Plagiarism detection across languages: a comprehensive study of Arabic and English-to-Arabic long documents

Plagiarism detection

Plagiarism detection systems aim to identify similarities between textual content, which may occur within the same language (monolingual plagiarism) or across different languages (cross-lingual plagiarism). These systems must overcome challenges such as linguistic diversity, disguised textual modifications, and scalability. The detection of cross-language plagiarism, especially in the context of machine translation and multilingual digital content, is becoming increasingly critical. Traditional monolingual methods struggle with sophisticated forms of plagiarism involving translation. While some methods attempt to use machine translation to unify texts into a common language, they are limited by translation inaccuracies. Recent approaches, therefore, focus on semantic and syntactic analysis, albeit at higher computational costs and requiring parallel corpora.

Monolingual plagiarism detection

Monolingual plagiarism detection traditionally relied on exact and approximate text matching techniques. However, recent research emphasizes hybrid approaches that integrate multiple similarity measures to improve robustness and accuracy.

Arabi & Akbari (2022) propose two approaches for monolingual plagiarism detection. The first combines FastText pre-trained word embeddings with term frequency-inverse document frequency (TF-IDF) weighting to form structural-semantic matrices of sentences. The second employs WordNet ontology alongside TF-IDF weighting to create similar matrices. Both approaches use a two-stage filtering process at the document and sentence levels, leveraging the Bag of Words (BoW) technique to narrow the search space. After filtering, structural and semantic matrices are formed, and their similarity is calculated using the Dice similarity measure. Documents exceeding a predefined similarity threshold are flagged as plagiarized. Validated on the PAN-PC-11 dataset, their FastText-based method achieved an F1-score of 90.97, while the WordNet-based method reached an F1-score of 90.00. These results highlight the effectiveness of combining lexical and semantic features for detecting sophisticated forms of plagiarism.

Another innovative approach is the multi-agent indexing system (MAIS) introduced by Zouaoui & Rezeg (2022). This index-based framework integrates semantic indexing with multi-agent systems, utilizing the Sorensen-Dice similarity measure for document retrieval. MAIS leverages Arabic ontology (ShemNet) to establish semantic closeness between entities in documents. The system proved effective at identifying complex types of plagiarism, such as synonym replacement and sentence scrambling, achieving superior performance on the AraPlagDet dataset.

Nagoudi, Cherroun & Alshehri (2018) investigated disguised plagiarism, focusing on advanced obfuscation techniques such as paraphrasing, word replacement, and structural manipulation in Arabic texts. Their research employs word embeddings and machine learning models, including support vector machines (SVM) and Random Forests, alongside sophisticated NLP techniques like bag-of-meanings and semantic similarity measures. The combination of lexical, syntactic, and semantic features proved effective in identifying complex plagiarism patterns.

Aliane & Aliane (2020) investigated the effectiveness of Siamese neural architectures for Arabic semantic textual similarity (STS) and plagiarism detection. Their study evaluated three Siamese-based models: bidirectional long short-term memory network (BiLSTM), convolutional neural network (CNN), and a transformer-based model using AraBERT. Each model consisted of a shared architecture for encoding sentence pairs and a comparison layer using similarity metrics like cosine similarity and dot product. They tested their models on two Arabic datasets: the PAN 2015 corpus for plagiarism detection and the Mawdoo3 question similarity dataset. The results demonstrated that the transformer model consistently outperformed the others, achieving a Pearson correlation of up to 0.93 on the PAN dataset.

Their work confirms the relevance of Siamese networks for Arabic STS and highlights the superiority of contextualised transformer-based embeddings in low-resource settings. While their focus was on sentence and short paragraph comparison, their findings provide a foundational baseline for extending Siamese architectures to longer document-level tasks, such as plagiarism detection. By demonstrating the robustness of AraBERT in the Siamese setting, this work paves the way for exploring more scalable transformer variants like Longformer to handle longer Arabic texts more effectively.

Saidi, Jarray & Alsuhaibani (2023) proposed the Siamese bidirectional encoder representations from transformers (SiameseBERT) a BERT-based Siamese network enhanced with a soft attention mechanism to assess semantic textual similarity (STS) for Arabic. Their approach embeds input sentences using various Arabic BERT models, including AraBERT and CAMeL-BERT, and applies a soft alignment mechanism to capture word-level interactions between sentence pairs. The model architecture integrates contextual embeddings through a Siamese structure, followed by attention-driven interaction, pooling, and prediction layers. Their experiments on benchmark Arabic STS datasets demonstrated the effectiveness of this architecture, achieving a high Pearson correlation score of up to 0.925 on the MSR-Paraphrase dataset.

The study highlights the strength of combining BERT embeddings, attention mechanisms, and Siamese networks for Arabic STS—a task crucial for applications such as plagiarism detection. It also shows that the inclusion of an attention mechanism consistently improves model performance across datasets. While their work focused on sentence-level similarity using transformer-based models limited by input length, it provides a strong foundation for exploring architectures like Longformer that address longer textual inputs, which are essential in document-level tasks such as Arabic plagiarism detection.

Cross-lingual plagiarism detection

Cross-lingual plagiarism detection refers to the identification of semantically equivalent or paraphrased content that has been translated between different languages. Unlike monolingual plagiarism, which compares text in the same language, cross-lingual detection must account for variations introduced by translation, including differences in syntax, morphology, and idiomatic expressions. This task typically involves a preprocessing step such as machine translation to bring texts into a common representation space, followed by semantic similarity assessment.

In the English-Arabic context, this challenge is further amplified due to the rich morphology of Arabic and structural asymmetry between the two languages. Therefore, effective cross-lingual detection systems must either incorporate a robust translation layer or utilize multilingual encoders capable of capturing aligned semantics across languages.

The following studies illustrate some of the techniques used in this emerging area of research. Alzahrani & Aljuaid (2022) address these challenges by employing deep neural networks equipped with semantic-rich features, such as topic similarity, named entity recognition (NER), semantic role labeling (SRL), and bag-of-meanings (BoM). Their models, configured with two or more hidden layers and Rectified Linear Unit (ReLU) activation functions, are trained for binary classification tasks using binary cross-entropy loss. Additionally, they implement multi-class classification for translation types (e.g., literal, paraphrased, summarized, independently written) using categorical cross-entropy. Their experiments demonstrated that all deep neural network models outperformed traditional classifiers like SVM or logistic regression, achieving a maximum classification accuracy of 97.01%. For cross-lingual semantic similarity prediction, they obtained an accuracy of 71.13%.

Alotaibi & Joy (2021) contribute to English-Arabic plagiarism detection by extracting semantic and syntactic features integrated into machine learning classifiers with multilingual encoders. Their system evaluates these features on the SemEval-2017 dataset, revealing that the support vector classifier (SVC) yielded the best F1-score of 0.879. This study underscores the potential of integrating linguistic features with robust classification models to address the structural divergence between English and Arabic.

Additional contributions include the work by Aljuaid (2020), who proposed a word-embedding and IDF-weighted model for English–Arabic plagiarism detection using comparable corpora, achieving over 82% sentence-level accuracy. Earlier, Aljohani & Mohd (2014) introduced an English-Arabic cross-language detection method using machine translation followed by a winnowing fingerprinting algorithm, reporting a recall of 81% and precision of 97%. These studies illustrate the diverse techniques—spanning embedding models, translation layers, and fingerprinting—that have been employed to address translation-induced variations in English–Arabic documents.

Advances in neural architectures have revolutionized text similarity analysis, enabling models to capture complex semantic relationships. For example, Unified transformer models like T5 (Raffel et al., 2020) simplify NLP by framing all tasks as text-to-text problems. Elmadany, Nagoudi & Abdul-Mageed (2023) extend this paradigm with Octopus, a multitask model tailored for Arabic NLP tasks. By leveraging task-specific fine-tuning, Octopus enhances performance on low-resource languages, making it a promising tool for future similarity analysis and text generation. The studies reviewed here illustrate a dynamic and evolving field, addressing critical challenges in monolingual and cross-lingual plagiarism detection, metric development, and neural model design. Despite significant progress, open challenges include scalability, computational efficiency, and the detection of obfuscated plagiarism, which remain areas for further exploration.

Gaps in related work and proposed contributions

The proposed research builds upon the foundations laid by previous studies while addressing significant gaps in the field of Arabic plagiarism detection. Table 1 provides a comprehensive comparison between the surveyed studies and the proposed work, focusing on dataset usage, methodology, ability to handle complex plagiarism types, scalability, cross-lingual capabilities, and evaluation rigor.

Plagiarism detection workflow

The proposed workflow aims to detect potential plagiarism across Arabic documents and cross-lingual sources. The system operates in three main stages:

• 1.

  Language standardization: Input text pairs (source and suspicious documents) are first analyzed to ensure they are both in Arabic. If one document is non-Arabic, a translation layer converts it into Arabic to maintain consistency.

• 2.

  Feature representation: The Arabic text pairs are passed through a transformer-based Siamese network architecture (using either Longformer or AraT5 encoders). These models are fine-tuned specifically for Arabic text to capture both semantic and structural relationships.

• 3.

  Classification: The embeddings generated by the Siamese network are compared and passed through a classification head to determine whether plagiarism exists and, if so, to classify the type of plagiarism (e.g., No Obfuscation, Artificial Obfuscation, Simulated Obfuscation).

The overall workflow is illustrated in Fig. 2, depicting the sequential stages from input document processing to final plagiarism classification.

This modular pipeline ensures robust handling of linguistic variations, cross-lingual text pairs, and complex plagiarism patterns, providing a flexible and scalable foundation for document similarity analysis.

Data pre-processing and feature engineering

Preprocessing is a crucial step to address dataset imbalances and adapt text inputs for the limitations of transformer-based architectures. The preprocessing phase involves the following key steps:

• 1.

  Context size management: Due to the limited context window of Longformer (4,096 tokens) and AraT5 (1,024 tokens), each document was truncated to fit within these token limits. Careful attention was given to ensure that plagiarized text offsets remain within these truncated contexts to prevent loss of critical information.

• 2.

  No-plagiarism label creation: A “No Plagiarism” label was introduced by extracting non-overlapping text segments from the dataset where no plagiarism indicators were present. This additional label enhances the model’s ability to differentiate between genuine and plagiarized content.

• 3.

  Tokenization and encoding: Text pairs were tokenized using their respective model tokenizers (LongformerTokenizer or T5Tokenizer) to ensure compatibility with the downstream architecture.

The effects of these preprocessing steps are evident in the adjusted label distributions shown in Fig. 3. These steps collectively ensure that the dataset is optimized for transformer-based architectures, mitigating common pitfalls such as sequence truncation and class imbalance.

Translation layer

The integration of a translation layer in our plagiarism detection system addresses the challenge of cross-lingual plagiarism, particularly when source and suspicious documents are written in different languages. This layer ensures linguistic uniformity by converting non-Arabic text into Arabic, allowing seamless processing by Arabic-specific NLP models. Given the structural and syntactic complexities of the Arabic language, selecting an optimal translation system plays a pivotal role in preserving both semantic and contextual fidelity during translation.

In this study, we evaluated multiple state-of-the-art translation models, including both commercial APIs and open-source large language models (LLMs), based on their suitability for the Arabic language. These models represent different architectures, ranging from sequence-to-sequence neural networks to transformer-based systems, and exhibit varying strengths in handling context-rich translations, idiomatic expressions, and morphological variations.

Commercial translation systems, such as Google Translate, are widely recognized for their robust multilingual capabilities and are built upon extensive neural machine translation (NMT) frameworks (Wu et al., 2016). However, their performance often varies across languages, especially for morphologically rich languages like Arabic. Similarly, Argos Translate, an open-source translation toolkit, leverages pre-trained neural network models optimized for multilingual tasks and provides customization capabilities for domain-specific translations (Finlay & Argos Translate Contributors, 2025).

In contrast, recent advancements in transformer-based models, such as mBART-Large-50-Many-to-Many-MMT, have shown promise in addressing cross-lingual tasks by extending pre-training to multiple languages (Liu et al., 2020). mBART employs a denoising autoencoder mechanism for sequence-to-sequence tasks, which improves its ability to maintain syntactic and semantic fidelity during translation. Another significant contribution comes from models like Gemini Flash 1.5 and SILMA-9B-Instruct-v1.0, both fine-tuned on multilingual corpora and optimized for translation-specific tasks. Gemini Flash 1.5, a generative transformer-based language model, incorporates extensive contextual embeddings to address nuances specific to Arabic, while SILMA-9B-Instruct-v1.0 adapts its decoder architecture for multilingual translation scenarios (Qian et al., 2024).

Additionally, Marefa-MT-En-Ar, a specialized English-to-Arabic translation model, was included in our evaluation due to its focus on Arabic’s unique linguistic features. Marefa-MT utilizes domain-specific vocabulary and fine-tuned attention mechanisms to ensure that translations retain grammatical coherence and stylistic appropriateness in Arabic text (marefa-nlp, 2023).

The translation layer, therefore, serves as a critical intermediary component, ensuring that non-Arabic content can be seamlessly integrated into our Arabic-focused plagiarism detection pipeline. By systematically evaluating multiple translation models, we have ensured that the selected system not only meets linguistic accuracy requirements but also minimizes semantic drift, thereby enhancing the reliability of the overall detection process.

Plagiarism detection layer

In this section, we present the core architecture utilized for detecting plagiarism across Arabic documents. At the heart of our system lies a Siamese neural network (SNN) (Bromley et al., 1993), a well-established framework for similarity tasks, adapted here to address the nuances of Arabic textual data.

Siamese network

SNNs have emerged as a powerful architecture for measuring the similarity between two inputs. Originally proposed for tasks such as signature verification (Bromley et al., 1993), they have been effectively adapted for applications in text similarity and plagiarism detection. In our implementation, a Siamese network is employed to compare two text documents and classify their relationship into predefined categories, such as “No Plagiarism,” “Artificial Obfuscation,” and “Simulated Obfuscation.”

Architecture overview The SNN consists of two identical sub-networks that share the same architecture, weights, and parameters. Each sub-network processes one of the input documents independently, transforming it into a fixed-size embedding vector. These embeddings are then concatenated and passed through fully connected dense layers for classification.

The Siamese neural network architecture employed in this study offers several key advantages for plagiarism detection tasks. By utilizing shared weights across both input branches, the network ensures symmetric processing of document pairs, reducing computational overhead and guaranteeing consistent feature extraction. Additionally, the architecture excels in capturing semantic similarity, making it highly effective in identifying subtle differences between documents, such as paraphrased text or structural variations.

The integration of Transformer-based encoders (AraT5 and Longformer) enhances the network’s capability to process both monolingual and cross-lingual text pairs. The embeddings generated by these encoders play a pivotal role in capturing intricate structural and semantic relationships. Furthermore, the adoption of weighted cross-entropy loss and Dice loss effectively mitigates class imbalance issues, improving the network’s predictive accuracy across different types of plagiarism.

This carefully designed architecture, illustrated in Fig. 4, provides a robust foundation for accurately detecting and categorizing plagiarism in Arabic and cross-lingual text datasets.

1. Document encoding: The two input documents, denoted as $X_1$ and $X_2$, are passed through a shared transformer-based encoder, such as Longformer (Beltagy, Peters & Cohan, 2020) or AraT5 (Elmadany, Nagoudi & Abdul-Mageed, 2023). The encoder produces contextual embeddings for each document:

(1) $$E_1=f_{\theta }(X_1),E_2=f_{\theta }(X_2)$$where:

– $f_{\theta }$ represents the shared transformer encoder, and

– $E_1$, $E_2$ are the contextual embeddings of the input documents.

2. Mean pooling: To produce a fixed-size representation for each document, mean pooling is applied across the token embeddings:

(2) $$E_1=\frac{1}{n}\sum _{i=1}^n{A}_{1i},E_2=\frac{1}{m}\sum _{j=1}^m{A}_{2j}$$where:

– $n$ and $m$ are the token lengths of the documents.

– $A_{1i}$ and $A_{2j}$ represent the embeddings of individual tokens in $X_1$ and $X_2$, respectively.

The mean pooling operation ensures that variable-length input sequences are mapped to fixed-size vector embeddings.

3. Embedding concatenation: The resulting document embeddings, $E_1$ and $E_2$, are concatenated into a single vector:

(3) $$E=[E_1\Vert E_2]$$

This concatenated vector E serves as the combined representation of the document pair.

4. Dense layers and classification: The concatenated embeddings are passed through a series of dense layers, where transformations are applied to extract high-level semantic relationships:

(4) $$H=\mathrm{R}\mathrm{e}\mathrm{L}\mathrm{U}(W_{1}E+b_1)$$

(5) $$H^{\mathrm{\prime }}=\mathrm{D}\mathrm{r}\mathrm{o}\mathrm{p}\mathrm{o}\mathrm{u}\mathrm{t}(H)$$

(6) $$y=\mathrm{S}\mathrm{o}\mathrm{f}\mathrm{t}\mathrm{m}\mathrm{a}\mathrm{x}(W_2{H}^{\mathrm{\prime }}+b_2)$$where:

– $W_1$ and $W_2$ are weight matrices for the dense layers.

– $b_1$ and $b_2$ are bias terms.

– H represents the hidden layer activations after applying the ReLU non-linearity.

– H′ represents the output after dropout regularization, and

– $y$ is the output probability distribution over the predefined classes.

5. Similarity representation: In our implementation, similarity between document embeddings is primarily learned implicitly through the dense layers, which are optimized for a classification objective rather than explicit similarity computation. The dense layers are trained to map the embeddings into a space where relationships between document pairs correspond to their similarity class labels (e.g., No Plagiarism, Artificial Obfuscation, Simulated Obfuscation). Our approach relies on the final softmax classification layer to determine similarity categories. This layer inherently learns a more complex representation of semantic and structural similarities during the training process, guided by the chosen loss functions (weighted cross-entropy or Dice loss).

6. Loss function: The network is optimized using either weighted cross-entropy loss or Dice loss, both of which address the class imbalance issue in our dataset as will be detailed in section “Loss functions”.

7. Prediction: The final predicted class is determined by selecting the maximum probability from the output layer:

(7) $$\hat{y}=\arg max(y)$$

This Siamese network, combined with transformer-based encoders and an effective loss function, provides a robust foundation for accurately detecting and classifying plagiarism types across Arabic and cross-lingual datasets.

Evaluation metrics

Precision: Precision evaluates the proportion of correctly predicted positive instances out of all predicted positive instances. It measures the model’s accuracy in making positive predictions and is defined as:

(12) $$\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}=\frac{\mathrm{T}\mathrm{P}}{\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{P}}$$where TP is the number of true positives (correctly predicted positive instances), and FP is the number of false positives (incorrectly predicted positive instances).

Recall: Recall (or Sensitivity) measures the proportion of correctly predicted positive instances out of all actual positive instances. It reflects the model’s ability to identify all relevant instances and is defined as:

(13) $$\mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}=\frac{\mathrm{T}\mathrm{P}}{\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{N}}$$where TP is the number of true positives, and FN is the number of false negatives (actual positive instances incorrectly predicted as negative).

F1-score: F1-score is the harmonic mean of precision and recall, providing a balanced measure of a model’s accuracy when dealing with imbalanced datasets. It is particularly useful when the dataset has a skewed distribution of classes. F1-score is defined as:

(14) $$\mathrm{F}1\text{-}\mathrm{s}\mathrm{c}\mathrm{o}\mathrm{r}\mathrm{e}=2\cdot \frac{\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}\cdot \mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}}{\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}+\mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}}$$

These metrics were chosen to comprehensively evaluate the models, as they provide insights into different aspects of performance. Precision emphasizes minimizing false positives, Recall focuses on reducing false negatives, and F1-score offers a balanced evaluation, especially when dealing with class imbalances in the dataset.

Hardware and software infrastructure

All experiments were conducted on a system equipped with an AMD Ryzen 9 5900X processor, 128 GB of RAM, and an NVIDIA RTX 3090 GPU. The experiments were implemented using Python version 3.11.0. To maintain a consistent environment for all experiments, the same hardware and software configurations were used throughout the study.

Translation layer results

The performance of each translation model was evaluated using the previously mentioned metrics, and the results are summarized in Table 2.

From the results, Argos-Translate achieved the highest scores across all metrics, indicating its superior ability to generate accurate and contextually appropriate translations for our specific use case. In contrast, Gemini 1.5 flash showed relatively lower scores, suggesting limitations in handling the required linguistic and contextual nuances for this task.

These findings highlight the importance of selecting suitable translation model based on their performance metrics to enhance the effectiveness of the Arabic plagiarism detection process.

Plagiarism detection results

Strengths

The proposed system integrates Arabic-specific transformer models alongside scalable architectures capable of handling long text sequences, combining the linguistic strengths of AraT5 and the extended contextual abilities of Longformer. The incorporation of Siamese neural networks allowed for effective pairwise document comparison, while weighted cross-entropy and Dice loss addressed dataset imbalance challenges, enhancing model reliability. Furthermore, the translation layer bridges the gap for cross-lingual plagiarism detection, ensuring a broader applicability of the system across multilingual contexts.

Limitations

Despite the promising results, the system faced limitations in both data availability and computational efficiency. The ExAraCorpusPAN2015 dataset, while valuable, remains limited in both size and label diversity, posing challenges in generalizing findings across broader domains. Additionally, the reliance on Longformer, which was not pre-trained specifically on Arabic data, introduced inherent limitations in capturing Arabic linguistic subtleties. Fine-tuning helped mitigate this to an extent, but an Arabic-pretrained Longformer remains a future need. Moreover, computational resource constraints restricted exploration into more extensive model ensembles or training deeper configurations.

Opportunities for future research

Future directions can focus on developing larger, domain-diverse Arabic plagiarism datasets with balanced class distributions to overcome dataset limitations. Training an Arabic-specific Longformer model from scratch could address the contextual nuances of long Arabic text sequences more effectively. Additionally, integrating hybrid architectures that combine transformer-based semantic analysis with statistical and linguistic features might further enhance performance. Exploring advanced multi-modal techniques, where textual data is complemented with metadata (e.g., author style, citation patterns), could open new avenues for detecting more sophisticated forms of plagiarism.

In summary, this study establishes a robust foundation for Arabic plagiarism detection, addressing critical technical and methodological challenges, while laying the groundwork for future advancements in the field.