Modeling Spatial Knowledge from Verbal Descriptions Lamia Belouaer, David Brosset, Christophe Claramunt Naval Academy Research Institute, BP 600, 29240, Brest Naval, France {lamia.belouaer, david.brosset, christophe.claramunt}@ecole-navale.fr Abstract. Over the past few years, several alternative approaches have been suggested to represent the spatial knowledge that emerges from nat- ural environments. This paper introduces a rule-based approach whose objective is to generate a spatial semantic network derived from several humans reporting a navigation process in a natural environment. Verbal descriptions are decomposed and characterized by a graph-based model where actions and landmarks are the main abstractions. A set of rules implemented as first order predicate calculus are identified and applied, and allow to merge the common knowledge inferred from route descrip- tions. A spatial semantic network is derived and provides a global and semantic view of the environment. The whole approach is illustrated by a case study and some preliminary experimental results. Keywords: navigation knowledge, verbal description, spatial semantic network. 1 Introduction Spatial cognition is one of the most fundamental experiences of humans inter- acting in their environment [6]. All of us are in our every day life navigating in known or unknown environments, and trying to infer new spatial knowledge that will help us to act and behave appropriately. Amongst many spatial concepts, maps have been long used as valuable references for structuring a representation of our environment. Maps can be modeled in different forms: either qualitative or quantitative, metric or topological. The representations that emerge from maps are often sufficient for many applications, but in some cases they might not be appropriate. This is especially the case for human navigating with a low knowledge of their environment, particularly in natural and poorly structured environments. The spatial knowledge that can be inferred from human percep- tion of such environments cannot be directly described using conventional map representations. In fact, navigation in these environments should be structured enough to infer new spatial knowledge that will allow humans to navigate and reach their destinations [9]. Moreover, when perceived and described appropri- ately, navigation can generate a useful representation of the environment. The spatial knowledge that emerges might be even analyzed in order to infer a gen- eralized representation of the environment. Of course, humans perceive different