Non-task-oriented dialogue systems are required to chat with users in accordance with their interests. In this study, we propose a neural network-based method for estimating speakers’ levels of interest from dialogues. Our model first converts given utterances into utterance vectors using a word sequence encoder with word attention. Afterward, our novel attention approach, sentence-specific sentence attention extracts useful information for estimating the level of interest. Additionally, we introduce a new pre-training method for our model. Experimental results indicated that it was most effective to use topic-specific sentence attention and proposed pre-training in combination.