TL;DR
Experimental results indicate that the proposed approach achieves competitive performance compared to existing plagiarism detection systems, and the comparative analysis highlights the strengths and limitations of different word embedding models across datasets.
The findings underscore the potential of advanced NLP techniques to overcome language-specific challenges, providing a foundation for future research in multilingual plagiarism detection and enhancing the development of tools for other languages facing similar challenges.
In this study, a hybrid approach to semantic text similarity combining distributed word embeddings with classical lexical similarity measures is developed. Analyzed are the limitations of modern deep learning models, namely computational overhead and weak interpretability in resource-constrained environments. Proposed is a hybrid architecture that integrates Word2Vec distributed representations with cosine and Jaccard lexical similarity metrics. Investigated is a weighted fusion mechanism that combines vector-based semantic distances with set-theoretic token overlap for robust scoring. Developed is a three-stage processing pipeline covering text preprocessing, sentence embedding generation, and similarity computation and fusion. Established is a tunable weighting parameter that experimentally balances semantic depth against lexical matching precision. Conducted are experimental evaluations on benchmark semantic textual similarity and paraphrase detection datasets using classification metrics. Determined is that the proposed hybrid model attains higher correlation with human judgment than standalone or traditional baselines. Demonstrated is a notable reduction of error rates for exact lexical matches frequently missed by vector-only models. Presented is an efficient and scalable solution that balances computational performance with semantic accuracy for practical tasks.
This study introduces a multi-similarity neural network framework for paraphrase detection, an important task in natural language processing that identifies whether two sentences convey the same meaning using different expressions. The proposed method combines various similarity measures, such as string-based similarity, semantic similarity, and embedding-based similarity, with a deep learning classifier. The framework is structured as a three-phase pipeline: preprocessing, extraction of multiple similarity features, and classification through a neural network. It employs more than 168 string similarity algorithms, semantic measures derived from WordNet, and several pre-trained embedding models to compute similarity scores. These features are aggregated and supplied to a deep neural network to determine whether sentence pairs are paraphrases. The model was evaluated on the Microsoft Research Paraphrase Corpus (MSRP) using accuracy and F1-score as performance metrics. The experimental results indicate that the proposed framework achieves 81.74% accuracy and an F1 Score of 86.6%, surpassing several existing approaches. Overall, the results suggest that integrating diverse similarity measures with neural networks enhances the identification of both explicit and nuanced paraphrases, thereby supporting advancements in text analysis and plagiarism detection systems.
This study introduces an external plagiarism detection framework built on an artificial neural network model and a lexical feature extraction framework adapted to the linguistic features of Arabic, verifying its effectiveness for Arabic plagiarism detection.
Plagiarism remains a serious threat to academic integrity, while most available detection tools operate through literal word matching and are therefore easily defeated by paraphrasing. This study implements the pre-trained IndoBERT model without fine-tuning as a semantic-similarity screening tool for Indonesian-language student assignment documents, using a two-level scheme: a chunk-level similarity score (200 words with a 50-word overlap) as a document-level indicator, and one-to-one sentence alignment as a locator of similar passages. A 768-dimensional embedding is formed through attention-masked mean pooling followed by L2 normalization, and vectors are compared using cosine similarity at a 0.92 threshold. Evaluation was conducted on 20 pairs drawn from 15 documents, of which only 11 were authentic student submissions while 4 were controls derived from a single source document through graded paraphrasing. Similarity scores decreased across the four control conditions ordered by intended paraphrase intensity, namely 100%, 66.67%, 42.34%, and 26.67%, whereas same-theme original pairs averaged 5.93%. Against a proxy sentence-level ground truth, the system obtained 84.79% accuracy, 96.82% precision, 61.13% recall, and a 74.94% F1-score with a 1.19% false positive rate. A User Acceptance Testing exercise with six respondents yielded an index of 82.56%, describing only the participating group rather than general acceptance. These results provide preliminary evidence that the two-level scheme is feasible as an initial semantic-similarity indicator, with the final judgment of plagiarism remaining with the examiner.
Text semantic similarity calculation is a core technology in the field of natural language processing, and it is crucial in scenarios such as question answering systems and information retrieval. Traditional methods based on keyword matching or static word vectors are difficult to capture deep semantics of text and cannot effectively solve the problem of lexical ambiguity. This study aims to systematically evaluate the performance and advantages of the BERT model in text semantic similarity calculation. By constructing a multi-source dataset containing academic literature, online text, and standard evaluation data, and after preprocessing such as text cleaning and word segmentation standardization, a pre-training + finetuning training paradigm was adopted, and the model was trained by setting appropriate hyperparameters. The F1 score and correlation coefficient were used as core indicators to conduct comparative experiments with traditional models such as Word2Vec and GloVe. The results show that the BERT model achieves F1 scores of 0.84 and 0.91 on the CNSE and CNSS datasets, respectively, which are significantly better than traditional models, and it can effectively utilize contextual information to eliminate ambiguity and capture deep semantic relationships. This study validates the effectiveness of the BERT model in semantic similarity calculation, providing more accurate technical support for related application scenarios, and laying the foundation for subsequent model optimization and lightweighting research.