Summary: Recently, computerized dialogue agents are studied actively. However, most of the studies of the dialogue agents are related to the task-oriented dialogue agents. That is, there are few studies of non-task-oriented dialogue agents. The analysis of dialogues between humans is necessary to make a practical system that can present enthusiastic dialogues. In this paper, we propose an automatic method that evaluates "enthusiastic" utterances in a dialogue using Conditional Random Fields (CRF). We use word co-occurence that extracted from large scale corpora as a feature to measure strength between utterances. As a result, our method indicates as much performance as human determinations.