This work presents an eficient and scalable framework for detecting plagiarism in large document collections using sentence embedding models and fast approximate nearest neighbor search. Each document is segmented into overlapping chunks using a sliding window approach and encoded into dense semantic vectors using the pretrained intfloat/e5-base-v2 model. To accelerate semantic comparison, we first apply document-level ifltering using global embeddings, followed by chunk-level matching via FAISS for GPU-accelerated top-k retrieval. The proposed system significantly reduces runtime through embedding reuse and candidate pruning, while maintaining strong detection performance on large-scale benchmark datasets. It is fully modular, supports both CPU and GPU execution, and is compatible with the TIRA evaluation platform. Source code is publicly available at https://github.com/koppen777/plagiarism-detectio.
Automatic plagiarism detection has become increasingly important in the era of large-scale digital content and machine-generated text. The PAN shared task provides a benchmark for evaluating systems on challenging plagiarism scenarios involving paraphrasing, translation, and AI-generated rewriting. Given a suspicious document and a pool of source documents, the goal is to accurately locate plagiarized passages and align them with their original sources.
While traditional approaches often rely on lexical overlap or character-based similarity, such methods are limited when facing paraphrased or semantically modified text. Recent work has explored the use of sentence embeddings to capture semantic similarity; however, many of these approaches are either computationally expensive or require exhaustive pairwise comparisons. We observed that models like all-MiniLM-L6-v2 were not robust enough for this task, producing low recall. These limitations call for a more eficient and scalable solution.
To address this, we propose a lightweight and eficient two-stage plagiarism detection system that combines sentence embeddings with FAISS-based retrieval. We segment suspicious and source documents into overlapping chunks using a sliding window strategy and embed them using the intfloat/e5-base-v2 model, which we found to outperform MiniLM in this task. Document-level embeddings are used to iflter unrelated sources, followed by chunk-level approximate nearest neighbor search via FAISS to detect plagiarism. We further optimize performance by tuning key parameters such as window size, stride, similarity threshold, and the number of top-K candidates. Our system is compatible with TIRA and supports both CPU and GPU execution, making it suitable for large-scale evaluation environments such as Kaggle. This work was carried out by the Foshan University Artificial Intelligence Laboratory, which focuses on natural language processing and computer vision research.
The PAN plagiarism detection shared task has evolved into a standard benchmark for evaluating text
reuse systems under realistic and adversarial scenarios. In recent years, top-performing systems have
shifted from lexical fingerprinting methods toward semantic-aware techniques. For instance, the
topranked system in PAN 2023 [
The 2025 edition of the PAN shared task [
Broadly, existing methods for semantic plagiarism detection fall into three categories. The first category
uses traditional lexical features (e.g., n-gram overlap, edit distance), which are fast but vulnerable to
paraphrasing. The second category leverages supervised classifiers (e.g., BERT or RoBERTa fine-tuned
on similarity datasets), which ofer strong performance but require labeled training data and are slow
at inference. The third category, which our method belongs to, applies sentence embedding models
such as Sentence-BERT [
A major limitation of many prior systems is their reliance on exhaustive pairwise comparison between all suspicious and source chunks, which becomes infeasible at scale. Furthermore, many public implementations either neglect document-level filtering or use unoptimized thresholds, leading to suboptimal trade-ofs between precision and recall. Our work addresses these issues by incorporating document-level embedding filtering, parameter tuning, and embedding reuse to maximize detection quality while minimizing runtime.
Our plagiarism detection system adopts a two-stage semantic retrieval pipeline, as shown in Figure 1. The system is designed to eficiently identify semantic reuse across a large number of document pairs by leveraging sentence embeddings and FAISS-based approximate nearest neighbor (ANN) search. The method consists of the following key components:
Each suspicious and source document is preprocessed through sentence splitting using regular expressions. Specifically, we apply a heuristic rule that splits the text at punctuation followed by whitespace (e.g., period, exclamation mark, or question mark), which corresponds to the regular expression (?<=[.!?])\s+. This lightweight rule performs reasonably well on English text and avoids the overhead of full syntactic parsing.
After sentence segmentation, we apply a sliding window mechanism to generate overlapping textual chunks of sentences (we set = 6 with stride 2), which are used as the basic comparison units in the downstream embedding and retrieval process.
We encode each chunk using a pretrained Sentence Transformer model. After evaluating several candidates, we selected intfloat/e5-base-v2 as it provides stronger performance than all-MiniLM-L6-v2 in semantic detection tasks. Each document also has a full-text embedding to support coarse filtering.
To reduce the number of comparisons, we compute a global embedding for each suspicious document and retrieve its Top-K most similar source documents based on cosine similarity. This document-level ifltering reduces the chunk comparison space from thousands to just a dozen documents per query.
We use FAISS [
Matched chunks are mapped back to character ofsets using sentence-level ofsets. Overlapping or adjacent segments are merged if they are within a certain character distance (e.g., 30 chars). The final predicted plagiarized spans are then saved in the XML format required by the PAN evaluation system.
We evaluated our system on the pan25-generated-plagiarism-detection-validation dataset released by the PAN 2025 shared task organizers. The dataset contains suspicious documents and source documents, with manually annotated plagiarism cases provided in XML format.
Each document is preprocessed using sentence tokenization based on regular expressions. We use a sliding window of 6 sentences with a stride of 2 to generate overlapping textual chunks. Sentence embeddings are computed using the intfloat/e5-base-v2 model, chosen for its strong semantic retrieval performance. FAISS is configured with inner product indexing (normalized vectors), and top-5 chunk matches are retrieved for each suspicious chunk. To reduce search space, we first compute a document-level embedding and filter the top-5 candidate source documents.
All experiments are conducted on the Kaggle platform using GPU for embedding and CPU for FAISS retrieval. Embeddings are cached and reused to reduce runtime overhead. Table 1 summarizes the key hyperparameters. Evaluation is performed using the oficial PAN evaluation script, which computes precision, recall, and F1-score based on the overlap of predicted and ground-truth spans. All XML outputs follow the PAN format and are submitted through TIRA for validation.
We observed that increasing the window size and lowering the similarity threshold improved recall without sacrificing much precision. Document-level filtering using global embeddings significantly reduced runtime, from 30+ hours to under 3 hours on the same hardware. Compared to a naive baseline with no filtering and smaller embedding model, our approach ofers better scalability and accuracy.
We presented an eficient two-stage plagiarism detection system based on sentence embeddings and FAISS-based retrieval. By combining document-level filtering and chunk-level approximate matching, our method achieves a strong balance between accuracy and computational eficiency. The use of pretrained sentence encoders such as intfloat/e5-base-v2, along with optimized hyperparameters and embedding reuse, significantly improves detection quality while reducing runtime. Our system is fully compatible with the TIRA evaluation platform and can be deployed on resource-constrained environments such as Kaggle. Our implementation code is publicly available at https://github.com/ koppen777/plagiarism-detection to foster reproducibility and further research in plagiarism detection.
This work is supported by the Social Science Foundation of Guangdong Province, China (No.
GD24CZY02), and by the organizers of the PAN 2025 shared task [
During the preparation of this work, the author(s) used ChatGPT (GPT-4) in order to: grammar and language polishing. After using this tool, the author(s) reviewed and edited the content as needed and take full responsibility for the publication’s content.