Accurate Generated Text Detection Based on Deep Layer-wise Relevance Propagation

Text generation models (TGMs) excel in producing text that matches the style of human language reasonably well. Such TGMs (e.g., GAN-based text generation model) can be misused for fake news and fake product reviews that can look authentic and fool humans. To distinguish the generated texts by TGM from the real texts, the latest detection models leverage learning algorithms with a variety of features to improve the detection performance against generated texts. However, the features included in these models are usually not relevant to the deep semantic correlation and fail to achieve accurate detection of the generated texts. To handle this problem, we propose a deep interpretable model to achieve accurate detection of the generated texts. In detail, a feature extraction method based on Layer-wise Relevance Propagation is proposed to improve the feature mining performance for detection of generated texts. Experiments based on real-world datasets demonstrate that the proposed model achieves more accurate generated text detection and significantly outperforms state-of-the-art detection schemes.