Improve natural language inference performance (#1234)

* improve natural language inference performance

* update release notes

* update comment

* add new test

* lint fix

Author: thehomebrewnerd

docs/source/release_notes.rst

@@ -8,13 +8,14 @@ Future Release
     * Enhancements
     * Fixes
     * Changes
+        * Update inference process to only check for NaturalLanguage if no other type matches are found first (:pr:`1234`)
     * Documentation Changes
-        * Updating contributing doc with Spark installat instructions (:pr:`1232`)
+        * Updating contributing doc with Spark installation instructions (:pr:`1232`)
     * Testing Changes
         * Enable auto-merge for minimum and latest dependency merge requests (:pr:`1228`, :pr:`1230`, :pr:`1233`)
 
     Thanks to the following people for contributing to this release:
-    :user:`gsheni`, :user:`willsmithorg`
+    :user:`gsheni`, :user:`thehomebrewnerd`, :user:`willsmithorg`
 
 v0.11.0 Dec 22, 2021
 ====================

woodwork/tests/type_system/test_ltype_inference.py

@@ -1,3 +1,5 @@
+from mock import patch
+
 import woodwork as ww
 from woodwork.accessor_utils import _is_koalas_series
 from woodwork.logical_types import (
@@ -15,6 +17,12 @@
     Unknown,
 )
 from woodwork.tests.testing_utils import to_pandas
+from woodwork.type_sys.type_system import (
+    DEFAULT_INFERENCE_FUNCTIONS,
+    DEFAULT_RELATIONSHIPS,
+    DEFAULT_TYPE,
+    TypeSystem,
+)
 from woodwork.utils import import_or_none
 
 UNSUPPORTED_KOALAS_DTYPES = [
@@ -125,6 +133,47 @@ def test_natural_language_inference(natural_language):
             assert isinstance(inferred_type, NaturalLanguage)
 
 
+@patch("woodwork.type_sys.inference_functions.natural_language_func")
+def test_nl_inference_called_on_no_other_matches(nl_mock, pandas_natural_language):
+    assert isinstance(
+        ww.type_system.infer_logical_type(pandas_natural_language[0]), NaturalLanguage
+    )
+    new_type_sys = TypeSystem(
+        inference_functions=DEFAULT_INFERENCE_FUNCTIONS,
+        relationships=DEFAULT_RELATIONSHIPS,
+        default_type=DEFAULT_TYPE,
+    )
+    new_type_sys.inference_functions[NaturalLanguage] = nl_mock
+    _ = new_type_sys.infer_logical_type(pandas_natural_language[0])
+    assert nl_mock.called
+
+
+@patch("woodwork.type_sys.inference_functions.natural_language_func")
+def test_nl_inference_called_with_unknown_type(nl_mock, pandas_strings):
+    assert isinstance(ww.type_system.infer_logical_type(pandas_strings[0]), Unknown)
+    new_type_sys = TypeSystem(
+        inference_functions=DEFAULT_INFERENCE_FUNCTIONS,
+        relationships=DEFAULT_RELATIONSHIPS,
+        default_type=DEFAULT_TYPE,
+    )
+    new_type_sys.inference_functions[NaturalLanguage] = nl_mock
+    _ = new_type_sys.infer_logical_type(pandas_strings[0])
+    assert nl_mock.called
+
+
+@patch("woodwork.type_sys.inference_functions.natural_language_func")
+def test_nl_inference_not_called_with_other_matches(nl_mock, pandas_integers):
+    assert isinstance(ww.type_system.infer_logical_type(pandas_integers[0]), Integer)
+    new_type_sys = TypeSystem(
+        inference_functions=DEFAULT_INFERENCE_FUNCTIONS,
+        relationships=DEFAULT_RELATIONSHIPS,
+        default_type=DEFAULT_TYPE,
+    )
+    new_type_sys.inference_functions[NaturalLanguage] = nl_mock
+    _ = new_type_sys.infer_logical_type(pandas_integers[0])
+    assert not nl_mock.called
+
+
 def test_categorical_inference_based_on_dtype(categories_dtype):
     """
     This test specifically targets the case in which a series can be inferred

woodwork/type_sys/type_system.py

@@ -315,11 +315,26 @@ def get_inference_matches(types_to_check, series, type_matches=[]):
                 get_inference_matches(check_next, series, type_matches)
             return type_matches
 
-        type_matches = get_inference_matches(self.root_types, series)
+        # Don't include NaturalLanguage as we only want to check that if
+        # no other matches are found
+        types_to_check = [
+            ltype for ltype in self.root_types if ltype != NaturalLanguage
+        ]
+        type_matches = get_inference_matches(types_to_check, series)
 
         if len(type_matches) == 0:
-            # If no matches, set type to default type (Unknown)
-            logical_type = self.default_type
+            # Check if this is NaturalLanguage, otherwise set
+            # type to default type (Unknown). Assume that a column
+            # can only be natural language if it is not already a
+            # match for another type. Also improves performance by
+            # limiting the times the natural language inference function
+            # is called.
+            if self.inference_functions.get(
+                NaturalLanguage
+            ) and self.inference_functions[NaturalLanguage](series):
+                logical_type = NaturalLanguage
+            else:
+                logical_type = self.default_type
         elif len(type_matches) == 1:
             # If we match only one type, return it
             logical_type = type_matches[0]