Communication Aspects: How Landmarks Enrich the Communication Between Human and Machine - Landmarks: GIScience for Intelligent Services - page 172

and mistakes in the human-computer interaction [ 31 ] . Rather, the machine should
always pick what is (objectively) correct and best—a perfect super-human so to
speak.
In the remainder of this chapter, we will first address ways of how machines may
produce references to landmarks (in Sect. 6.2 ) because this appears to be easier than
understanding references to landmarks, which we will discuss in Sect. 6.3 . Finally,
in Sect. 6.4 , we will point to some studies that have investigated the usability and
usefulness of introducing landmarks into human-computer interaction.
For a machine, it is easier to successfully produce landmarks than to understand
them for computational and cognitive reasons. Computationally, a machine may
rely on the data structures and algorithms presented in Chaps. 4 and 5 —given that
sufficient data is available. As we have discussed, this is hardly ever uniformly the
case for any given environment (see also the final discussion in the next chapter).
Accordingly, some fallback strategies in communication are required in case no
landmark is available.
These computational methods ensure that the machine will pick landmark
references that are salient and relevant. The pick may not necessarily be actually
optimal though, since data may not be complete and any computational approach
out of necessity makes some simplifying assumptions. However, generally this will
not be an issue, since in contrast to machines people are very good at adapting
to their communication partner. Simply by mentioning a geographic object in the
communication, the communication partner will pay attention to this object, making
it more relevant and, thus, increasing its salience [ 57 ] . The object becomes a
landmark by virtue of being mentioned in the communication. This will compensate
for the machine potentially picking not the most suited landmark reference. People
will still be able to recognize and understand the chosen one in most cases. So, in
short, in producing landmark references, machines can rely on a well-defined, but
likely incomplete, set of landmark candidates to choose from—determined by the
underlying data sets and algorithms—and on the cognitive facilities of the human
communication partner, which enable them to flexibly adapt to the chosen landmark
references.
On the other hand, this human flexibility and ability to adapt is the reason why
understanding landmark references is so difficult for a computer. Even though we
can assume a limited context in which interaction happens, namely negotiating some
spatial descriptions (e.g., where something is located or how to get to some place),
the variety of geographic objects people may select from and the variety in the ways
they may describe this selection is immensely wide. This means that the computer
needs much greater flexibility compared to the case of producing landmarks, and
machines are not very good at adaptation. Thus, either the vocabulary needs to be
restricted for successful understanding of landmark references—which takes away
many advantages of landmark-based communication—or a sophisticated mecha-
nism for resolving (arbitrary) landmark references is required, including detailed
data about the environment and advanced parsing functionality. Alternatively,
machine learning mechanisms may be employed to learn from communication
with a human partner; however, to date we are not aware of any approach actually
following down that path.