Communication Aspects: How Landmarks Enrich the Communication Between Human and Machine - Landmarks: GIScience for Intelligent Services - page 179

sequence depends on the chosen optimization criterion. Typically, G UARD aims
for compact route directions, i.e., chooses the minimal number of chunks to cover
the whole route. References to landmarks are integrated into these instructions
already in G UARD 's first step, employing the landmark handling approach by
Richter [ 56 ] discussedinChap. 5 . In fact, landmarks are an important element of
spatial chunking (see [ 43 ] for an in-depth discussion). They enable structuring route
directions into chunks, such as 'follow Parbury Lane until you reach the end' (from
the CORAL example above) or 'turn left at the church' (which may still be a long
way down the road).
The chunk sequence is used as input for the language generation module (via
the dialog management, but we can ignore that for now). The language generation
module is a full-fledged NLG system based on the pCRU framework [ 5 ] . This
framework—probabilistic context-free representational underspecification is its full
name—allows for resolving the issue that getting from a semantic representation,
such as the chunks produced by G UARD , to a specific linguistic expression is almost
always underspecified. There are many ways to express a change of direction to the
left, for example, 'turn left', 'turn to the left', 'make a left turn', 'left', or 'go left'.
With pCRU, these variations can be formalized in a context-free grammar , 2 and each
possible variation can be assigned a probability of being generated by the system.
The kiosk system specifically uses landmarks as (intermediate) destinations along
the route. The route is segmented using landmarks along the way. For example, the
system may produce the following instructions to reach a particular office: 'Turn
around and go straight until the first corridor on your left. Turn left and go straight
until door B3180 is at your left.' As you can see, these instructions are similar
in structure to the CORAL instructions listed in Table 6.1 , but have at least one
landmark reference in every instruction step.
Dethlefs et al. [ 20 ] also use pCRU in their route direction generation approach.
They create different kinds of direction—turn-by-turn directions for unfamiliar
environments and destination descriptions for familiar ones. In that, they combine
and adapt the approach to determining the salience of streets by Tomko et al. [ 61 ]
and the idea of landmark categories by Duckham et al. [ 22 ] to determine the
content of the route directions, and then use pCRU and aggregation mechanisms to
create natural language route directions with linguistic variation in the instructions.
Tab le 6.2 lists examples for turn-by-turn instructions and destination descriptions
generated by their system for a route from the Richmond South Post Office to the
Richmond Cricket Ground (Richmond is a suburb of Melbourne, Australia).
2 We will not explain context-free grammar any further in this topic other than saying that it is a
concept of formal language theory. A grammar essentially is a mapping from some nonterminal
symbols (e.g., the semantic representation of a left turn) to some (string of) terminal symbols
(e.g., the words 'left', or 'go left'); the grammar is context-free if a specific nonterminal symbol
V always maps to the same string of terminal symbols w , regardless of its surrounding symbols
(the context). For more details on formal languages refer to a textbook on theoretical computer
science or theoretical linguistics, for example, Hopcroft, J.E., Motwani, R., Ullman, J.D. (2006).
Introduction to Automata Theory, Languages, and Computation (3rd ed.). Addison-Wesley.