Coordinating spatial perspective in discourse

Understanding and generating spatial descriptions such as “to the left of” and “above” is crucial for any situated conversational agent such as a robot used in rescue operations. The semantics of spatial descriptions are complex and involve (i) perceptual knowledge obtained from scene geometry (Regier and Carlson, 2001), (ii) world knowledge about the objects involved (Coventry et al., 2001), and (iii) shared knowledge that is established as the common ground in discourse. Dialogue partners coordinate all three types of meaning when describing and interpreting visual scenes. One example of (iii) is the assignment of perspective or the reference frame (RF). For example, the table may be “to the left of the chair”, “to the right of the chair”, “behind the chair” or “South of the chair”. The RF, may be described linguistically “from your view” or “from there” but in a spontaneous conversation it is frequently omitted. Instead, it is integrated in the content of the conversation as a discourse variable which is applied over several turns and even over several speakers. The RF may also be inferred from the perceptual context if given some configuration of the scene a spatial description is true only in that RF. It follows that when interpreting and generating spatial descriptions humans rely on verification of spatial templates in different RFs which requires considerable computational complexity (Steels and Loetzsch, 2009). The perspective is grounded by some point in the scene called the viewpoint (VPT). There are three ways in which the VPT is set in human languages (Levinson, 2003): (i) relative RF: by some third object distinct from the located and reference objects (the speaker, the hearer, the sofa); (ii) intrinsic RF: by the reference object itself (the chair); or (iii) extrinsic RF: by some global reference point (the North). Sometimes (mostly for route descriptions) a distinction is made between speaker-oriented (egocentric) and external (allocentric) perspective or between route and survey perspective but this is a less specific distinction. The geometric spatial templates are projected within the framework defined by the VPT (Maillat, 2003).