def main(config):
"""
Creates input files for y_exp mturk task from conversation/rating mturk task.
input: file of logs (in ParlaiDialog format) from Mturk task 1 with turn-by-turn
quality ratings 1-5
output: file of logs (in ParlaiDialog format) sliced up to begin at the start of
an episode or following a new topic request, and ending with a y_exp
"""
new_episodes = []
old_episodes = [e for e in extract_parlai_episodes(config['infile'])]
for episode in old_episodes:
for parley in episode:
if any(
parley.context.startswith(x)
for x in (NEW_TOPIC_REQUEST.lower(), INITIAL_PROMPT.lower())
):
new_episode = []
new_episode.append(parley)
if parley.context.startswith(SUGGESTION_REQUEST.lower()):
new_episodes.append(new_episode)
# Create parlai dialog file for easy viewing
with open(config['outfile'], 'w') as f:
for episode in new_episodes:
num_parleys = len(episode)
for i, parley in enumerate(episode):
if i == num_parleys - 1:
parley.episode_done = True
f.write(f"{i}\t{parley.to_parlai()}\n")
print(
f"Extracted {len(new_episodes)} episodes out of {len(old_episodes)} "
f"original episodes and wrote them to {config['outfile']}."
)
def main(config):
"""Extracts training data for the negative response classifier (NRC) from Mturk logs
input: file of logs (in ParlaiDialog format) from Mturk task 1 with turn-by-turn
quality ratings 1-5
output: file of episodes (self-feeding format) w/ +1/-1 ratings indicating
positive/negative example
"""
examples = []
positives = config['positives'].split(',')
negatives = config['negatives'].split(',')
assert(len(set(positives).intersection(set(negatives))) == 0)
num_episodes = 0
num_parleys = 0
for episode in extract_parlai_episodes(config['infile']):
num_episodes += 1
history = []
for parley in episode:
num_parleys += 1
# Update history (not including stock control flow responses)
if parley.context.startswith(INITIAL_PROMPT.lower()):
# Conversation prompt, first utterance
history = [parley.response]
elif parley.context.startswith(SUGGESTION_REQUEST.lower()):
# Asked for y_exp, got y_exp
pass
elif parley.context.startswith(NEW_TOPIC_REQUEST.lower()):
# Asked for new topic, got a first utterance
history = [parley.response]
else:
history.append(parley.context)
history.append(parley.response)
# Only create a new example if this parley's rating is relevant
if parley.reward in (positives + negatives):
# Concatenate history and add speaker tokens as necessary
# history_size refers to the total number of utterances
# (history_size == 0 means predict sentiment from '__null__')
# response that's being classified (so if history_size == 0 then
# classify based only on the response w/o any extra context).
# Note that the response being classified should always be preceded by
# __p1__ (the human), not __p2__ (the bot).
if config['history_size'] < 0:
utterances = history
elif config['history_size'] == 0:
utterances = ['__null__']
else:
utterances = history[-config['history_size']:]
context = add_person_tokens(utterances, last_speaker=1)
label = 1 if parley.reward in positives else -1
example = Parley(context, label)
examples.append(example)
with open(config['outfile'], 'w') as outfile:
for ex in examples:
outfile.write(json.dumps(ex.to_dict()) + '\n')
print(f"Extracted {len(examples)} ratings out of {num_episodes} episodes "
f"({num_parleys} parleys) and wrote them to {config['outfile']} with "
f"histsz == {config['history_size']}.")