Understanding Navigational Instructions:
Michael Levit, Deb Roy

About a decade ago, University of Edinburgh and its partners put together a collection of transcribed human/human dialogs in which both participants were given similar maps. One of the participants, the instruction giver, had to navigate the other particpant along a path that was drawn only on the giver's map. Since then, this corpus (commonly known as MapTask Corpus) has been used for various tasks such as sleep deprivation studies, "games and moves" classification experiments and others. To our knowledge, there has been no attempts to model actual understanding on the instruction receiver end. The long-term goal of this study is to fill this gap. To achieve this goal, we introduce Navigational Information Units (NIUs) that describe individual intervals of the paths and are drawn from several categories (such as "moves" and "positions"). Our first aim is to decompose these NIUs into a number of possibly independent constituents and ground the latter in terms of the objects on the maps. Having such models at hand, we can further employ the notion of context and try and replicate the entire paths as a sequence of extracted NIUs.