Fixed problems with non-projective trees

spacy/tests/test_visualization.py

@@ -1,4 +1,5 @@
 import pytest
+import deplacy
 from spacy.visualization import Visualizer
 from spacy.tokens import Span, Doc
 
@@ -113,6 +114,8 @@ def test_dependency_tree_non_projective(en_vocab):
         deps=["dep"] * 9,
     )
     dep_tree = Visualizer.render_dependency_tree(doc[0 : len(doc)], True)
+    for line in dep_tree:
+        print(line)
     assert dep_tree == [
         "<╗    ",
         "═╩═══╗",
@@ -159,3 +162,45 @@ def test_dependency_tree_input_not_span(en_vocab):
 
     with pytest.raises(AssertionError):
         Visualizer.render_dependency_tree(doc[1:3], True)
+
+def test_dependency_tree_highly_nonprojective(en_vocab):
+    """Test a highly non-projective tree (colloquial Polish)."""
+    doc = Doc(
+        en_vocab,
+        words=[
+            "Owczarki",
+            "przecież",
+            "niemieckie",
+            "zawsze",
+            "wierne",
+            "są",
+            "bardzo",
+            ".",
+        ],
+        heads=[5, 5, 0, 5, 5, None, 4, 5],
+        deps=["dep"] * 8,
+    )
+    dep_tree = Visualizer.render_dependency_tree(doc[0 : len(doc)], True)
+    assert dep_tree == [
+        "═╗<╗",
+        " ║<╣",
+        "<╝ ║",
+        "<══╣",
+        "═╗<╣",
+        "═══╣",
+        "<╝ ║",
+        "<══╝",
+    ]
+    dep_tree = Visualizer.render_dependency_tree(doc[0 : len(doc)], False)
+    assert dep_tree == [
+        "╔>╔═",
+        "╠>║ ",
+        "║ ╚>",
+        "╠══>",
+        "╠>╔═",
+        "╠═══",
+        "║ ╚>",
+        "╚══>",
+    ]
+
+

spacy/visualization.py

@@ -37,6 +37,11 @@ ROOT_LEFT_CHARS = {
 }
 
 
+class TableColumn:
+    def __init__(self, entity: str, width: int, overflow_strategy: str = "truncate"):
+        pass
+
+
 class Visualizer:
     @staticmethod
     def render_dependency_tree(sent: Span, root_right: bool) -> list[str]:
@@ -68,16 +73,38 @@ class Visualizer:
         # start with the root column
         indices_in_current_column = [i for i, h in enumerate(heads) if h is None]
         while len(indices_in_current_column) > 0:
+            assert (
+                len(
+                    [
+                        i
+                        for i in indices_in_current_column
+                        if i in all_indices_ordered_by_column
+                    ]
+                )
+                == 0
+            )
             all_indices_ordered_by_column = (
                 indices_in_current_column + all_indices_ordered_by_column
             )
-            indices_in_current_column = [
-                i
-                for index_in_current_column in indices_in_current_column
-                for i in children_lists[index_in_current_column]
-                if i not in all_indices_ordered_by_column
-            ]
-        # -1: root token with no arrow; 0: length not yet set
+            indices_in_next_column = []
+            # The calculation order of the horizontal lengths of the children
+            # on either given side of a head must ensure that children
+            # closer to the head are processed first.
+            for index_in_current_column in indices_in_current_column:
+                following_children_indices = [
+                    i
+                    for i in children_lists[index_in_current_column]
+                    if i > index_in_current_column
+                ]
+                indices_in_next_column.extend(following_children_indices)
+                preceding_children_indices = [
+                    i
+                    for i in children_lists[index_in_current_column]
+                    if i < index_in_current_column
+                ]
+                preceding_children_indices.reverse()
+                indices_in_next_column.extend(preceding_children_indices)
+            indices_in_current_column = indices_in_next_column
         horizontal_line_lengths = [
             -1 if heads[i] is None else 1
             # length == 1: governed by direct neighbour and has no children itself