Prior Knowledge Enhanced Topic Model For Research Topic Evolution Analysis

Extracting and analyzing topics from the scientific literature is crucial for identifying the research hotspots, under-standing the developmental trajectory, and further revealing and predicting the evolutionary trends of a research field. Existing topic models like LDA and ETM can effectively extract topics from large text corpora. However, they struggle with incor-porating the domain-specific prior knowledge that is inherent and can be easily obtained in the scientific literature (e.g., the keywords in the paper), causing weekly targeted but over generic topic extraction and representation. To address this problem, we propose a Prior Attention Topic Model (PATM), which leverages inherent keywords in the scientific literature as prior knowledge for more targeted topic extracting and representation. Specifically, we first construct a prior attention weight matrix to guide the document embedding to pay more attention on the keywords, and then, a multi-view Variational Autoencoder (VAE) is utilized to separately encode different literature types, combining their unique characteristics while sharing one latent space. At last, the topic-word probability martix learned from the VAE are regarded as the topic representations. The experimental results and case study demonstrate that the proposed PATM is effective in extracting topics in the scientific literature, including papers and patents, and by analyzing these topics, we clearly reveal the thematic evolution of the research about emerging information technologies empowering auditing.