[addresses] Adding address-level component dropout to AddressComponents (returns an ordering so the client formatter can potentially emit multiple addresses with different components dropped out). Adding PO box and category probabilities to config

resources/parser/default.yaml

@@ -7,6 +7,83 @@ languages:
     # Replace user-tagged admin components with the non-local language version
     replace_non_local_probability: 0.4
 
+# Dependencies for including each component in an "address"
+# Two-way dependencies are not an issue
+component_dependencies:
+    house:
+        dependencies: []
+
+    road:
+        dependencies:
+            - house
+            - house_number
+            - suburb
+            - city_district
+            - city
+            - postcode
+
+    house_number:
+        dependencies:
+            - road
+
+    entrance:
+        dependencies:
+            - house_number
+
+    staircase:
+        dependencies:
+            - house_number
+
+    level:
+        dependencies:
+            - house_number
+
+    unit:
+        dependencies:
+            - house_number
+
+    postcode:
+        dependencies: []
+
+
+# Each component is dropped out separately and a new address
+# is added to the training set. These are only the address-level
+# components. Places/boundaries are taken care of elsewhere.
+dropout:
+    attention:
+        probability: 0.8
+    care_of:
+        probability: 0.8
+    house:
+        probability: 0.6
+    house_number:
+        probability: 0.5
+    road:
+        probability: 0.4
+    entrance:
+        probability: 0.8
+    staircase:
+        probability: 0.8
+    level:
+        probability: 0.6
+    unit:
+        probability: 0.5
+    postcode:
+        probability: 0.6
+
+po_box:
+    probability: 0.1 
+    # Note: these probabilities all independent (don't need to sum to 1)
+    drop_address_probability: 0.8 # drop house number, road, etc.
+    drop_places_probability: 0.1 # drop place names
+    drop_postcode_probability: 0.3 # drop postal code
+
+category:
+    # Same thing for category queries
+    drop_address_probability: 0.8 # drop house number, road, etc.
+    drop_places_probability: 0.1 # drop place names
+    drop_postcode_probability: 0.3 # drop postal code
+
 neighborhood:
     # Usually in Germany, may have e.g. name:prefix=Ortsteil
     add_prefix_probability: 0.5