I think you misunderstood, in a lot of ways, what Haag & Nagel (2000; what you call Paper 1) did and how Arens, Gerber & Nagel (2008; Paper 2) extended it. Fig. 1 of AGN08 is a good summary of HN00. What HN00 did was build a system that could watch a video of an intersection, detect cars, and translate the car behavior into a conceptual framework. As inspiration for their system, they used their idea of how humans represent the task:
ﬁve levels of representation seem to be involved: (i) a representation of the geometry of spatiotemporal developments in the road traﬃc scene, comprising both a 2D-one in the image plane and a 3D-one relating to the depicted scene, (ii) a representation of driving maneuvers closely coupled to particular traﬃc situations, (iii) a conceptual representation of visible bodies, their attributes, and their elementary movements, (iv) generic conceptual representations of spatiotemporal body conﬁgurations and their expected temporal developments, and (v) one or more versions of a natural language representation of developments centred around the current point in time.
In other words, the goal of HN00 was to look at a 2D image of an intersection, from it construct a 2D/3D representation of the scene. In that scene identify and label objects and describe them in a conceptual language called SIT++. Once in that conceptual representation (as situation trees) they could conduct logical inference (using Fuzzy Metric
Temporal Horn Logic) on their representation in order to decide what the agents they identified are trying to do.
Note that HN00 involved no natural language processing (NLP) at all. Although they did have to use lots of pattern recognition and various machine learning algorithm that would be familiar to NLP practitioners. However, their domain was transforming a visual scene into a conceptual (not natural language) internal representation.
How did AGN08 extend beyond this? They changed what they wanted to do. Their task was not simply to view a scene and transform it into an internal representation, but to then output that internal representation in a natural language description. Thus, they were adding a natural language generation system to HN00. Generating natural language from an internal representation is obviously an important part of NLP.
In the process of adding this functionality, AGN08 had to extend the internal representation in several ways. This was due to the fact that more internal information was required to generate good natural language output, and because they wanted to deal with more complex scenes than HN00. The paper focuses on this aspect of the work (extending the internal representation) and only tangentially touches on natural language output. They go into detail of the natural language output in:
R. Gerber, Naturlichsprachliche Beschreibung von Straßenverkehrsszenen durch Bildfolgenauswertung. Dissertation,Fakultat fur Informatik der Universitat Karlsruhe (TH), Karlsruhe, January 2000
Unfortunately, I am not willing to learn German and read a whole thesis in order to give you a more complete answer about the details. Before you try to do that yourself (hopefully you already know German) or look into more recent papers, I recommend learning some basics of NLP. A good source is the following question:
Looking for a good beginners reference to learn computational linguistics