As the number of social network services (SNSs) and their users grow, so does the complexity of individual networks as well as the amount of information to be consumed by the users. Unlike ordinary bloggers who tend to focus on posting rather long, unidirectional content, on the other hand, the SNS users exchange shorter and more instantaneous messages and contents more inter-actively to form conversations. In order to alleviate the information overload problem and help SNS users manage their social networks, we show that more refined networks of several kinds can be identified over the explicit, connectivity-based social network by analyzing the conversational content based on the topical diversity and topical purity of the conversations, which was revealed by an LDA-based algorithm applied to about 1.5 million conversations including about 5 million tweets posted by more than 2,000 Twitter users and about 0.3 million their conversational partners. The resulting “semantic” social networks help identifying more meaningful social relationships among the users interconnected by a “syntactic” social network and determining new properties such as different types of relationships based on topic diversity and topic purity or different types of users based on types of relationships they have. We expect that the proposed method can lead to more meaningful SNS analyses and hence to more practical values to SNS in general.