in theoretical computer science, "semantic" is meant to indicate a property of a language.
in particular, it describes the set of words/symbols that are disposed in a particular order to describe a well-formed program.
always in theoretical computer science, it is often introduced alongside the study of automata, in order to demonstrate the differences and limitations of various theoretical models.
in the paper you linked it is mentioned how the logic that we can run on protocols like Bitcoin could be expressed using a simple state machine, the simplest of automatas, which are easier to describe and write but less powerful than models such as a PDAs and Turing machines
so when it is said that the UTxO tends to produce a simpler semantic model from a purely practical point of view it can be seen as a way to express the complexity needed to describe a program.
A simpler semantic model doesn't necessarily mean that is less powerful, only that is easier to express. (whereas a less powerful semantic model is condition sufficient but not necessary for a simpler semantic model)