In this paper, we describe our intuitions about how language tech- nologies can contribute to create new ways to enhance the accessi- bility of exhibits in cultural contexts by exploiting the knowledge about the history of our senses and the link between perception and language. We evaluate the performance of fve multi-class classifcation models for the task of sensory recognition and introduce the DEEP Sensorium (Deep Engaging Experiences and Practices - Sensorium), a multidimensional dataset that combines cognitive and afective features to inform systematic methodologies for augmenting ex- hibits with multi-sensory stimuli. For each model, using diferent feature sets, we show that the features expressing the afective dimension of words combined with sub-lexical features perform better than uni-dimensional training sets.