A paraphrase model of a question text inputted by a user is learned, and a paraphrase expression is generated in real time. When information in text set storage unit is updated, text pair extracting unit extracts a paraphrase text pair from the text set storage unit and stores it in text pair storage unit. Model learning unit learns a question text paraphrase model from the paraphrase text pair in text pair storage unit, and stores it in model storage unit. Text pair extracting unit extracts a paraphrase text pair again from the text set storage unit by using the question text paraphrase model which the model storage unit possesses, and stores it in the text pair storage unit. In case where the stored paraphrase text pair is the same as the paraphrase text pair stored in the text pair storage unit, learning of the question text paraphrase model is ended. Candidate creating unit reads the question text paraphrase model from the model storage unit and generates a paraphrase candidate of the inputted question text.