A Multi-level Semantic Fusion Approach for News Reprint Pattern Detection

News reprint analysis is gradually becoming a hot research topic. Most existing studies in news reprint analysis mainly focus on mining news reprint relations, there is little study exploring the detection of the patterns of news reprint yet. To fill this gap, we aim to identify news reprint patterns such as word variant, sentence conversion, and topic restatement, which can provide deep insights into the news propagation mechanism. The challenge of this question lies in how to dig domain information and obtain the deep semantic representation of news to discover the writing styles of various reprint patterns. This paper proposes a media domain information-driven multi-level semantic fusion (MDID-MLSF) approach. It simultaneously considers news media information and extracts word-sentence-paragraph-level news semantic information by the interactive matching mechanism. Specifically, Word Mover's Distance (WMD) algorithm and Multilayer Perceptron (MLP) are employed to obtain semantic information at the word level and sentence level. Then, the model utilizes attention-based hierarchical Bi-LSTM to obtain the paragraph level semantic information. Finally, news media information and different levels' semantic information are jointly modeled to detect the patterns of news reprint. We empirically evaluate the performance of the proposed model on a real-world dataset, the experimental results demonstrate the efficacy of the model.