diff --git a/docling/backend/msword_backend.py b/docling/backend/msword_backend.py
index c0324f6fb1..6d5d24c1c8 100644
--- a/docling/backend/msword_backend.py
+++ b/docling/backend/msword_backend.py
@@ -2490,13 +2490,15 @@ def _get_paragraph_position(self, paragraph_element):
             for attr_name in ["y", "top", "positionY", "y-position", "position"]:
                 value = elem.get(attr_name)
                 if value:
-                    try:
-                        # Remove any non-numeric characters (like 'pt', 'px', etc.)
-                        clean_value = re.sub(r"[^0-9.]", "", value)
-                        if clean_value:
-                            return float(clean_value)
-                    except (ValueError, TypeError):
-                        pass
+                    # Read the leading number and ignore a unit suffix like
+                    # 'pt' or 'px'. The sign is part of it: a drawing anchored
+                    # above or left of its reference point carries a legal
+                    # negative offset, and dropping the minus turns it into a
+                    # position further down. A value with no number at all
+                    # falls through to the next candidate.
+                    match = re.search(r"-?\d*\.?\d+", value)
+                    if match:
+                        return float(match.group())
 
             # Check for position in transform attribute
             transform = elem.get("transform")
diff --git a/tests/test_backend_msword.py b/tests/test_backend_msword.py
index 314fa1c6f3..69a726cec6 100644
--- a/tests/test_backend_msword.py
+++ b/tests/test_backend_msword.py
@@ -2472,6 +2472,76 @@ def test_footnotes_and_endnotes_are_not_dropped():
     )
 
 
+def _position_of(value: str, attr: str = "y") -> float | int | None:
+    """Read one position attribute back through _get_paragraph_position."""
+    from lxml import etree
+
+    from docling.backend.msword_backend import MsWordDocumentBackend
+
+    backend = MsWordDocumentBackend.__new__(MsWordDocumentBackend)
+    elem = etree.fromstring(
+        f'<p xmlns:w="http://schemas.openxmlformats.org/wordprocessingml/2006/main"'
+        f' {attr}="{value}"/>'
+    )
+
+    return MsWordDocumentBackend._get_paragraph_position(backend, elem)
+
+
+@pytest.mark.parametrize("attr", ["y", "top", "positionY", "y-position"])
+def test_get_paragraph_position_keeps_a_negative_sign(attr: str):
+    """A drawing anchored above its reference point has a legal negative offset.
+
+    The position feeds the top-to-bottom sort of floating textbox paragraphs, so
+    dropping the sign moves that drawing to the wrong end of the document.
+    """
+    assert _position_of("-45.5", attr=attr) == -45.5
+
+
+def test_get_paragraph_position_still_strips_units():
+    assert _position_of("45.5pt") == 45.5
+
+
+@pytest.mark.parametrize(
+    ("value", "expected"), [("2e-5", 2.0), ("-2e-5", -2.0), ("1.5e-3in", 1.5)]
+)
+def test_get_paragraph_position_reads_exponent_notation_as_its_mantissa(
+    value: str, expected: float
+):
+    """A value in exponent notation is read, not dropped and not inflated.
+
+    Deleting every character outside the digits, the period and the minus leaves
+    "2-5" for "2e-5", which float() rejects, so the offset was swallowed by the
+    except and the drawing fell back to its source line. Deleting the minus too,
+    as the code did before this branch, leaves "25": an offset a million times
+    too large. Matching the number reads what is written, sign included. Word
+    writes these lengths with a unit suffix rather than an exponent, so this is
+    about the parser staying honest on input it did not expect.
+    """
+    assert _position_of(value) == expected
+
+
+@pytest.mark.parametrize(("value", "expected"), [(".5", 0.5), ("-.5", -0.5)])
+def test_get_paragraph_position_keeps_a_decimal_without_its_leading_zero(
+    value: str, expected: float
+):
+    r"""A decimal written as ".5" keeps its magnitude and its sign.
+
+    This one guards the pattern rather than the fix: `-?\d+(?:\.\d+)?` requires a
+    digit before the period, so it would match the "5" in ".5" and return 5.0,
+    ten times too large and positive for "-.5".
+    """
+    assert _position_of(value) == expected
+
+
+def test_get_paragraph_position_falls_through_a_value_with_no_number():
+    """An attribute like y="auto" holds no offset, so the next hint decides.
+
+    re.search returns None there, and the source line is the documented last
+    resort of this method. A single parsed element is on line 1.
+    """
+    assert _position_of("auto") == 1
+
+
 @pytest.mark.parametrize(
     ("note_body", "expected"),
     [