Masterkolloquium

Over the last few years major advances in the development of Artificial In-
telligence were recorded. Googles DeepMind[WIG18] manages to generate
convincing photos of burgers, dogs, and butterflies, while OpenAI[ope18]
and AlphaStar[tea19], another DeepMind project, managed to beat the worlds
best players of two highly complex games. According to the Institute for
Economic Policy the number of Artificial Intelligence patents filed has in-
creased drastically[WBSL18]. Artificial Intelligence applications are well
equipped to deal with big data, but there is one sector where this sort
solution is sparsely or not at all used: The space sector. Space missions
produce a lot of data and with advancements in technology the amount
of data is bound to increase even further. Flight controllers have a hard
time going through that data to find the single value they were looking for.
Slow reaction times to critical mission scenarios is a consequence of that.
Visionspace, a NewSpace company and experienced partner for satellite
and ground station monitoring proposed an AI solution. A chatbot was pro-
totyped with the ability to quickly process queries in the form of English
questions on space mission data. This prototype was developed with RASA,
an open source Framework for the creation of contextual AI assistants. The
next step is to assess the Natural Language Processing capabilities of that
prototype and improve them. The sector of space missions contains many
domain specific terms and each mission uses yet another set of mission ex-
plicit terms to describe its data models. The domain specific language and
the initial lack of distinct sample queries have proven to be the biggest
hurdle. A number of sample generation methods were investigated or dis-
cussed, but ultimately none of them have proven especially useful. The
domain specific terminology and limited abilities of the prototype made it
necessary to instruct potential sample generators to a degree where they
were biased by the instructor and would either produce samples very sim-
ilar to existing ones or lead to an over-saturation of edge cases. Another
aspect of space missions is discussed in this thesis. Space missions are crit-
ical by nature, there is often no chance of recovery if something goes bad.
That raises the question whether flight controllers will or even should trust
an AI assessment. Classification of natural language is of statistical nature
and is destined to fail sometimes. Blind trust in AI will eventually lead to
catastrophe while no trust in it makes it useless. We need the user to trust
the AI and catch its errors. Research has shown that trust assessment of
AI systems is an ability that can be attained by a user of any expertise if
the AI system explains its decision making[RSG16]. Removing part of the
black box from our AI solution and enhancing the user experience in gen-
eral gives us the ability to collect unbiased samples directly from the Flight
Controllers in missions. It would be bad if the mission loses trust in this
project due to the limited capabilities of the Natural Language Understand-
ing model. This is why the design of trust assessment features and quality
of life improvements in the Natural Language Processing are the main focus
of this thesis.