update crosswoz data

README.md

@@ -13,13 +13,13 @@ CrossWOZ is the first large-scale Chinese Cross-Domain Wizard-of-Oz task-oriente
   - Run `python preprocess.py` in the current directory. Need `../../crosswoz/` as the original data.
 - **Main changes of the transformation:**
   - Add simple description for domains, slots, and intents.
-  - switch intent&domain of `General` dialog acts => domain == 'General' and intent in ['thank','bye','greet','welcome']
+  - Switch intent&domain of `General` dialog acts => domain == 'General' and intent in ['thank','bye','greet','welcome']
   - Binary dialog acts include: 1) domain == 'General'; 2) intent in ['NoOffer', 'Request', 'Select']; 3) slot in ['酒店设施']
   - Categorical dialog acts include: slot in ['酒店类型', '车型', '车牌']
   - Non-categorical dialogue acts: others. assert intent in ['Inform', 'Recommend'] and slot != 'none' and value != 'none'
   - Transform original user goal to list of `{domain: {'inform': {slot: [value, mentioned/not mentioned]}, 'request': {slot: [value, mentioned/not mentioned]}}}`, stored as `user_state` of user turns.
   - Transform `sys_state_init` (first API call of system turns) without `selectedResults` as belief state in user turns.
-  - Transform `sys_state` (last API call of system turns) to `db_query` with domain states that contain non-empty `selectedResults`. The `selectedResults` are saved as `db_results`. Both stored in system turns.
+  - Transform `sys_state` (last API call of system turns) to `db_query` with domain states that contain non-empty `selectedResults`. The `selectedResults` are saved as `db_results` (only contain entity name). Both stored in system turns.
 - **Annotations:**
   - user goal, user state, dialogue acts, state, db query, db results.
   - Multiple values in state are separated by spaces, meaning all constraints should be satisfied.
@@ -36,10 +36,10 @@ Chinese
 
 | split      |   dialogues |   utterances |   avg_utt |   avg_tokens |   avg_domains |   cat slot match(state) | cat slot match(goal)   |   cat slot match(dialogue act) |   non-cat slot span(dialogue act) |
 |------------|-------------|--------------|-----------|--------------|---------------|-------------------------|------------------------|--------------------------------|-----------------------------------|
-| train      |        5010 |        84660 |     16.9  |        20.55 |          3.02 |                   99.67 | -                      |                            100 |                             94.39 |
+| train      |        5012 |        84674 |     16.89 |        20.55 |          3.02 |                   99.67 | -                      |                            100 |                             94.39 |
 | validation |         500 |         8458 |     16.92 |        20.53 |          3.04 |                   99.62 | -                      |                            100 |                             94.36 |
 | test       |         500 |         8476 |     16.95 |        20.51 |          3.08 |                   99.61 | -                      |                            100 |                             94.85 |
-| all        |        6010 |       101594 |     16.9  |        20.54 |          3.03 |                   99.66 | -                      |                            100 |                             94.43 |
+| all        |        6012 |       101608 |     16.9  |        20.54 |          3.03 |                   99.66 | -                      |                            100 |                             94.43 |
 
 6 domains: ['景点', '餐馆', '酒店', '地铁', '出租', 'General']
 - **cat slot match**: how many values of categorical slots are in the possible values of ontology in percentage.

data.zip

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:ddc310bba9fada91f266bd455d50b064ec897f069b960c6ca63447cefb8355bb
-size 15918515
+oid sha256:128b43faafb54f423b91240ee12e6e0eb1966d6dcdd2c8b903e91e0e254f7376
+size 15920814

preprocess.py

@@ -432,9 +432,6 @@ def preprocess():
             split = 'validation'
     
         for ori_dialog_id, ori_dialog in data.items():
-            if ori_dialog_id in ['10550', '11724']:
-                # skip error dialog
-                continue
             dialogue_id = f'{dataset}-{split}-{len(dialogues_by_split[split])}'
 
             # get user goal and involved domains
@@ -456,15 +453,20 @@ def preprocess():
             }
 
             for turn_id, turn in enumerate(ori_dialog['messages']):
+                # skip error turns
                 if ori_dialog_id == '2660' and turn_id in [8,9]:
-                    # skip error turns
                     continue
                 elif ori_dialog_id == '7467' and turn_id in [14,15]:
-                    # skip error turns
                     continue
                 elif ori_dialog_id == '11726' and turn_id in [4,5]:
-                    # skip error turns
                     continue
+                elif ori_dialog_id == '10550' and turn_id == 6:
+                    dialogue['user_state_final'] = dialogue['turns'][-2]['user_state']
+                    break
+                elif ori_dialog_id == '11724' and turn_id == 8:
+                    dialogue['user_state_final'] = dialogue['turns'][-2]['user_state']
+                    break
+
                 speaker = 'user' if turn['role'] == 'usr' else 'system'
                 utt = turn['content']