Text analysis using a neural topic model with word embedding

Topic models have been widely used in various fields to find topics of documents or words. It has been known that topic models have difficulties for model modification and extension, and also have poor performance for corpus with large vocabularies. To overcome these problems, neural topic models have been suggested. However, both models still have limitations in that it is difficult to find the meaning of topics. To resolve this problem, embedded topic model(ETM) has been proposed that combined a neural topic model and word embedding. ETM expresses words and topics through word embedding, thus can be visualized using dimension reduction. In this paper, we introduced ETM and evaluation methods, and apply them to 100,000 research paper summary data. For the analysis, various combination of hyper parameters are considered, and the selected model is compared with other topic model. As a result, it is shown that ETM has better performance. The meaning of topics are investigated by using top words of each topic and the plot of topic and word embeddings through dimension reduction, and topics seems to represent the fields of research papers.