In this paper, the authors describe an extension to an approach previously
discussed for personalization of a natural language system in the automotive
domain that allows reasoning under uncertainty with incomplete preference
structures. Therefore, the concept of an “information stream”
is defined as an underlying model for real-time recommendation learned
from previous speech queries. The stream captures contextual data based
on implicit feedback from the user’s speech utterances.
Furthermore, a formative
user study is discussed. Each study iteration has been based on a prototype
that allows the user to utter natural language queries in the restaurant
domain. The system responds with a ranked list of restaurant recommendations
in relation to the user’s context. Several driving scenarios
with varying contexts have been analyzed (e.g. weekday/ weekend, route
destinations, traffic). Users could inspect the result lists and indicate
the most preferred item. In addition to quantitative data gained from
this interaction, feedback on relevance of context features and on
the UI concept was collected in a post-study interview for each iteration.
Based on the study findings, we outline the contextual features found
to be most relevant for speech-based interaction in automotive applications.
These findings will be integrated into an existing hybrid recommendation
model.