@@ -95,6 +95,9 @@ As mentioned before, all evalution metrics confirmed our initial study conjectur
Since this last paragraph contains all the necessary information about the text, we consider only that one in our project. This, of course, makes the task more difficult to approach as it excludes the context prior to the last paragraph and thus disregards any reference occurring in the last paragraphs to prior arguments. Especially for texts of argumentative nature, this can negatively impact the system performance in a significant way, as prior context can be of relevance in some cases.
3. Underperformance for infrequent classes
As in all of Machine Learning tasks, system results will be negatively effected if instances for some output classes are barely existent in the test dataset. In our case, we realized quite early during annotation that this could be the case for the Maslow category "physiological needs" as we had very few instances of this class in our data. The results proved our hypothesis regarding this aspect, as the system performed poorly on the instances that were annotated for "physiological needs". Similarly, Reiss motives that rarely occurred in our test dataset performed roughly equally poor due to the same problem: Underfitting due to lack of sufficient data. To resolve this problem, one would have to ensure a more balanced dataset or even consider leaving out a few classes and restructuring the output classes at use.
