Wet/dry capacity are confused especially often if there is no barig or ban2 to tell them apart. These should be easy to separate if we take the counted object into account.
How best to achieve this?
curate a list of expected systems for each counted object
use the synsets from semantic.py to determine which system is most likely
manually label some data and train a classifier
As first step, maybe best to get a confusion matrix (via manual evaluation) to focus on the most confused pairs.
Wet/dry capacity are confused especially often if there is no barig or ban2 to tell them apart. These should be easy to separate if we take the counted object into account.
How best to achieve this?
As first step, maybe best to get a confusion matrix (via manual evaluation) to focus on the most confused pairs.