def _get_utter_from_dict(vocab: Dict[str, int], image_url_id: Dict[str,
int],
utter_dict: dict) -> Utterance:
"""Extract Utterance object from JSON dict.
Args:
vocab (Dict[str, int]): Vocabulary.
image_url_id (Dict[str, int]): Image URL to index.
utter_dict (dict): JSON dict.
Returns:
Utterance: Extracted Utterance.
"""
utter = utter_dict.get('utterance')
_speaker: str = utter_dict.get('speaker')
_utter_type: str = utter_dict.get('type')
_text: str = utter.get('nlg')
_pos_images: List[str] = utter.get('images')
_neg_images: List[str] = utter.get('false images')
# Some attributes may be empty.
if _text is None:
_text = ""
if _utter_type is None:
_utter_type = ""
if _pos_images is None:
_pos_images = []
if _neg_images is None:
_neg_images = []
# Convert speaker into an integer.
speaker: int = -1
if _speaker == 'user':
speaker = USER_SPEAKER
elif _speaker == 'system':
speaker = SYS_SPEAKER
assert speaker != -1
# Convert utterance type into an integer.
utter_type: int = DatasetConfig.utterance_type_dict.get(_utter_type, 0)
# We don't care the type of system response.
if speaker == SYS_SPEAKER:
utter_type = 0
# Convert text into a list of integers.
words: List[str] = word_tokenize(_text)
text: List[int] = [vocab.get(word.lower(), UNK_ID) for word in words]
# Images
pos_images: List[int] = [
image_url_id.get(img, 0) for img in _pos_images
]
neg_images: List[int] = [
image_url_id.get(img, 0) for img in _neg_images
]
utter = Utterance(speaker, utter_type, text, pos_images, neg_images)
return utter
def get_knowledge_items(dialog: Dialog, ordinal_number: Dict[int, int],
task: int) -> List[TidyDialog]:
"""Get items for knowledge task from a single dialog.
Args:
dialog (Dialog): Dialog.
ordinal_number (Dict[int, int]): Ordinal numbers.
task (int): Task.
Returns:
List[TidyDialog]: Extracted tidy dialogs.
"""
expected_utter_types = {}
if task == KNOWLEDGE_STYLETIP_SUBTASK:
expected_utter_types = DatasetConfig.utterance_knowledge_styletip_types
elif task == KNOWLEDGE_ATTRIBUTE_SUBTASK:
expected_utter_types = DatasetConfig.utterance_knowledge_attribute_types
elif task == KNOWLEDGE_CELEBRITY_SUBTASK:
expected_utter_types = DatasetConfig.utterance_knowledge_celebrity_types
dialogs: List[TidyDialog] = []
utterances = get_init_pad_utters()
context_size = DatasetConfig.dialog_context_size
utter_type = None
has_shown = False
products = []
selected_products = []
for utter in dialog:
pos_images = [image for image in utter.pos_images if image > 0]
if utter.speaker == USER_SPEAKER:
utterances.append(TidyUtterance(utter))
selected_products = get_products(ordinal_number, utter.text,
products)
utter_type = utter.utter_type
elif utter.speaker == SYS_SPEAKER:
desc = task == KNOWLEDGE_ATTRIBUTE_SUBTASK and has_shown and \
utter_type in DatasetConfig.utterance_recommend_types and \
len(utter.text) > 10
if utter_type in expected_utter_types or desc:
if desc:
selected_products = get_products(ordinal_number,
utter.text, products)
utterances = utterances[-context_size:]
text = copy.deepcopy(utter.text)
special_utter = Utterance(utter.speaker, utter.utter_type,
text, selected_products, [])
special_utter = TidyUtterance(special_utter)
dialogs.append(copy.deepcopy(utterances + [special_utter]))
utter_type = None
utterances.append(TidyUtterance(utter))
has_shown = False
if pos_images:
products = pos_images
if utter_type in DatasetConfig.utterance_recommend_types:
has_shown = True
return dialogs
def get_init_pad_utters() -> List[TidyUtterance]:
"""Get initial padding utterances.
Returns:
List[TidyUtterance]
"""
utters: List[TidyUtterance] = []
for i in range(DatasetConfig.dialog_context_size):
if (DatasetConfig.dialog_context_size - i - 1) % 2 == 0:
speaker = SYS_SPEAKER
else:
speaker = USER_SPEAKER
utter = Utterance(speaker, -1, [], [])
utters.append(TidyUtterance(utter))
return utters
def standardized_dialog(dialog: Dialog) -> Dialog:
"""Standardized raw dialog.
Args:
dialog (Dialog): Raw dialog.
Returns:
Dialog: Standard dialog.
"""
std_dialog: Dialog = []
for utter in dialog:
if not std_dialog and utter.speaker != USER_SPEAKER:
std_dialog.append(Utterance(USER_SPEAKER, -1, [], [], []))
if not std_dialog or utter.speaker != std_dialog[-1].speaker:
std_dialog.append(utter)
else:
std_dialog[-1].utter_type = utter.utter_type
std_dialog[-1].text += utter.text
std_dialog[-1].pos_images += utter.pos_images
#std_dialog[-1].neg_images += utter.neg_images
return std_dialog
def _get_utter_from_dict(vocab: Dict[str, int],
obj_id: Dict[str, int],
utter_dict: dict,
utter_coref: dict,
speaker: str) -> Utterance:
"""Extract Utterance object from JSON dict.
Args:
vocab (Dict[str, int]): Vocabulary.
obj_id (Dict[str, int]): Obj_ID to index.
utter_dict (dict): JSON dict.
utter_coref (dict): JSON dict.
Returns:
Utterance: Extracted Utterance.
"""
if speaker == 'sys':
_speaker: str = 'system'
_utter_type: str = (ast.literal_eval(utter_dict.get('system_transcript_annotated')))[0]['intent'].split(':')[0]
_text: str = utter_dict['system_transcript']
if speaker == 'user':
_speaker: str = 'user'
_utter_type: str = (ast.literal_eval(utter_dict.get('system_transcript_annotated')))[0]['intent'].split(':')[0]
_text: str = utter_dict['transcript']
object_list = []
for state in utter_dict.get('belief_state'):
for sub_state in state['slots']:
for sub_sub_state in sub_state:
if "OBJECT" in sub_sub_state:
object_list.append(int(sub_sub_state[-1]))
_pos_obj: List[str] = []
for product in object_list:
for key in utter_dict:
if utter_dict[key] == product:
_pos_obj.append(key)
#_pos_obj: List[str] = list(utter_coref.keys())
#_neg_obj: List[str] = []
# Some attributes may be empty.
if _text is None:
_text = ""
if _utter_type is None:
_utter_type = ""
if _pos_obj is None:
_pos_obj = []
#if _neg_obj is None:
# _neg_obj = []
# Convert speaker into an integer.
speaker: int = -1
if _speaker == 'user':
speaker = USER_SPEAKER
elif _speaker == 'system':
speaker = SYS_SPEAKER
assert speaker != -1
# Convert utterance type into an integer.
utter_type: int = DatasetConfig.utterance_type_dict.get(_utter_type, 0)
# We don't care the type of system response.
if speaker == SYS_SPEAKER:
utter_type = 0
# Convert text into a list of integers.
words: List[str] = word_tokenize(_text)
text: List[int] = [vocab.get(word.lower(), UNK_ID) for word in words]
# Images
pos_obj: List[str] = _pos_obj
#neg_obj: List[int] = [obj_id.get(img, 0)
# for img in _neg_obj]
utter = Utterance(speaker, utter_type, text, pos_obj) #, neg_obj)
return utter