The base suggested-element modifier specifies that the reduce node will be the bottom-most node in a singlet chain. That is, subsequent rules using the singlet element modifier will not look "down" past a node labeled with the base attribute.
Base and singlet provide a mechanism for layering multiple attributes and syntactic and semantic ambiguities directly into the parse tree.
The example will help clarify the use of the base modifier.
# The rule below reduces a node labeled _noun to an _np. It
places an internal flag called "base" on the _np node.
# Subsequent rules (e.g., in subsequent passes) that are searching for
_noun will not find it under an _np created with the current rule.
@RULES
_np [base] <- _noun @@
# Say a subsequent pass has this rule. Even
though the s (or singlet) modifier says to look "within" a node
for _noun, it won't find it within an _np
# labeled with a base flag, as created by the rule above.
@RULES
_np <- _adj _noun [s] @@