While many major scientific publishers utilise systems to detect verbatim text reuse, these tools are often architecturally limited, leaving the peer-review process vulnerable to sophisticated threats like paraphrased content and coordinated fraud among different reviewers.

This thesis aimed to address these limitations by developing a more advanced, scalable detection system capable of understanding semantics. This project focuses on a dual-method prototype that combines a lexical n-gram baseline with a semantic similarity model. The semantic model uses embeddings and a vector database to perform a global, cross-reviewer comparison of all reports in a historical corpus

By operating on the level of semantics instead of just lexical matching, this approach makes it possible to uncover sophisticated misconduct, including paraphrased reports from different reviewers. The prototype's performance was validated through a large-scale test and qualitative interviews with research integrity experts.