Topic Analysis of Japanese Sentences Using Sentence Embeddings

Previously, we proposed a topic analysis method that clusters sentence embeddings generated by Japanese Sentence-BERT and assigns a topic-word list to each cluster according to the values of a topic-word evaluation function. For a word in each cluster, the function is defined as a linear combination of the word’s cluster-based TFIDF and the cosine similarity between the word embedding and cluster centroid with hyperparameter α. In this study, we propose a method to automatically determine an appropriate value for α and evaluate the topic-word list for each cluster given by the determined α using topic coherence and topic diversity. Furthermore, we define a new metric called cluster-recall to evaluate whether a topic-word list adequately expresses the content of the corresponding cluster and evaluate the topic-word list attached to each cluster using cluster-recall. Automatically attached topic-word lists make it easy to understand the topic analysis results of a large-scale sentence dataset.