use actual event names for featurization instead of constants

rasa/architecture_prototype/config_to_graph.py

@@ -289,7 +289,7 @@ def _core_config_to_train_graph_schema(
             "uses": StoryToTrainingDataConverter,
             "fn": "convert_for_training",
             "config": {},
-            "needs": {"story_graph": "load_stories"},
+            "needs": {"story_graph": "load_stories", "domain": "load_domain"},
             "persistor": False,
         }
         nlu_train_graph_schema, nlu_outs = _nlu_config_to_train_graph_schema(
@@ -353,14 +353,14 @@ def _core_config_to_predict_graph_schema(
                 "uses": StoryToTrainingDataConverter,
                 "fn": "convert_for_inference",
                 "config": {},
-                "needs": {"tracker": "add_parsed_nlu_message"},
+                "needs": {"tracker": "add_parsed_nlu_message",},
                 "persistor": False,
             },
             "create_e2e_lookup": {
                 "uses": MessageToE2EFeatureConverter,
                 "fn": "convert",
                 "config": {},
-                "needs": {"messages": nlu_e2e_out,},
+                "needs": {"messages": nlu_e2e_out},
                 "persistor": False,
             },
             **nlu_e2e_predict_graph_schema,

rasa/architecture_prototype/graph_components.py

@@ -7,6 +7,7 @@
 from rasa.architecture_prototype.interfaces import ComponentPersistorInterface
 from rasa.core.channels import CollectingOutputChannel, UserMessage
 from rasa.shared.constants import DEFAULT_DATA_PATH, DEFAULT_DOMAIN_PATH
+from rasa.shared.core.constants import DEFAULT_ACTION_NAMES
 from rasa.shared.core.domain import Domain
 from rasa.shared.core.events import ActionExecuted, UserUttered, Event
 from rasa.shared.core.generator import TrackerWithCachedStates
@@ -106,10 +107,28 @@ def generate(
 
 
 class StoryToTrainingDataConverter(GraphComponent):
-    def convert_for_training(self, story_graph: StoryGraph) -> TrainingData:
-        messages = []
+    def convert_for_training(
+        self, story_graph: StoryGraph, domain: Domain
+    ) -> TrainingData:
+        user_actions = [
+            ActionExecuted(user_action) for user_action in domain.user_actions
+        ]
+        end_to_end_actions = [
+            ActionExecuted(action_text=action_text)
+            for action_text in domain.action_texts
+        ]
+        default_actions = [
+            ActionExecuted(default_action) for default_action in DEFAULT_ACTION_NAMES
+        ]
+        user_events = []
         for step in story_graph.story_steps:
-            messages += self._convert_tracker_to_messages(step.events)
+            user_events += [
+                event for event in step.events if isinstance(event, UserUttered)
+            ]
+
+        messages = self._convert_tracker_to_messages(
+            user_actions + end_to_end_actions + default_actions + user_events
+        )
 
         # Workaround: add at least one end to end message to initialize
         # the `CountVectorizer` for e2e. Alternatives: Store information or simply config
@@ -148,6 +167,7 @@ def convert_for_inference(self, tracker: DialogueStateTracker) -> List[Message]:
 
 class MessageToE2EFeatureConverter(GraphComponent):
     """Collects featurised messages for use by an e2e policy."""
+
     def convert(
         self, messages: Union[TrainingData, List[Message]]
     ) -> Dict[Text, Message]:
@@ -156,8 +176,8 @@ def convert(
         additional_features = {}
         for message in messages:
             key = next(
-                k
-                for k in message.data.keys()
+                v
+                for k, v in message.data.items()
                 if k in {ACTION_NAME, ACTION_TEXT, INTENT, TEXT}
             )
             additional_features[key] = message

rasa/core/featurizers/single_state_featurizer.py

@@ -89,9 +89,7 @@ def _create_entity_tag_specs(
         ]
 
     def prepare_for_training(
-        self,
-        domain: Domain,
-        bilou_tagging: bool = False,
+        self, domain: Domain, bilou_tagging: bool = False,
     ) -> None:
         """Gets necessary information for featurization from domain.
 
@@ -221,10 +219,12 @@ def _extract_state_features(
         e2e_features: Optional[Dict[Text, Message]] = None,
     ) -> Dict[Text, List["Features"]]:
         key = next(
-            k for k in sub_state.keys() if k in {ACTION_NAME, ACTION_TEXT, INTENT, TEXT}
+            v
+            for k, v in sub_state.items()
+            if k in {ACTION_NAME, ACTION_TEXT, INTENT, TEXT}
         )
+        # TODO: We need a fallback for unexpected user texts during prediction time
         parsed_message = e2e_features[key]
-        assert parsed_message
 
         # remove entities from possible attributes
         attributes = set(
@@ -247,9 +247,7 @@ def _extract_state_features(
         return output
 
     def encode_state(
-        self,
-        state: State,
-        e2e_features: Optional[Dict[Text, Message]] = None,
+        self, state: State, e2e_features: Optional[Dict[Text, Message]] = None,
     ) -> Dict[Text, List["Features"]]:
         """Encode the given state.
 
@@ -289,9 +287,7 @@ def encode_state(
         return state_features
 
     def encode_entities(
-        self,
-        entity_data: Dict[Text, Any],
-        bilou_tagging: bool = False,
+        self, entity_data: Dict[Text, Any], bilou_tagging: bool = False,
     ) -> Dict[Text, List["Features"]]:
         """Encode the given entity data.
 
@@ -335,9 +331,7 @@ def encode_entities(
         }
 
     def _encode_action(
-        self,
-        action: Text,
-        e2e_features: Optional[Dict[Text, Message]] = None,
+        self, action: Text, e2e_features: Optional[Dict[Text, Message]] = None,
     ) -> Dict[Text, List["Features"]]:
         if action in self.action_texts:
             action_as_sub_state = {ACTION_TEXT: action}
@@ -349,9 +343,7 @@ def _encode_action(
         )
 
     def encode_all_actions(
-        self,
-        domain: Domain,
-        e2e_features: Optional[Dict[Text, Message]] = None,
+        self, domain: Domain, e2e_features: Optional[Dict[Text, Message]] = None,
     ) -> List[Dict[Text, List["Features"]]]:
         """Encode all action from the domain.
 
@@ -381,9 +373,7 @@ def __init__(self) -> None:
         )
 
     def _extract_state_features(
-        self,
-        sub_state: SubState,
-        sparse: bool = False,
+        self, sub_state: SubState, sparse: bool = False,
     ) -> Dict[Text, List["Features"]]:
         # create a special method that doesn't use passed interpreter
         name_attribute = self._get_name_attribute(set(sub_state.keys()))

tests/architecture_prototype/graph_schema.py

@@ -82,7 +82,7 @@
         "uses": StoryToTrainingDataConverter,
         "fn": "convert_for_training",
         "config": {},
-        "needs": {"story_graph": "load_stories"},
+        "needs": {"story_graph": "load_stories", "domain": "load_domain"},
         "persistor": False,
     },
     "process_core_WhitespaceTokenizer_0": {