A Method of Labeling to the Topic in a Chat Dialogue

Abstract

In this research, we propose a method of labeling a topic word for the sections in a chat dialogue. First, we investigated how humans label topic words. As a result, it was found that the words included in the first sentence in which the topic started and the words associated with the words appearing evenly in the section were given as a topic words. It was also found that non-characteristic words that appear in other sections were selected as the topic word when labeling is based on the frequency of appearance of words in the section. We first obtain the words with high TF-IDF in the first sentence of the section as the candidate for the topic word. The words with high DF value that appear regardless of the topic are set to exclude words as the stop word. We compared the label using the proposed method with labels given by humans. It was confirmed that the accuracy of labeling was improved.