This paper describes annotation of face images in keywords based on latent semantic indexing, and experimental results in FIARS. Two latent semantic spaces are constructed from visual and symbolic features. These features are corresponding to lengths of some places of a face and keywords. One latent semantic space is constructed from visual features, the other space is constructed from both features. The former space is used for retrieving similar face images, and the latter for seeking keywords to a given face image. Moreover, the two types of visual futures are utilized. One is specified in terms of the lengths of face parts, and the other in terms of points on the outlines of a face and its parts. As an experiment, recall and precision ratios of assigned keywords are measured using the two types of the visual features.