Topic mining of scientific literature can accurately capture the contextual structure of a topic, track research hotspots within a field, and improve the availability of information about the literature. This paper introduces a multi-dimensional topic mining method based on a hierarchical semantic graph model. The main innovations include (1) the hierarchical extraction of feature terms and construction of a corresponding semantic graph and (2) multi-dimensional topic mining based on graph segmentation and structure analysis. The process of semantic graph construction is based primarily on hierarchical feature term extraction, which can effectively reveal the hierarchical structural distribution of feature terms within documents. Our graph model also takes into account the complementarity of content-and context-related feature terms in documents while avoiding the loss of textual information. In addition, the multi-dimensional features of the topic can be mined effectively via an in-depth analysis of the constructed graph, resulting in a quantitative visualization of the many-to-many association between the topic and feature terms. A variety of experiments on existing document datasets demonstrate that the proposed approach is able to outperform state-of-the-art methods in terms of accuracy and efficacy. INDEX TERMS Topic mining, multi-dimensional topic, hierarchical semantic graph.

Traditional plagiarism detection is based primarily on methods of character matching or topic similarity. Another promising methodology remains largely unexplored: employing deep mining to establish a contextual hierarchy among themes. This paper proposes a semantic approach to measuring the extent of plagiarism, based on a hierarchical graph model. The main innovations are as follows: (1) hierarchical extraction of topic feature terms and elucidation of a corresponding graph structure; (2) graph similarity calculation based on the maximum common subgraph. This semantic-measure method goes beyond semantic detection of topics to take into account the context of topic feature terms, as well as the hierarchical structure by which those topics are related. This contextual-hierarchical perspective should, in turn, improve the accuracy of plagiarism detection. In addition, by mining the implicit relationships between hierarchical feature terms, our method can detect plagiarized documents with similar themes but using different topic words: a potential boon to plagiarism detection recall. In an experiment conducted on a dataset from Chinese paper database CNKI, the semantic-measure method indeed demonstrates accuracy and recall superior to those achieved with current state-of-the-art methods.