Merge pull request #6549 from RasaHQ/slow-crf-entity-extractor-training

Fix slow training of CRFEntityExtractor when using Entity Roles and Groups

changelog/6549.bugfix.rst

@@ -0,0 +1 @@
+Fix slow training of ``CRFEntityExtractor`` when using Entity Roles and Groups.

rasa/nlu/extractors/crf_entity_extractor.py

@@ -406,14 +406,18 @@ def _create_features_for_token(
  # get the features to extract for the token we are currently looking at
  current_feature_idx = pointer_position + half_window_size
  features = configured_features[current_feature_idx]
+
+ prefix = prefixes[current_feature_idx]
+
  # we add the 'entity' feature to include the entity type as features
  # for the role and group CRFs
+ # (do not modify features, otherwise we will end up adding 'entity'
+ # over and over again, making training very slow)
+ additional_features = []
  if include_tag_features:
- features.append("entity")
-
- prefix = prefixes[current_feature_idx]
+ additional_features.append("entity")
 
- for feature in features:
+ for feature in features + additional_features:
  if feature == "pattern":
  # add all regexes extracted from the 'RegexFeaturizer' as a
  # feature: 'pattern_name' is the name of the pattern the user