Abstract: Abstract: To address the problems of highly realistic AI-generated text caused by the rapid development of large language models and the performance degradation of traditional detection methods, an intelligent text detection approach based on multi-feature adaptive fusion was proposed. Firstly, a language style feature set covering text statistical features, language structural features, and language uncertainty features was constructed to capture the differences between real texts and generated texts; meanwhile, deep semantic features of texts were extracted using independent encoding technology. Based on these, a dual-path mapping feature adaptive fusion strategy was designed: language style features and deep semantic features were first fused at a primary level, then further integrated through deep learning to enhance the adaptability and representational power of the fused features. Experimental results demonstrate that the proposed method achieves detection accuracies of 98.1% on the Chinese experimental dataset and 98.5% on the English TuringBench dataset. Compared with the best-performing baseline method J-Guard, improvements of 2.1 and 2.4 percentage points are observed on the Chinese and English datasets, verifying the effectiveness of the proposed approach. Focused on the issue that the rapid development of Large Language Model results in highly realistic AI-generated text and performance degradation of traditional detection methods, a detection method for AI-generated texts based on multi-feature adaptive fusion is proposed. This method first constructs a linguistic style feature set integrating textual statistical features, linguistic structural features, and linguistic uncertainty features to capture the discrepancies between authentic and AI-generated texts. Simultaneously, an independent encoding technique is employed to extract deep semantic features of the text. Subsequently, a dual-path mapping feature adaptive fusion strategy is designed: initially fusing the linguistic style features with deep semantic features, followed by secondary fusion based on deep learning mechanisms to enhance adaptive feature fusion capability. Experimental results demonstrate that the proposed method achieves detection accuracies of 98.1% and 98.5% on a self-constructed Chinese dataset and the TuringBench English dataset, respectively, representing average improvements of 6.67% and 7.29% compared with baseline models.

Key words: Keywords: AI-generated text, feature fusion, generated text detection, text classification, artificial intelligence

摘要： 摘 要: 针对大语言模型快速发展致使智能生成文本信息高度拟真、传统检测方法性能下降的问题，提出一种基于多特征自适应融合的智能生成文本检测方法。该方法首先构建涵盖文本统计特征、语言结构性特征及语言不确定性特征的语言风格特征集，以捕捉真实文本与生成文本的差异；并利用独立编码技术提取文本的深层语义特征。在此基础之上，设计一种双路映射特征自适应融合策略，先将语言风格特征与深层文本语义特征初步融合，再基于深度学习方法进行二次融合，增强特征自适应融合能力。实验结果表明：在中文实验数据集与英文TuringBench数据集上的检测准确率分别达到98.1%和98.5%。与基线方法中性能表现最好的J-Guard相比，本文方法在中英文实验数据集上的准确率分别提升了2.1与2.4个百分点，验证了所提方法的有效性。

关键词: 关键词: 智能生成文本, 特征融合, 生成文本检测, 文本分类, 人工智能