In an age where AI can generate and manipulate text that is often indistinguishable from human writing, which characteristics can help us identify human and machine text more accurately? This is a central question at the heart of a new research project from PD Dr. Terry Ruas at the University of Göttingen, titled “Paraphrase Types: A New Paradigm for Paraphrase Generation and Detection“, which has received funding from the German Research Foundation (DFG) for 36 months.
While current large language models, such as ChatGPT and Gemini, can produce text that appears human-like, they often miss the underlying linguistic nuances that connect different phrasings of the same idea. A direct example of this limitation is the inability of such models to reliably explain why and how two text segments are similar, or the linguistic characteristics that distinguish similar texts. This project aims to give AI a deeper understanding of the text it generates and manipulates.
The key innovation of the project lies in the concept of “paraphrase types.” Instead of treating paraphrasing as a simple binary choice (either two sentences mean the same thing or they don’t) Dr. Ruas proposes shifting the paradigm in paraphrase generation and detection to a more granular one using specific linguistic changes. These can range from simple word swaps (lexicon) to altering sentence structure (syntax), to more complex changes such as switching from active to passive voice (diathesis). “This approach has great practical relevance because a better understanding of AI generation methods could, among other things, revolutionize plagiarism detection,” says Dr. Jan Philip Wahle, who wrote his dissertation on the topic and now works as a project leader in the research group.
The project team will collaborate with a team of researchers in Germany and abroad, including researchers from the University of Göttingen, GWDG, University of Wuppertal, LMU Munich, the National Institute of Informatics in Japan, Federal University of ABC and University Center FEI in Brazil, and the National Research Council in Canada.
Press release by Göttinger Tageblatt: https://www.goettinger-tageblatt.de/beruf-und-bildung/regional/wer-schrieb-den-text-wirklich-goettinger-projekt-sucht-versteckte-muster-von-ki-und-menschen-GBBROHMF7ZA5NPZ4CYMZ4LIUGQ.html
