PoSLemma: How Traditional Machine Learning and Linguistics Preprocessing Aid in Machine Generated Text Detection

Resumen

With the release of several Large Language Models (LLMs) to the public, concerns have emerged regarding their ethical implications and potential misuse. This paper proposes an approach to address the need for technologies that can distinguish between text sequences generated by humans and those produced by LLMs. The proposed method leverages traditional Natural Language Processing (NLP) feature extraction techniques focusing on linguistic properties, and traditional Machine Learning (ML) methods like Logistic Regression and Support Vector Machines (SVMs). We also compare this approach with an ensemble of Long-Short Term Memory (LSTM) networks, each analyzing different paradigms of Part of Speech (PoS) taggings. Our traditional ML models achieved F1 scores of 0.80 and 0.72 in the respective analyzed tasks.