A couple optimizations - checking against None for deps saves on multiple steps in the non-enhanced graph case, and checking for any Space in misc optimizes the processing of Spaces in words which don't have Space (which is the most common case by far in most treebanks)

Author: AngledLuffa

stanza/models/common/doc.py

@@ -1250,6 +1250,15 @@ def consolidate_whitespace(self):
         """
         Remove whitespace misc annotations from the Words and mark the whitespace on the Tokens
         """
+        # most tokens have no whitespace annotation at all, and searching the
+        # misc strings for one is much cheaper than splitting each of them
+        # into pieces.  every prefix this method looks for - SpaceAfter,
+        # SpacesAfter, SpacesBefore - begins with "Space", and every test it
+        # makes on them is case sensitive, so this skips exactly the tokens
+        # which would have fallen through all of those tests anyway
+        if not (self._misc and "Space" in self._misc) and not any(word._misc and "Space" in word._misc for word in self.words):
+            return
+
         found_after = False
         found_before = False
         num_words = len(self.words)
@@ -1498,9 +1507,33 @@ def __init__(self, sentence, word_entry):
         if self._misc is not None:
             init_from_misc(self)
 
-        # use the setter, which will go up to the sentence and set the
-        # dependencies on that graph
-        self.deps = word_entry.get(DEPS, None)
+        # a Word being constructed cannot have any existing in-edges to clear,
+        # so we skip the general setter and go straight to adding the edges
+        deps = word_entry.get(DEPS, None)
+        if deps is not None:
+            self._init_deps(deps)
+
+    def _init_deps(self, value):
+        """Add the enhanced dependencies of a freshly built Word
+
+        The deps setter has to clear the Word's existing parents first, which
+        costs a has_node and an in_edges lookup per Word.  Neither can find
+        anything during construction, so this skips them
+        """
+        graph = self._sent._enhanced_dependencies
+        if graph is None:
+            graph = nx.MultiDiGraph()
+            self._sent._enhanced_dependencies = graph
+
+        if isinstance(value, str):
+            value = value.split("|")
+        if all(isinstance(x, str) for x in value):
+            value = [x.split(":", maxsplit=1) for x in value]
+        for parent, dep in value:
+            parent = tuple(map(int, parent.split(".", maxsplit=1)))
+            if len(parent) == 1:
+                parent = parent[0]
+            graph.add_edge(parent, self.id, dep)
 
     @property
     def manual_expansion(self):