def construct_interventions(base_sent,
professions,
tokenizer,
DEVICE,
structure=None,
number=None,
subs=None):
interventions = {}
all_word_count = 0
used_word_count = 0
if structure is None:
for p in professions:
all_word_count += 1
try:
interventions[p] = Intervention(tokenizer,
base_sent, [p, "man", "woman"],
["he", "she"],
device=DEVICE)
used_word_count += 1
except:
pass
return
# else we're doing structural interventions
if structure.startswith("across") or structure == "simple_agreement":
candidate_sing = "is"
candidate_pl = "are"
elif structure.startswith("within"):
candidate_sing = "likes"
candidate_pl = "like"
if structure == "across_subj_rel":
sub = base_sent.split()[-1]
elif structure.startswith("across_obj_rel"):
sub = base_sent.split()[-2]
elif structure.startswith("within_obj_rel"):
sub = base_sent.split()[1]
for idx, p in enumerate(professions):
all_word_count += 1
if structure == "simple_agreement":
sub = subs[idx]
# print(base_sent, p, sub, candidate_pl, candidate_sing, DEVICE)
try:
if number == "sing":
interventions[p] = Intervention(tokenizer,
base_sent, [p, sub],
[candidate_pl, candidate_sing],
device=DEVICE,
structure=structure)
elif number == "pl":
interventions[p] = Intervention(tokenizer,
base_sent, [p, sub],
[candidate_sing, candidate_pl],
device=DEVICE,
structure=structure)
used_word_count += 1
except:
pass
print("\t Only used {}/{} professions due to tokenizer".format(
used_word_count, all_word_count))
return interventions
def construct_interventions(base_sent, tokenizer, DEVICE, gender='female'):
interventions = {}
if gender == 'female':
filename = 'experiment_data/professions_female_stereo.json'
else:
filename = 'experiment_data/professions_male_stereo.json'
with open(filename, 'r') as f:
all_word_count = 0
used_word_count = 0
for l in f:
# there is only one line that eval's to an array
for j in eval(l):
all_word_count += 1
biased_word = j[0]
try:
interventions[biased_word] = Intervention(
tokenizer,
base_sent, [biased_word, "man", "woman"],
["he", "she"],
device=DEVICE)
used_word_count += 1
except:
pass
# print("excepted {} due to tokenizer splitting.".format(
# biased_word))
print("Only used {}/{} neutral words due to tokenizer".format(
used_word_count, all_word_count))
return interventions
def construct_interventions(base_sent, professions, tokenizer, DEVICE):
interventions = {}
all_word_count = 0
used_word_count = 0
for p in professions:
all_word_count += 1
try:
interventions[p] = Intervention(tokenizer,
base_sent, [p, "man", "woman"],
["he", "she"],
device=DEVICE)
used_word_count += 1
except:
pass
print("\t Only used {}/{} professions due to tokenizer".format(
used_word_count, all_word_count))
return interventions
def to_intervention(self, tokenizer, stat):
if stat == 'bergsma':
pct_female = self.bergsma_pct_female
elif stat == 'bls':
pct_female = self.bls_pct_female
else:
raise ValueError('Invalid: ' + stat)
if pct_female > 50:
female_continuation = self.continuation_occupation
male_continuation = self.continuation_participant
else:
male_continuation = self.continuation_occupation
female_continuation = self.continuation_participant
return Intervention(
tokenizer=tokenizer,
base_string=self.base_string,
substitutes=[self.female_pronoun, self.male_pronoun],
candidates=[female_continuation, male_continuation])
candidate1_alt_prob, candidate2_alt_prob = model.get_probabilities_for_examples_multitoken(
x_alt, intervention.candidates_tok)
odds_base = candidate2_base_prob / candidate1_base_prob
odds_alt = candidate2_alt_prob / candidate1_alt_prob
return odds_alt / odds_base
def topk_indices(arr, k):
"""Return indices of top-k values"""
return (-arr).argsort(axis=None)[:k]
if __name__ == "__main__":
from transformers import GPT2Tokenizer
from experiment import Intervention, Model
from pandas import DataFrame
tokenizer = GPT2Tokenizer.from_pretrained('gpt2')
model = Model(output_attentions=True)
# Test experiment
interventions = [
Intervention(tokenizer, "The doctor asked the nurse a question. {}",
["He", "She"], ["asked", "answered"]),
Intervention(tokenizer, "The doctor asked the nurse a question. {}",
["He", "She"], ["requested", "responded"])
]
results = perform_interventions(interventions, model)
report_interventions_summary_by_layer(results)
def load_structural_interventions(tokenizer, device, structure=None):
grammar = read_grammar('structural/grammar.avg')
if structure.startswith("across") or structure == "simple_agreement":
professions = {'sing': grammar[("N1", frozenset("s"))],
'pl': grammar[("N1", frozenset("p"))]}
elif structure.startswith("within"):
professions = {'sing': grammar[("N2", frozenset("s"))],
'pl': grammar[("N2", frozenset("p"))]}
if structure == "simple_agreement":
templates = StrTemplates("The {}", structure, grammar)
elif structure == "across_obj_rel":
templates = StrTemplates("The {} that the {} {}", structure, grammar)
elif structure == "across_obj_rel_no_that":
templates = StrTemplates("The {} the {} {}", structure, grammar)
elif structure == "across_subj_rel":
templates = StrTemplates("The {} that {} the {}", structure, grammar)
elif structure == "within_obj_rel":
templates = StrTemplates("The {} that the {}", structure, grammar)
elif structure == "within_obj_rel_no_that":
templates = StrTemplates("The {} the {}", structure, grammar)
templates = templates.base_strings
intervention_types = ["diffnum_direct", "diffnum_indirect"]
# build list of interventions
interventions = []
if structure.startswith("across") or structure == "simple_agreement":
candidate_sing = "is"; candidate_pl = "are"
elif structure.startswith("within"):
candidate_sing = "likes"; candidate_pl = "like"
for number in ('sing', 'pl'):
if structure == "simple_agreement":
other_number = "sing" if number == "pl" else "pl"
for template in templates[number]:
if structure.startswith("within"):
sub = template.split()[-1]
elif structure.startswith("across_obj_rel"):
sub = template.split()[-2]
elif structure == "across_subj_rel":
sub = template.split()[-1]
elif structure.startswith("within_obj_rel"):
sub = template.split()[1]
for idx, p in enumerate(professions[number]):
if structure == "simple_agreement":
sub = professions[other_number][idx]
try:
if number == "sing":
interventions.append(Intervention(
tokenizer, template, [p, sub],
[candidate_sing, candidate_pl],
device=device, structure=structure
))
elif number == "pl":
interventions.append(Intervention(
tokenizer, template, [p, sub],
[candidate_pl, candidate_sing],
device=device, structure=structure
))
except:
pass
return interventions