A spoken dialogue system was developed for the literature retrieval, and discussions were conducted mainly from the viewpoints of speech reply generation. The system was designed taking the following three points into consideration: appropriate amount of information included in a sentence, appropriate guidance from the system to users, and inappropriateness of using text-to-speech conversion devices for reply speech generation. In order to generate wide variety of output sentences, they were converted from concept representations. The generated sentences were given as sequences of phone and prosodic labels so that they can be directly fed to the speech synthesizer. System evaluation was conducted to compare the system-oriented and user-oriented dialogue modes, and also to find out appropriate ellipsis levels of output sentences. The results indicated the version at the middle of two extremes was the best for the both cases, though they will be different depending on various factors, such as whether the users having good experiences on dialogue systems, whether the system having a display, and so on.