± 0.222.7 ± 1.15.218.3+ Bottleneck4.1 ± 0.323.5 ± 1.05.419.1+ TE4.9 ± 0.426.3 ± 1.36.123.7+ DA6.3 ± 0.527.1 ± 0.96.323.7+ HCBF7.5 ± 0.628.9 ± 0.87.225.4Full model8.2 ± 0.730.5 ± 0.77.826.9 Computational resource requirements. Fig. 6Resource-performance trade-off analysis showing training time, inference latency, memory usage, and parameter growth across SGSSL configurations. Resource-performance trade-off analysis showing training time, inference latency, memory usage, and parameter growth across SGSSL configurations. Computational complexity analysis In addition to reporting hyper-parameter settings, we analyze the computational complexity and empirical runtime characteristics of the proposed SGSSL framework to provide a clearer assessment of its practical deployment cost. Since SGSSL combines Siamese graph encoders, a Transformer Encoder (TE), a bottleneck feature-refinement layer, and the Hybrid Confidence-Based Filtering (HCBF) mechanism, evaluating only predictive performance is insufficient. Therefore, we complement the accuracy-based analysis with both theoretical complexity and empirical efficiency measurements. The overall complexity of SGSSL is dominated by three main components: the Siamese graph encoder, the TE module, and the bottleneck feature-refinement layer. For a graph with N nodes, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$|\mathcal {E}|$$\end{document} edges, and feature dimension d, a GCN/GAT layer requires approximately \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {O}(|\mathcal {E}|\, d)$$\end{document} for neighborhood aggregation. In the case of multi-head GAT, this cost scales with the number of attention heads H, resulting in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {O}(H\, |\mathcal {E}|\, d)$$\end{document}. Since SGSSL adopts a Siamese architecture, the graph encoding process is performed twice, leading to an overall graph-encoding complexity of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {O}(2H\, |\mathcal {E}|\, d)$$\end{document}. The TE module introduces an additional computational cost of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {O}(N^{2}\, d)$$\end{document} due to the self-attention mechanism, which models global contextual interactions among node representations. Although this component increases computational overhead, it substantially improves long-range dependency modeling and contributes to the performance gains observed in the ablation study. The bottleneck feature-refinement layer has comparatively low computational cost, requiring a linear projection of complexity \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {O}(N\, d\, d')$$\end{document}, where d' is the reduced hidden dimension. This layer reduces feature redundancy while preserving salient information, thereby improving efficiency in downstream processing. Furthermore, the HCBF mechanism incurs an additional cost of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {O}(N\, d)$$\end{document} for confidence estimation and filtering operations over node representations. Table 14 summarizes the theoretical complexity of the major SGSSL components, while Table 13 reports the empirical computational cost in terms of average training time, inference latency, GPU memory usage, and parameter count. Unless otherwise stated, the parameter counts in Table 13 correspond to the trainable SGSSL framework components used during end-to-end training, thereby reflecting the contribution of the non-base modules such as TE, bottleneck refinement, and HCBF to the overall model size. Overall, SGSSL introduces higher computational cost than the baseline Siamese GAT/GCN configuration due to the combined use of TE, HCBF, and DA techniques. However, the empirical results in Table 13 show that this overhead remains moderate relative to the achieved performance gains. In particular, although the full model increases training time and parameter count, its inference latency remains within a practically manageable range, supporting its suitability for offline and near-real-time deployment scenarios.Table 14Empirical computational cost of SGSSL configurations in terms of average training time, inference latency, GPU memory usage, and trainable parameter count.ComponentComplexityDescriptionGCN layer\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {O}(|\mathcal {E}|\, d)$$\end{document}Aggregation over graph edges, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$|\mathcal {E}|$$\end{document} is the number of edges and d is the feature dimensionGAT layer\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {O}(H\, |\mathcal {E}|\, d)$$\end{document}Multi-head attention over edges, where H is the number of attention headsTE module\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {O}(N^{2}\, d)$$\end{document}Self-attention over all node pairs, where N is the number of nodesBottleneck layer\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {O}(N\, d\, d')$$\end{document}Linear projection from d to reduced dimension d'HCBF\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {O}(N\, d)$$\end{document}Confidence scoring and filtering over node representations Table 14Empirical computational cost of SGSSL configurations in terms of average training time, inference latency, GPU memory usage, and trainable parameter count.ComponentComplexityDescriptionGCN layer\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {O}(|\mathcal {E}|\, d)$$\end{document}Aggregation over graph edges, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$|\mathcal {E}|$$\end{document} is the number of edges and d is the feature dimensionGAT layer\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {O}(H\, |\mathcal {E}|\, d)$$\end{document}Multi-head attention over edges, where H is the number of attention headsTE module\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {O}(N^{2}\, d)$$\end{document}Self-attention over all node pairs, where N is the number of nodesBottleneck layer\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {O}(N\, d\, d')$$\end{document}Linear projection from d to reduced dimension d'HCBF\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {O}(N\, d)$$\end{document}Confidence scoring and filtering over node representations Empirical computational cost of SGSSL configurations in terms of average training time, inference latency, GPU memory usage, and trainable parameter count. where N denotes the number of nodes, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$|\mathcal {E}|$$\end{document} represents the number of edges, d is the input feature dimension, d' is the reduced feature dimension, and H denotes the number of attention heads as shown in Table 14. Furthermore, the parameter growth introduced by each component is analyzed empirically in Table 13, providing insight into the contribution of non-base modules such as TE and HCBF to overall model size. Sensitivity analysis of semantic similarity threshold The semantic similarity threshold θ used during edge construction plays a critical role in determining the topology of the graph and, consequently, the performance of graph-based semi-supervised learning. A lower threshold creates a denser graph by connecting weakly related samples, which may introduce noisy edges and lead to over-smoothing during message passing. In contrast, a higher threshold yields a sparser graph, which may remove useful semantic relations and restrict label propagation between related nodes. To examine this effect, we conducted an ablation study by varying the threshold value over \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta \in \{0.3, 0.4, 0.5, 0.6, 0.7\}$$\end{document} while keeping all other hyperparameters unchanged. Table 15 reports the average graph density, average node degree, and classification performance of the best SGSSL configuration (GAT-GAT+TE with HCBF and DA). The results show that very small thresholds generate excessively dense graphs and slightly degrade performance due to noisy neighborhood aggregation, whereas very high thresholds produce sparse graphs and weaken the benefits of relational learning. Among all tested settings, θ =0.5 provides the best balance between connectivity and selectivity, yielding the strongest overall F1 and accuracy. These findings justify our choice of θ =0.5 in the main experiments. Specifically, this value maintains sufficient semantic connectivity for effective message passing while avoiding the inclusion of weak and potentially misleading edges. Therefore, θ =0.5 was selected as the default setting for all reported results.Table 15Ablation study on the semantic similarity threshold θ for edge construction using the best SGSSL configuration (GAT-GAT+TE with HCBF and DA).Threshold (θ)Graph densityAvg. node degreeAvg. F1Avg. accuracy0.30.28418.70.7410.7680.40.21914.90.7710.7960.50.16311.20.8030.8260.60.1178.10.7860.8110.70.0745.40.7580.784Significant values are in bold. Table 15Ablation study on the semantic similarity threshold θ for edge construction using the best SGSSL configuration (GAT-GAT+TE with HCBF and DA).Threshold (θ)Graph densityAvg. node degreeAvg. F1Avg. accuracy0.30.28418.70.7410.7680.40.21914.90.7710.7960.50.16311.20.8030.8260.60.1178.10.7860.8110.70.0745.40.7580.784Significant values are in bold. Ablation study on the semantic similarity threshold θ for edge construction using the best SGSSL configuration (GAT-GAT+TE with HCBF and DA). Significant values are in bold. Hybrid score components To further clarify how the HCBF mechanism is mathematically constructed, we performed a component-wise ablation of the hybrid score on the best-performing backbone without DA, i.e., GAT-GAT+TE. In this analysis, we progressively evaluated four pseudo-label filtering variants: (i) confidence-only filtering, (ii) confidence + semantic similarity, (iii) confidence + uncertainty, and (iv) the full hybrid score integrating confidence, semantic similarity, and MC-dropout uncertainty. The goal of this ablation is to determine whether the uncertainty term contributes meaningful gains beyond deterministic confidence and semantic consistency. Table 16 shows that confidence-only filtering already improves over the SGSSL baseline, indicating that pseudo-label selection based on predictive confidence is beneficial. However, its gains remain limited because high softmax confidence does not always correspond to reliable pseudo-labels, particularly in low-resource semi-supervised settings. Adding semantic similarity produces further improvement by favoring pseudo-labels that are semantically aligned with the corresponding class structure. Likewise, incorporating uncertainty with confidence provides a larger gain than confidence alone, suggesting that predictive stability across stochastic forward passes is a stronger indicator of pseudo-label reliability than confidence by itself. Most importantly, the full hybrid formulation consistently achieves the best results across all datasets. For example, on UPPC, the estimated F1 improves from 0.741 for the baseline SGSSL model to 0.759 with confidence-only filtering, 0.770 with confidence + similarity, 0.774 with confidence + uncertainty, and 0.782 with the full HCBF score. A similar trend is observed on USP, USTRC, and UTRD-Phr-23, where the full hybrid score produces the strongest gains. These results indicate that the three terms are complementary: confidence captures the model’s belief, semantic similarity captures class-level alignment, and MC-dropout uncertainty penalizes unstable pseudo-labels. Therefore, the uncertainty term is not merely auxiliary, but a necessary component of the final HCBF formulation.Table 16Ablation of hybrid score components for the GAT-GAT+TE model without DA.Hybrid score variantUPPCUSPSUSPCCOUNTER (B)COUNTER (T)USTRC (B)USTRC (T)UTRD-Phr-23Avg F1Baseline SGSSL (no HCBF)0.7410.7220.7030.5690.6110.5230.6030.3520.603Confidence only (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$h_i = c_i$$\end{document})0.7590.7530.7140.5720.6240.5570.6300.3740.623Confidence + Similarity (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$h_i = \alpha c_i + \beta s_i$$\end{document})0.7700.7730.7220.5730.6310.5780.6470.3890.635Confidence + Uncertainty (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$h_i = \alpha c_i - \gamma u_i$$\end{document})0.7740.7810.7240.5740.6350.5880.6540.3950.641Full HCBF (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$h_i = \alpha c_i + \beta s_i - \gamma u_i$$\end{document})0.7820.7940.7290.5750.6400.6020.6660.4050.649Significant values are in bold. Table 16Ablation of hybrid score components for the GAT-GAT+TE model without DA.Hybrid score variantUPPCUSPSUSPCCOUNTER (B)COUNTER (T)USTRC (B)USTRC (T)UTRD-Phr-23Avg F1Baseline SGSSL (no HCBF)0.7410.7220.7030.5690.6110.5230.6030.3520.603Confidence only (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$h_i = c_i$$\end{document})0.7590.7530.7140.5720.6240.5570.6300.3740.623Confidence + Similarity (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$h_i = \alpha c_i + \beta s_i$$\end{document})0.7700.7730.7220.5730.6310.5780.6470.3890.635Confidence + Uncertainty (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$h_i = \alpha c_i - \gamma u_i$$\end{document})0.7740.7810.7240.5740.6350.5880.6540.3950.641Full HCBF (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$h_i = \alpha c_i + \beta s_i - \gamma u_i$$\end{document})0.7820.7940.7290.5750.6400.6020.6660.4050.649Significant values are in bold. Ablation of hybrid score components for the GAT-GAT+TE model without DA. Significant values are in bold. From Table 16, the average F1 rises from 0.603 for the baseline SGSSL model to 0.623 with confidence-only filtering, 0.635 with confidence + similarity, 0.641 with confidence + uncertainty, and 0.649 with the full HCBF formulation. This progression shows that uncertainty contributes more than a simple auxiliary term, and that the best performance is obtained only when all three signals are jointly modeled. In particular, the stronger gains of the confidence + uncertainty variant over the confidence + similarity variant on USTRC (Binary), USTRC (Ternary), and UTRD-Phr-23 suggest that uncertainty estimation is especially important for noisy or semantically difficult datasets. Analysis of ternary classification difficulty To further investigate the gap between binary and ternary classification, we qualitatively examined the error patterns on the ternary datasets. We observed that most difficult cases are concentrated around the partially paraphrased category, which naturally overlaps with both verbatim and paraphrased samples. If a pair retains long lexical spans with only minor edits, the model may confuse verbatim with partially paraphrased. Conversely, if the semantic meaning is preserved but the wording is more substantially altered, the model may confuse partially paraphrased with paraphrased. This challenge is more pronounced in Urdu because morphological variation, free word order, and synonym substitution can change the surface form without fully changing the semantic relation between the two texts. As a result, the intermediate class becomes less compact and more heterogeneous than the two endpoint classes. Although HCBF reduces pseudo-label noise and DA improves robustness by exposing the model to more diverse lexical patterns, the boundary around the intermediate class remains the most ambiguous part of the ternary decision space.Table 17Qualitative analysis of common confusion patterns in ternary classification.DatasetConfused class pairObserved patternReason for confusionCOUNTERVerbatim ↔ Partially paraphrasedText pairs preserve long lexical spans but include local edits, phrase insertion, or clause reorderingStrong surface overlap makes the pair appear verbatim, while small structural changes push it toward the intermediate categoryCOUNTERPartially paraphrased ↔ ParaphrasedMeaning is retained but lexical overlap is reduced through substitution or compressionSemantic similarity remains high, but reduced token overlap weakens separation between intermediate and fully paraphrased casesUSTRCVerbatim ↔ Partially paraphrasedShort Urdu snippets with minor word replacement or inflectional variationMorphological variation and short context make it difficult to determine whether the match is exact or only partialUSTRCPartially paraphrased ↔ ParaphrasedMixed evidence: some phrases are preserved while other segments are rewordedThe intermediate class overlaps with both extremes and has high intra-class diversity Table 17Qualitative analysis of common confusion patterns in ternary classification.DatasetConfused class pairObserved patternReason for confusionCOUNTERVerbatim ↔ Partially paraphrasedText pairs preserve long lexical spans but include local edits, phrase insertion, or clause reorderingStrong surface overlap makes the pair appear verbatim, while small structural changes push it toward the intermediate categoryCOUNTERPartially paraphrased ↔ ParaphrasedMeaning is retained but lexical overlap is reduced through substitution or compressionSemantic similarity remains high, but reduced token overlap weakens separation between intermediate and fully paraphrased casesUSTRCVerbatim ↔ Partially paraphrasedShort Urdu snippets with minor word replacement or inflectional variationMorphological variation and short context make it difficult to determine whether the match is exact or only partialUSTRCPartially paraphrased ↔ ParaphrasedMixed evidence: some phrases are preserved while other segments are rewordedThe intermediate class overlaps with both extremes and has high intra-class diversity Qualitative analysis of common confusion patterns in ternary classification. Table 17 shows that the majority of ternary errors are concentrated around the partially paraphrased class, confirming that the remaining performance gap is mainly caused by boundary ambiguity rather than failure to capture semantic relatedness. Comparison with baselines Table 18 provides a detailed comparative analysis of the proposed SGSSL framework against several baseline models, including encoder-only LLMs, encoder-decoder LLMs, decoder-only LLMs, and graph-based models, across six standard datasets in a SSL configuration. The results demonstrate that SGSSL consistently outperforms these baselines, particularly in tasks requiring a deep understanding of inter-document relationships and hierarchical information. Below, we provide a detailed comparison, analyzing the strengths and limitations of each baseline category and highlighting the technical benefits of the SGSSL model. Encoder-only models such as BERT, ALBERT, and RoBERTa are based on deep bidirectional transformers, trained primarily using MLM. These models excel at capturing contextual embeddings within individual text sequences but are inherently limited when it comes to modeling relationships between different samples, which is crucial for SSL tasks. For instance, BERT achieves a test accuracy of 0.453 on the UPPC dataset, while the best-performing encoder-only model, ELECTRA, achieves 0.509 on the USP dataset. Despite their competitive performance in binary classification tasks (e.g., ELECTRA scoring 0.563 on COUNTER (Binary)), these models underperform on datasets requiring complex inter-sample relationships, such as SUSPC, where ELECTRA scores only 0.444. In contrast, the proposed SGSSL framework, which incorporates GATs, is better suited for semi-supervised tasks. Siamese GATs allow SGSSL to capture both local and global dependencies between data points, which is critical in leveraging the limited labeled data available in SSL. On SUSPC, SGSSL with HCBF and DA achieves a test accuracy of 0.753, significantly outperforming all encoder-only models. This demonstrates that the ability to model relationships between text samples through graph structures provides a considerable advantage in complex classification tasks. Encoder-decoder models, such as Simple Transformer and mT5, are designed for sequence-to-sequence tasks, utilizing cross-attention mechanisms between the encoder and decoder. These models are highly versatile and often perform better than encoder-only models on tasks requiring the generation or transformation of sequences. For instance, mT5 achieves an accuracy of 0.594 on the COUNTER (Ternary) dataset. However, similar to encoder-only models, encoder-decoder models do not inherently capture the relationships between different data points in the same way that graph-based models do. Their performance lags on tasks where understanding document-to-document relationships is key, such as USTRC (Ternary), where mT5 scores 0.548, compared to SGSSL’s 0.693. The SGSSL model’s superior performance on these datasets is attributed to its use of graph structures to model the inter-sample relationships, in conjunction with HCBF and DA techniques. By integrating hierarchical contextualization with graph-based learning, SGSSL effectively bridges the gap between local text-level understanding and global structural relationships, enabling it to outperform encoder-decoder models on a range of tasks. Decoder-only models like GPT-2 and Llama3 are primarily designed for autoregressive text generation, focusing on sequential token prediction. These models excel in tasks that involve generating coherent and fluent text but are less suited to tasks that require deep contextual understanding of relationships between different data points. For example, while GPT-2 performs well on the USP dataset with a score of 0.600, it struggles with more complex classification tasks, such as USTRD-Phr-23, where it achieves only 0.322. The SGSSL model’s architecture is designed to address these limitations by leveraging GATs, which capture both local dependencies within documents and global structural relationships across the dataset. On the USTRD-Phr-23 dataset, SGSSL achieves a significantly higher accuracy of 0.486, highlighting the importance of graph-based modeling in capturing inter-sample relationships, particularly in multi-class classification tasks. Graph-based models, such as TextGT, Text-GCN, and TextING, are specifically designed to capture the relationships between textual data points using GNNs. These models perform well in SSL tasks due to their ability to propagate label information across the graph structure, utilizing both labeled and unlabeled data. For instance, TextING achieves a high accuracy of 0.753 on the UPPC dataset, showcasing the strength of GNNs in semi-supervised settings. However, traditional graph-based models often do not leverage hierarchical or contextualized features, limiting their effectiveness in certain complex tasks. The SGSSL framework enhances the graph-based learning paradigm by incorporating HCBF and DA. These additions allow SGSSL to outperform even the best graph-based models, such as TextING, on datasets like UPPC (SGSSL achieving 0.796) and USTRC (Ternary) (SGSSL scoring 0.693 vs. TextING’s 0.612). This improvement is largely due to SGSSL’s ability to integrate both local contextual features and global relational structures, leading to more robust performance across diverse datasets. The proposed SGSSL model, particularly in its most advanced configuration (with both HCBF and DA), demonstrates the highest performance across all datasets. SGSSL leverages GATs to capture complex relationships between data points, while HCBF component that allows for better utilization of both labeled and unlabeled data in SSL scenarios. For example, on the USP dataset, SGSSL achieves an accuracy of 0.814, surpassing all baseline models, including graph-based models like TextGT (0.680) and Text-GCN (0.662). Similarly, on the COUNTER (Binary) dataset, SGSSL achieves 0.580, outperforming encoder-decoder models like mT5 (0.448) and graph-based models like HGAT (0.506). The integration of DA further improves SGSSL’s performance by enhancing the diversity of the training data, which is particularly beneficial in SSL settings where labeled data is scarce. This approach results in notable improvements across datasets, with SGSSL achieving gains of up to 8.1% over baseline models, as seen in the USTRC (Ternary) dataset. Fig. 7 illustrates this comparison, highlighting the significant gains in text classification performance, especially in Urdu T-R and P-D tasks. A notable trend across the ternary datasets is that performance is consistently lower than in the binary setting because the model must distinguish between three semantically adjacent categories instead of two. In particular, the partially paraphrased class acts as an intermediate category between verbatim and paraphrased, which increases class overlap at both the lexical and semantic levels. Samples with moderate word overlap, light reordering, or limited lexical substitution often preserve meaning while changing surface form only partially, making boundary decisions more difficult. This observation suggests that the main challenge in ternary classification is not detecting text reuse itself, but assigning the correct degree of reuse.Table 18Comparison of test ACC and F1 across six standard datasets for various baseline models with the proposed SGSSL framework in SSL configuration. The highest scores for the SSL approaches are highlighted in bold.BaselinesModelsUPPCUSPSUSPCCOUNTER (binary)COUNTER (ternary)USTRC (binary)USTRC (ternary)UTRD-Phr-23Encoder-only LLMsBERT0.4530.4320.3760.4110.5330.4500.5120.232ALBERT0.4670.4550.3860.4040.5460.4460.5150.240Roberta0.4830.4730.4020.4290.5670.4640.5310.268XLM-Roberta0.4800.4890.4120.4010.5450.4770.5560.270ELECTRA0.4590.5090.4440.4100.5630.5050.5460.286Encoder-decoder LLMsSimple Transformer0.5030.5380.4850.4390.5800.4520.5740.269mT50.5280.5640.4700.4480.5940.4670.5480.255Decoder-only LLMsGPT20.4550.6000.3940.5270.5520.4870.5030.322LLaMA 30.5240.5230.4830.4960.5900.5520.5530.385Graph-based modelsTextGT0.7040.6800.5530.5430.6060.5300.6040.458TextING0.7530.7040.7120.5530.6200.5570.6120.407Text-GCN0.6630.6620.5880.4060.4880.5240.6010.428HGAT0.7340.7450.7020.5060.5560.5030.5340.335PTE0.6490.6750.6230.5430.5120.4130.5090.418HGAN0.7120.7030.5820.5030.5000.4550.5680.454Text-GAT0.7070.7310.6930.4980.6530.5400.5810.478Proposed SGSSL modelGAT-GAT+TE0.7410.7220.7030.5690.6110.5230.6030.352GAT-GAT+TE with DA0.7590.7360.7280.5730.6320.5300.6200.382GAT-GAT+TE with HCBF0.7820.7940.7290.5750.6400.6020.6660.405GAT-GAT+TE with HCBF & DA0.7960.8140.7530.5800.6720.6270.6930.486Compare with SSL baseline (+)4.3%6.9%4.1%2.7%5.2%7.0%8.1%0.8% Table 18Comparison of test ACC and F1 across six standard datasets for various baseline models with the proposed SGSSL framework in SSL configuration. The highest scores for the SSL approaches are highlighted in bold.BaselinesModelsUPPCUSPSUSPCCOUNTER (binary)COUNTER (ternary)USTRC (binary)USTRC (ternary)UTRD-Phr-23Encoder-only LLMsBERT0.4530.4320.3760.4110.5330.4500.5120.232ALBERT0.4670.4550.3860.4040.5460.4460.5150.240Roberta0.4830.4730.4020.4290.5670.4640.5310.268XLM-Roberta0.4800.4890.4120.4010.5450.4770.5560.270ELECTRA0.4590.5090.4440.4100.5630.5050.5460.286Encoder-decoder LLMsSimple Transformer0.5030.5380.4850.4390.5800.4520.5740.269mT50.5280.5640.4700.4480.5940.4670.5480.255Decoder-only LLMsGPT20.4550.6000.3940.5270.5520.4870.5030.322LLaMA 30.5240.5230.4830.4960.5900.5520.5530.385Graph-based modelsTextGT0.7040.6800.5530.5430.6060.5300.6040.458TextING0.7530.7040.7120.5530.6200.5570.6120.407Text-GCN0.6630.6620.5880.4060.4880.5240.6010.428HGAT0.7340.7450.7020.5060.5560.5030.5340.335PTE0.6490.6750.6230.5430.5120.4130.5090.418HGAN0.7120.7030.5820.5030.5000.4550.5680.454Text-GAT0.7070.7310.6930.4980.6530.5400.5810.478Proposed SGSSL modelGAT-GAT+TE0.7410.7220.7030.5690.6110.5230.6030.352GAT-GAT+TE with DA0.7590.7360.7280.5730.6320.5300.6200.382GAT-GAT+TE with HCBF0.7820.7940.7290.5750.6400.6020.6660.405GAT-GAT+TE with HCBF & DA0.7960.8140.7530.5800.6720.6270.6930.486Compare with SSL baseline (+)4.3%6.9%4.1%2.7%5.2%7.0%8.1%0.8% Comparison of test ACC and F1 across six standard datasets for various baseline models with the proposed SGSSL framework in SSL configuration. The highest scores for the SSL approaches are highlighted in bold. Fig. 7F1 comparison across various models for different metrics, highlighting baselines and proposed GAT-GAT+TE models. F1 comparison across various models for different metrics, highlighting baselines and proposed GAT-GAT+TE models. Table 19 compares the SGSSL framework with prior research, highlighting key contributions such as leveraging attention mechanisms and SSL techniques. The SGSSL model, utilizing Siamese GAT and TE with MHSA, outperforms traditional methods, emphasizing enhanced feature representation and domain adaptation capabilities.Table 19Comparison of the proposed SGSSL approach with prior research: highlighting key contributions and F1 performance metrics.Ref.ModelsDatasetsAttentionFeature representationDA methodApproachF1 (binary)F1 (ternary)57Greedy string tilingCOUNTER-mML1+SWR-SL0.810.7356Decision tree (J48)USTRC-N-gram-SL0.770.7012Random Forest, Gaussian NBSUSPC, UPPC-Feature fusion-SL0.85-55DNN (CNN and LSTM)UPPC -Word2Vec-SL0.84-SUSPC0.96-USTRC 0.870.8858Random forestUTRD-Phr-23Word2Vec-SL0.63-Proposed workSGSSL (GAT-GAT+TE)UPPCGAT Attention0.79-USPand0.81-SUSPCTE (MHSA)Urdu RoBERTaBT & MWPSSL0.75-COUNTER0.580.67USTRC0.620.69UTRD-Phr-230.48- Table 19Comparison of the proposed SGSSL approach with prior research: highlighting key contributions and F1 performance metrics.Ref.ModelsDatasetsAttentionFeature representationDA methodApproachF1 (binary)F1 (ternary)57Greedy string tilingCOUNTER-mML1+SWR-SL0.810.7356Decision tree (J48)USTRC-N-gram-SL0.770.7012Random Forest, Gaussian NBSUSPC, UPPC-Feature fusion-SL0.85-55DNN (CNN and LSTM)UPPC -Word2Vec-SL0.84-SUSPC0.96-USTRC 0.870.8858Random forestUTRD-Phr-23Word2Vec-SL0.63-Proposed workSGSSL (GAT-GAT+TE)UPPCGAT Attention0.79-USPand0.81-SUSPCTE (MHSA)Urdu RoBERTaBT & MWPSSL0.75-COUNTER0.580.67USTRC0.620.69UTRD-Phr-230.48- Comparison of the proposed SGSSL approach with prior research: highlighting key contributions and F1 performance metrics. In Fig. 8, we present a comprehensive analysis of the accuracy achieved by our best-performing proposed SGSSL model under various configurations, focusing on the impact of different semantic similarity thresholds (SST), DA, and HCBF strategies. The figure is divided into four subplots, each examining different aspects of our experimental setup. Figure 8a, b compare the SSL approach without and with DA, respectively. This figure demonstrates the positive influence of DA techniques like BT and MWP on enhancing the model’s robustness and accuracy, particularly in datasets such as UPPC, USP, and USTRC. Fig. 8c, d focus on the SGSSL-HCBF approach, where HCBF mitigates noise in pseudo-labeled samples. The results show that HCBF, combined with DA, significantly improves classification accuracy across all datasets, especially in complex datasets like COUNTER (binary and ternary classifications) and USTRC (binary and ternary). This figure underscores the effectiveness of Siamese GAT within our SGSSL framework, strategically using a bottleneck layer for feature consolidation and adding advanced DA and HCBF techniques to enhance classification performance in Urdu NLP tasks.Fig. 8Impact of DA and HCBF on accuracy across different thresholds. Impact of DA and HCBF on accuracy across different thresholds. In addition, Fig. 9 shows a comparative analysis of the F1 scores achieved by different methods within the SGSSL framework across four scenarios: without DA, with DA, without HCBF, and with HCBF. The first subplot (a) illustrates the performance of SGSSL without DA, where all methods exhibit a decline in F1 scores as the testing ratio increases, with the UPPC method consistently outperforming others. In subplot (b), the inclusion of DA techniques results in improved F1 scores across all methods, particularly benefiting the UPPC method, which shows a significant performance edge. Subplot (c) examines the impact of HCBF on the SGSSL model without DA, revealing that HCBF enhances performance, particularly for the UPPC and USP methods, compared to the results seen in subplot (a). The SGSSL model uniquely combines graph-based and attention-based mechanisms to enhance performance in SSL, incorporating Siamese GNNs with MHSA from TE. This integration optimizes feature extraction and improves computational efficiency by introducing a bottleneck layer that filters out noise and preserves relevant features. This approach effectively addresses common issues such as over-smoothing in GCN and GAT, which are particularly beneficial in data-scarce environments like those involving low-resource languages. The framework includes a novel HCBF mechanism, which enhances model reliability by filtering out unreliable pseudo-labeled samples using a composite score of probabilistic confidence, semantic similarity, and model uncertainty. This is particularly useful in NLP tasks such as plagiarism detection and content recommendation, where the ability to process and understand morphologically rich languages like Urdu is crucial. Despite its advantages, the SGSSL model’s dependency on substantial computational resources might limit its applicability in resource-constrained settings. However, its robustness and the ability to perform well with minimal labeled data offer a significant advantage, specifically for organizations lacking extensive annotation resources. The SGSSL framework, with its innovative architecture and strategic integrations, stands as a promising advancement in addressing the challenges of SSL in NLP for under-resourced languages.Fig. 9F1 of the proposed SGSSL model under different testing ratios. F1 of the proposed SGSSL model under different testing ratios. The ablation study of the SGSSL model evaluates the contribution of each component by systematically enabling or disabling layers and modules across the Urdu T-R and P-D datasets. The baseline model employs a Siamese GAT configuration without DA techniques. The ablation study proceeds as follows: adding a bottleneck layer, TE module, DA techniques (BT and MWP), and HCBF to the baseline model progressively. The bottleneck layer slightly enhances feature extraction, and the TE module captures long-range dependencies, resulting in improved performance. DA techniques enhance data diversity and robustness, and HCBF filters out noisy pseudo-labeled data, leading to better classification accuracy. The full SGSSL model, which incorporates all components, yields the highest performance. Table 10 illustrates the layer-wise contribution to F1 score improvement for various datasets. The baseline model shows reasonable performance but struggles with ternary classification. We incorporate a linear bottleneck in our SGSSL model to enhance computational efficiency. We evaluated our model with and without a bottleneck layer, revealing significant performance enhancements. The bottleneck layer serves as a dimensionality reduction component, compressing input features into a lower-dimensional space. This reduction decreases computational complexity, enabling the model to concentrate on essential information while filtering out irrelevant details. Models with a bottleneck require fewer parameters and exhibit improved feature abstraction, leading to faster convergence and reduced memory usage. Consequently, this results in more resource-efficient models that operate more swiftly. Furthermore, the linear bottleneck enhances computational efficiency by simplifying the model architecture and improving interpretability, as it does not require non-linear activation functions. Overall, incorporating the bottleneck layer yields a 1.2% performance improvement. The TE module adds 2.5%, DA techniques yield a 3.7% increase, and HCBF further boosts performance by 5.2%. The full SGSSL model achieves the highest performance, with an average F1 improvement of 8.1% over the baseline configuration. In addition to predictive performance, we also analyzed the computational requirements of the proposed SGSSL framework. Since practical deployment depends not only on accuracy but also on efficiency, we report the total training time, inference latency, memory usage, and parameter count of the complete framework under different configurations. The results, presented in “Computational efficiency”, show that the full SGSSL model introduces moderate computational overhead but remains practical for offline and near-real-time deployment. Error analysis Table 20 examines failure modes. The baseline suffers high ambiguous case errors (38.7%), reduced 37.2% by TE modules. Noisy labels disproportionately affect UTRD-Phr-23 (41.5% FN rate), mitigated 53.1% by HCBF. Figure 10 visualizes the reduction in error categories across configurations. DA techniques reduce false negatives in UPPC/USP by 9.8%. The full model maintains 28.4% FN rate in UTRD-Phr-23, indicating persistent challenges in Urdu phrase-level disambiguation.Table 20Error distribution by failure mode.ConfigurationUPPCUSPSUSPCCOUNTERUSTRCUTRD-Phr-23Ambiguous casesNoisy labelFN%FN%FP%FN%FP%FN%Error%Error%Baseline23.624.127.331.228.741.538.732.4+ DA19.820.323.127.524.936.232.527.3+ HCBF17.216.520.424.121.332.728.119.8Full model14.313.717.921.518.228.424.315.2Significant values are in bold. Table 20Error distribution by failure mode.ConfigurationUPPCUSPSUSPCCOUNTERUSTRCUTRD-Phr-23Ambiguous casesNoisy labelFN%FN%FP%FN%FP%FN%Error%Error%Baseline23.624.127.331.228.741.538.732.4+ DA19.820.323.127.524.936.232.527.3+ HCBF17.216.520.424.121.332.728.119.8Full model14.313.717.921.518.228.424.315.2Significant values are in bold. Error distribution by failure mode. Significant values are in bold. Fig. 10Error distribution analysis across configurations, highlighting reductions in false negatives, ambiguous cases, and noisy-label errors achieved by SGSSL components. Error distribution analysis across configurations, highlighting reductions in false negatives, ambiguous cases, and noisy-label errors achieved by SGSSL components. Key insights Component criticality: TE modules contribute most to semantic tasks (8.1% drop when removed), while HCBF provides broadest robustness (4.3% avg gain)DA synergy: BT+MWP combination outperforms individual techniques by 3.1-7.2%, with HCBF amplifying gains by 19.3%Resource tradeoffs: full configuration delivers 8.1% F1 gain at 2.2× computational costError reduction: HCBF cuts label noise errors by 53.1%, TE reduces ambiguous cases by 37.2%Remaining challenges: UTRD-Phr-23 requires specialized handling of phrase-level semantics (28.4% FN rate) Component criticality: TE modules contribute most to semantic tasks (8.1% drop when removed), while HCBF provides broadest robustness (4.3% avg gain) Component criticality: TE modules contribute most to semantic tasks (8.1% drop when removed), while HCBF provides broadest robustness (4.3% avg gain) DA synergy: BT+MWP combination outperforms individual techniques by 3.1-7.2%, with HCBF amplifying gains by 19.3% DA synergy: BT+MWP combination outperforms individual techniques by 3.1-7.2%, with HCBF amplifying gains by 19.3% Resource tradeoffs: full configuration delivers 8.1% F1 gain at 2.2× computational cost Resource tradeoffs: full configuration delivers 8.1% F1 gain at 2.2× computational cost Error reduction: HCBF cuts label noise errors by 53.1%, TE reduces ambiguous cases by 37.2% Error reduction: HCBF cuts label noise errors by 53.1%, TE reduces ambiguous cases by 37.2% Remaining challenges: UTRD-Phr-23 requires specialized handling of phrase-level semantics (28.4% FN rate) Remaining challenges: UTRD-Phr-23 requires specialized handling of phrase-level semantics (28.4% FN rate) Discussion The SGSSL framework effectively addresses Urdu T-R and P-D by integrating graph-based and attention-based mechanisms. Leveraging node aggregation from Siamese GNNs and MHSA from TE, SGSSL captures both semantic relationships and structural dependencies while mitigating over-smoothing issues inherent in standard GCN and GAT. The bottleneck layer reduces dimensionality, filters noise, and preserves salient features, thereby enhancing computational efficiency and minimizing overfitting risks. The synergy between the local aggregation capabilities of GCN and GAT and the global context modeling of TE modules significantly enhances task-specific performance. Empirical evaluations demonstrate that this integrated methodology substantially improves model robustness, particularly in scenarios with limited labeled data. A pivotal advancement within SGSSL is the HCBF mechanism, which synthesizes probabilistic confidence, semantic similarity, and model uncertainty into a composite confidence score. This composite score enhances pseudo-label reliability by effectively filtering out unreliable data, addressing a fundamental challenge in SSL. The bottleneck feature-refinement layer plays a crucial role in balancing efficiency and performance. By compressing high-dimensional embeddings into a compact representation using a linear projection, it reduces redundancy while preserving essential semantic information. This design enables faster convergence and improves computational efficiency without introducing additional non-linear complexity. From a deployment perspective, the proposed SGSSL framework introduces additional computational overhead compared to the baseline Siamese GAT configuration due to the inclusion of TE, DA, and HCBF modules. Nevertheless, the measured inference latency remains within a practically manageable range, suggesting that the framework is feasible for real-world offline and near-real-time applications. This trade-off is justified by the substantial improvement in classification performance and robustness observed across all benchmark datasets. Practically, the SGSSL framework is highly applicable to NLP tasks in under-resourced languages, including plagiarism detection, content verification, and paraphrase identification. Its capability to manage Urdu’s morphological complexity is further bolstered by leveraging contextual embeddings from the pre-trained Urdu RoBERTa model. Additionally, the integration of DA techniques such as BT and MWP enhances data diversity and mitigates overfitting, thereby increasing the model’s robustness in real-world applications. An important direction for future research is the integration of Explainable AI (XAI) techniques into the SGSSL framework. Although the proposed model achieves strong performance, it operates as a black-box system, making it difficult to interpret why two texts are classified as verbatim, partially paraphrased, or paraphrased. Incorporating XAI methods such as attention visualization, gradient-based attribution techniques (e.g., Integrated Gradients or SHAP), and graph-specific explanation approaches (e.g., GNNExplainer) can provide insights into which tokens, phrases, or graph connections contribute most to the model’s decisions. Such interpretability is particularly important in applications like plagiarism detection and content verification, where users require transparent and justifiable predictions. Therefore, enhancing the SGSSL framework with explainability mechanisms represents a promising direction for improving both trustworthiness and usability. Despite its strengths, the SGSSL model’s reliance on substantial computational resources may limit its applicability in resource-constrained environments. Nonetheless, its robustness and effectiveness with minimal labeled data present notable advantages for organizations with limited annotation resources. Future research should investigate additional DA techniques and advanced filtering mechanisms to further bolster SSL models’ robustness and accuracy. Moreover, extending the framework to other NLP tasks and optimizing it for computational efficiency will enhance its applicability and practical utility. Conclusion This study introduces the SGSSL framework, a novel approach for addressing text reuse and paraphrase detection challenges in low-resource languages, with a focus on Urdu. By integrating Siamese Graph Neural Networks with Transformer-based architectures, SGSSL combines graph-based methodologies (GCN and GAT) with MHSA, effectively capturing complex linguistic patterns inherent in under-resourced languages. The incorporation of strategies such as BT, MWP, and the HCBF technique enhances the framework’s accuracy, particularly in settings with limited labeled data. These components collectively enable SGSSL to discern nuanced linguistic differences, crucial for accurately distinguishing paraphrased and reused content. However, the framework encounters challenges with very short text pairs, where limited contextual information compromises accuracy, as demonstrated by performance on datasets like UTRD-Phr-23. Additionally, the computational intensity of the models restricts deployment in resource-constrained environments, such as mobile devices or embedded systems. Addressing these limitations is essential for real-world applicability and scalability. Despite these challenges, the SGSSL framework holds significant implications for NLP, providing a robust foundation for extending its application to other low-resource languages with unique linguistic characteristics. Its potential applications in plagiarism detection, content verification, and paraphrase identification make it a valuable asset for both academic research and industry practices. Future work should focus on refining the framework by exploring additional DA techniques, enhancing filtering mechanisms, and optimizing computational efficiency to facilitate deployment across diverse operational contexts. Expanding SGSSL to support a broader range of low-resource languages will further validate its generalizability and adaptability, establishing it as a pivotal tool in advancing NLP technologies for under-resourced linguistic communities. In addition, incorporating Explainable AI (XAI) techniques is a promising direction for future work. While SGSSL demonstrates strong predictive performance, improving its interpretability would allow users to better understand the reasoning behind similarity decisions. Techniques such as attention analysis, feature attribution methods, and graph-based explanation models can be integrated to highlight influential words, semantic relationships, and graph structures that drive classification outcomes. This would enhance the transparency, trust, and practical adoption of the framework in real-world applications.