1.These are the implemented codes for our IJCAI 2024 accpeted paper (https://arxiv.org/abs/2403.03506).
2.For details about this paper.Please refer to our paper:
**Detecting AI-Generated Sentences in Realistic Human-AI Collaborative Hybrid Texts: Challenges, Strategies, and Insights**
BibTex:
@article{zeng2024detecting,
title={Detecting AI-Generated Sentences in Realistic Human-AI Collaborative Hybrid Texts: Challenges, Strategies, and Insights},
author={Zijie Zeng and Shiqi Liu and Lele Sha and Zhuang Li and Kaixun Yang and Sannyuya Liu and Dragan Gašević and Guanliang Chen},
year={2024},
journal={Proceedings of the 33rd International Joint Conference on Artificial Intelligence}
}
Dir A-Segment_Detection_Models contains all the segment detection models which detect segments by locating the boundaries between sentences of different authorship, .e.g., between human-written sentences and AI-written sentence.
Dir B-Segment_Classification_Models contains all the segment classification models which can perform classification over inputs such as a single sentence or a segment of several sentences.
Although we follow the pipeline of segment detection followed by segment classification to address sentence-level AI-content detection from hybrid texts, the segment detection task and sentence (not segment) classification task can actually be performed simultaneously.
For example, for an essay containing five sentences ⟨s1,s2,s3,s4,s5⟩, we can predict a label for each sentence using the segment classification model (treating each sentence as a segment). Let's say the prediction is <human,human,human,human,AI-written>. At the same time, we can apply the segment detection model to detect possible segments from ⟨s1,s2,s3,s4,s5⟩. Let's say two segments have been detected: ⟨s1,s2⟩ and ⟨s3,s4,s5⟩.
In this case, if we choose to trust the prediction of the segment detection model, then we should change the prediction of the last sentence (s5) from 'AI-written' to 'human-written' because the segmentation model believes that s3,s4,s5 should share the same label (i.e., segment classification based on sentence prediction results and the segment detection results).