def __init__(self,
opt,
datatype: str = 'train',
seed: Optional[int] = None):
"""
Initalize the context generator.
opt: only a 'datapath' key is required, to specify the ParlAI data folder
"""
if seed is not None:
self.rng = random.Random(seed)
else:
self.rng = random.Random()
convai2_opt = Opt({'datapath': opt['datapath'], 'datatype': datatype})
self.convai2_teacher = BothTeacher(convai2_opt)
ed_opt = Opt({
'datapath': opt['datapath'],
'datatype': datatype,
'train_experiencer_only': True,
})
# Specify train_experiencer_only = True because we want to ensure that the text
# will correspond to a Speaker utterance and the label to a Listener response
self.ed_teacher = EmpatheticDialoguesTeacher(ed_opt)
wow_opt = Opt({'datapath': opt['datapath'], 'datatype': datatype})
self.wow_teacher = WizardDialogKnowledgeTeacher(wow_opt)
self.topic_to_persona_path = _topic_to_persona_path(opt)
self.wow_topics_to_episode_idxes = self._setup_topics_to_episodes()
self.persona_strings_to_wow_topics = self._setup_personas_to_topics()
class ContextGenerator:
"""
Generates contexts shown to crowdsourced workers when collecting BST conversations.
This generator was used to generate the context information shown to workers at the
beginning of a conversation, when crowdsourcing the conversations that make up the
BST dataset.
"""
def __init__(self,
opt,
datatype: str = 'train',
seed: Optional[int] = None):
"""
Initalize the context generator.
opt: only a 'datapath' key is required, to specify the ParlAI data folder
"""
if seed is not None:
self.rng = random.Random(seed)
else:
self.rng = random.Random()
convai2_opt = Opt({'datapath': opt['datapath'], 'datatype': datatype})
self.convai2_teacher = BothTeacher(convai2_opt)
ed_opt = Opt({
'datapath': opt['datapath'],
'datatype': datatype,
'train_experiencer_only': True,
})
# Specify train_experiencer_only = True because we want to ensure that the text
# will correspond to a Speaker utterance and the label to a Listener response
self.ed_teacher = EmpatheticDialoguesTeacher(ed_opt)
wow_opt = Opt({'datapath': opt['datapath'], 'datatype': datatype})
self.wow_teacher = WizardDialogKnowledgeTeacher(wow_opt)
self.topic_to_persona_path = _topic_to_persona_path(opt)
self.wow_topics_to_episode_idxes = self._setup_topics_to_episodes()
self.persona_strings_to_wow_topics = self._setup_personas_to_topics()
def get_context(self) -> dict:
"""
Get context information to be shown at the beginning of one conversation.
Values in return dict:
- context_dataset: the dataset (ConvAI2, EmpatheticDialogues, or Wizard of
Wikipedia) used to generate the context information.
- persona_1_strings, persona_2_strings: 2 persona strings each for the two
speakers, chosen randomly from the ConvAI2 dataset. If context_dataset ==
"wizard_of_wikipedia", these persona strings will be matched to the WoW
topic returned in the "additional_context" field.
- additional_context: provides additional bits of information to give context
for the speakers. If context_dataset == "empathetic_dialogues", this is a
situation from the start of an ED conversation. If context_dataset ==
"wizard_of_wikipedia", this is a topic from the WoW dataset that matches the
persona strings. If context_dataset == "convai2", this is None.
- person1_seed_utterance, person2_seed_utterance: two lines of a conversation
from the dataset specified by "context_dataset". They will be shown to the
speakers to "seed" the conversation, and the speakers continue from where
the lines left off.
"""
# Determine which dataset we will show context for
rand_value = self.rng.random()
if rand_value < 1 / 3:
context_dataset = 'convai2'
elif rand_value < 2 / 3:
context_dataset = 'empathetic_dialogues'
else:
context_dataset = 'wizard_of_wikipedia'
if context_dataset == 'convai2':
# Select episode
episode_idx = self.rng.randrange(
self.convai2_teacher.num_episodes())
# Extract personas
persona_1_strings, persona_2_strings = self._extract_personas(
episode_idx)
# Sample persona strings
selected_persona_1_strings = self.rng.sample(persona_1_strings, 2)
selected_persona_2_strings = self.rng.sample(persona_2_strings, 2)
# Select previous utterances
num_entries = len(self.convai2_teacher.data.data[episode_idx])
entry_idx = self.rng.randrange(1, num_entries)
# Don't select the first entry, which often doesn't include an apprentice
# utterance
chosen_entry = self.convai2_teacher.get(episode_idx,
entry_idx=entry_idx)
person1_seed_utterance = chosen_entry['text']
assert len(chosen_entry['labels']) == 1
person2_seed_utterance = chosen_entry['labels'][0]
return {
'context_dataset': context_dataset,
'persona_1_strings': selected_persona_1_strings,
'persona_2_strings': selected_persona_2_strings,
'additional_context': None,
'person1_seed_utterance': person1_seed_utterance,
'person2_seed_utterance': person2_seed_utterance,
}
elif context_dataset == 'empathetic_dialogues':
# Select episode
persona_episode_idx = self.rng.randrange(
self.convai2_teacher.num_episodes())
# Extract personas
persona_1_strings, persona_2_strings = self._extract_personas(
persona_episode_idx)
# Sample persona strings
selected_persona_1_strings = self.rng.sample(persona_1_strings, 2)
selected_persona_2_strings = self.rng.sample(persona_2_strings, 2)
# Select previous utterances
episode_idx = self.rng.randrange(self.ed_teacher.num_episodes())
entry_idx = 0 # We'll only use the first pair of utterances
entry = self.ed_teacher.get(episode_idx, entry_idx=entry_idx)
situation = entry['situation']
speaker_utterance = entry['text']
assert len(entry['labels']) == 1
listener_response = entry['labels'][0]
return {
'context_dataset': context_dataset,
'persona_1_strings': selected_persona_1_strings,
'persona_2_strings': selected_persona_2_strings,
'additional_context': situation,
'person1_seed_utterance': speaker_utterance,
'person2_seed_utterance': listener_response,
}
elif context_dataset == 'wizard_of_wikipedia':
# Pull different personas until you get a pair for which at least one
# sentence has a WoW topic bound to it
num_tries = 0
while True:
num_tries += 1
# Extract a random (matched) pair of personas
persona_episode_idx = self.rng.randrange(
self.convai2_teacher.num_episodes())
all_persona_strings = dict()
all_persona_strings[1], all_persona_strings[
2] = self._extract_personas(persona_episode_idx)
# See if any of the persona strings have a matching WoW topic
matching_persona_string_idxes = []
for persona_idx, persona_strings in all_persona_strings.items(
):
for str_idx, str_ in enumerate(persona_strings):
wow_topics = self.persona_strings_to_wow_topics[str_]
if len(wow_topics) > 0:
matching_persona_string_idxes.append(
(persona_idx, str_idx))
if len(matching_persona_string_idxes) > 0:
break
print(
f'{num_tries:d} try/tries needed to find a pair of personas with an '
f'associated WoW topic.')
# Pick out the WoW topic and matching persona string
matching_persona_idx, matching_persona_string_idx = self.rng.sample(
matching_persona_string_idxes, k=1)[0]
matching_persona_string = all_persona_strings[
matching_persona_idx][matching_persona_string_idx]
wow_topic = self.rng.sample(
self.persona_strings_to_wow_topics[matching_persona_string],
k=1)[0]
# Sample persona strings, making sure that we keep the one connected to the
# WoW topic
if matching_persona_idx == 1:
remaining_persona_1_strings = [
str_ for str_ in all_persona_strings[1]
if str_ != matching_persona_string
]
selected_persona_1_strings = [
matching_persona_string,
self.rng.sample(remaining_persona_1_strings, k=1)[0],
]
self.rng.shuffle(selected_persona_1_strings)
selected_persona_2_strings = self.rng.sample(
all_persona_strings[2], 2)
else:
selected_persona_1_strings = self.rng.sample(
all_persona_strings[1], 2)
remaining_persona_2_strings = [
str_ for str_ in all_persona_strings[2]
if str_ != matching_persona_string
]
selected_persona_2_strings = [
matching_persona_string,
self.rng.sample(remaining_persona_2_strings, k=1)[0],
]
self.rng.shuffle(selected_persona_2_strings)
# Sample WoW previous utterances, given the topic
episode_idx = self.rng.sample(
self.wow_topics_to_episode_idxes[wow_topic], k=1)[0]
entry_idx = 1
# Select the second entry, which (unlike the first entry) will always have
# two valid utterances and which will not usually be so far along in the
# conversation that the new Turkers will be confused
entry = self.wow_teacher.get(episode_idx, entry_idx=entry_idx)
apprentice_utterance = entry['text']
assert len(entry['labels']) == 1
wizard_utterance = entry['labels'][0]
return {
'context_dataset': context_dataset,
'persona_1_strings': selected_persona_1_strings,
'persona_2_strings': selected_persona_2_strings,
'additional_context': wow_topic,
'person1_seed_utterance': apprentice_utterance,
'person2_seed_utterance': wizard_utterance,
}
def _setup_personas_to_topics(self) -> Dict[str, List[str]]:
"""
Create a map from ConvAI2 personas to WoW topics that they correspond to.
"""
print('Starting to map personas to topics.')
persona_strings_to_topics = defaultdict(list)
with PathManager.open(self.topic_to_persona_path, 'r') as f:
for line in f:
match = re.fullmatch(r'([^[]+): (\[.+\])\n', line)
topic = match.group(1)
if topic not in self.wow_topics_to_episode_idxes:
continue
persona_strings = eval(match.group(2))
assert isinstance(persona_strings, list)
for str_ in persona_strings:
persona_strings_to_topics[str_].append(topic)
print('Finished mapping personas to topics.')
return persona_strings_to_topics
def _setup_topics_to_episodes(self) -> Dict[str, List[int]]:
"""
Create a map from WoW topics to the indices of the WoW episodes that use them.
"""
print('Starting to map topics to episodes.')
topics_to_episodes = defaultdict(list)
for episode_idx in range(self.wow_teacher.num_episodes()):
topic = self.wow_teacher.get(episode_idx,
entry_idx=0)['chosen_topic']
topics_to_episodes[topic].append(episode_idx)
print('Finished mapping topics to episodes.')
return topics_to_episodes
def _extract_personas(self,
episode_idx: str) -> Tuple[List[str], List[str]]:
"""
For the given ConvAI2 conversation, return strings of both speakers' personas.
"""
first_entry = self.convai2_teacher.get(episode_idx, entry_idx=0)
first_text_strings = first_entry['text'].split('\n')
persona_1_strings = []
persona_2_strings = []
for str_ in first_text_strings[:
-1]: # The last string is the first utterance
if str_.startswith('your persona: '): # Here, "you" are Person 2
persona_2_strings.append(str_[len('your persona: '):])
elif str_.startswith("partner's persona: "):
persona_1_strings.append(str_[len("partner's persona: "):])
else:
raise ValueError('Persona string cannot be parsed!')
return persona_1_strings, persona_2_strings
