我们选取一段典型的天文学论文原文作为案例:“We measure the stellar mass function of galaxies at z=2.3 using deep HST WFC3 imaging. The sample consists of 420 galaxies with stellar masses ranging from 10^9 to 10^11 M⊙. We find that the mass function is well described by a Schechter function with a characteristic mass log(M*/M⊙)=10.8 and a faint-end slope α=-1.4. These results are consistent with previous studies (e.g., Marchesini et al. 2009).” 该段落存在两个问题:一是重复率较高(“mass function”出现三次),二是句式过于模板化(“We measure... We find that...”),容易被AI检测器标记。
我们采用双降工作流进行处理。首先,将固定术语(如“Schechter function”、“log(M*/M⊙)=10.8”)与可改写部分分离。然后,对可改写部分进行困惑度优化。原始段落的平均PPL为45,属于低困惑度区间。我们通过以下修改提升PPL至95:将“We measure”改为“This study presents measurements of”;将“We find that”改为“The analysis reveals that”;将“These results are consistent with”改为“The derived parameters align with those reported by”。同时,为避免重复,将第二个“mass function”替换为“the stellar mass distribution”。最终修改后的段落为:“This study presents measurements of the stellar mass function of galaxies at z=2.3, utilizing deep HST WFC3 imaging. The sample comprises 420 galaxies with stellar masses spanning 10^9 to 10^11 M⊙. The analysis reveals that the stellar mass distribution is well described by a Schechter function, with a characteristic mass log(M*/M⊙)=10.8 and a faint-end slope α=-1.4. The derived parameters align with those reported by Marchesini et al. (2009).” 经测试,修改后AI检测率从38%降至7%,重复率从22%降至9%,且所有术语和数据均得到保留。