Paraphrase Generation and Identification at paragraph-level

Abstract

The widespread availability of the Internet and the ease of accessing written content have significantly contributed to the rising incidence of plagiarism across various domains, including education. This behaviour directly undermines academic integrity, as evidenced by reports highlighting increased plagiarism in student work. Notably, students tend to plagiarize entire paragraphs more often than individual sentences, further complicating efforts to detect and prevent academic dishonesty. Additionally, advancements in natural language processing (NLP) have further facilitated plagiarism, particularly by using online paraphrasing tools and deep-learning language models designed to generate paraphrased text. These developments underscore the critical need to develop and refine effective paraphrase identification (PI) methodologies.
This thesis addresses one of the most challenging aspects of plagiarism detection (PD): identifying instances of plagiarism at the paragraph-level, with a particular emphasis on paraphrased paragraphs rather than individual sentences. By focusing on this level of granularity, the approach considers both intra-sentence and inter-sentence relationships, offering a more comprehensive solution to the detection of sophisticated forms of plagiarism. To achieve this aim, the research examines the influence of text length on the performance of NLP machine learning (ML) and deep learning (DL) models. Furthermore, it introduces ALECS-SS (ALECS – Social Sciences), a large-scale dataset of paragraph-length paraphrases, and develops three novel SALAC algorithms designed to preserve semantic integrity while restructuring paragraph content. These algorithms suggest a novel approach that modifies the structure of paragraphs while maintaining their semantics. The methodology involves converting text into a graph where each node corresponds to a sentence’s semantic vector, and each edge is weighted by a numerical value representing the sentence order probability. Subsequently, a masking approach is applied to the reconstructed paragraphs modifying the lexical elements while preserving the original semantic content. This step introduces variability to the dataset while maintaining its core meaning, effectively simulating paraphrased text. Human and automatic evaluations assess the reliability and quality of paraphrases, and additional studies examine the adaptability of SALAC across multiple academic domains. Moreover, state-of-the-art large language models (LLMs) are analysed for their ability to differentiate between human-written and machine-paraphrased text. This investigation involves the use of multiple PI datasets in addition to the newly established paragraph-level paraphrases dataset (ALECS-SS).
The findings demonstrate that text length significantly affects model performance, with limitations arising from dataset segmentation. Additionally, the results show that the SALAC algorithms effectively maintain semantic integrity and coherence across different domains, highlighting their potential for domain-independent paraphrasing. The thesis also analysed the state-of-the-art LLMs’ performance in detecting auto-paraphrased content and distinguishing them from human-written content at both the sentence and paragraph levels, showing that the models could reliably identify reworded content from individual sentences up to entire paragraphs. Collectively, these findings contribute to educational applications and plagiarism detection by improving how paraphrased content is generated and recognized, and they advance NLP-driven paraphrasing techniques by providing strategies that ensure that meaning and coherence are preserved in reworded material.