This is very unusual according to my experience, and I agree that it's difficult to interpret.
There is a single value for either precision or recall for a particular label, but since these tables are presented as confusion matrices the values cannot be precision/recall.
I notice that the matrices show percentages which sum to 100 across each row for the "recall" one and sum to 100 across each column for the "precision" one. Based on this observation my guess is that the values show:
- in the "recall" table, the percentage of instances predicted with class X (column) among true instances of class Y (row). Example: 12.23% of of instances where the true label is "carry" are labelled as "walk".
- in the "precision" table, the percentage of instances which are truly class X (row) among instances predicted as class Y (column). Example: 6.87% of the instances predicted as "walk" actually belong to class "carry".
In my opinion this kind of non-standard representation should be avoided unless there's a really good reason. In this case a regular confusion matrix would have been clearer.